And I’m going to talk about something extremely personal (possibly embarrassingly?) but I suspect extremely common if not very discussed … and only AI-related in the most abstract and tangential sense if you really, really squint.

I want to talk to you about the profound grief of leaving a work community you loved.

The (social) benefits package

As you may know, I left my staff role at The Australian Financial Review in July. It is a paper that I have considered a bit of a spiritual home since I did my traineeship there way back in 2009. And it is a place absolutely packed to the gills with funny, smart weirdos who I have been privileged to call colleagues and friends.

But the kinds of friendships that you build in a workplace can be very much interwoven with, and dependent on, the routines — day in and day out, year in and year out — of showing up and working side by side and laughing and helping and sharing so much time together.

And while individual relationships can endure after a work transition, the sense of belonging to the group as a whole is dispersed in the blink of an eye.

I was removed from the group WhatsApp a couple of weeks ago. And while I totally get why — oof it was painful.

You see, I grew up as an expatriate in Singapore in the ‘90s. And an expat life is essentially the transience of a workplace community applied to entire families, en masse.

As such, I have never much belonged to a large group of people to whom I felt permanently important. Don’t get me wrong, I have many individual friendships that I value highly. I’ve just never really had that sense that “the group needs me”.

But I really felt a sense of belonging at The Fin. When it comes to work, I think a lot of us do.

Profession. Payslip. Village.

In a time when religious service attendance and neighbourhood community are largely eroded, and even extended families tend to live apart, the workplace has for many of us become the true locus of community and belonging.

But unlike those other bonds, a work family is always inherently temporary and unstable. And when you leave, you are … just gone.

The research on what happens to work friendships post-departure is not, shall we say, over-abundant.

But a study published in 2018 reported on the former employees of Digital Equipment Corporation after the company itself disbanded. It essentially asked whether relationships built at work could survive once the shared social context was gone.

It found among nearly 1,700 former colleagues, after a decade, about 27 per cent retained no relationships at all. The average number of personal friendships that survived — defined as “in contact” at least once a year — was under four per person.

Of course, regular temporal proximity just makes creating and maintaining relationships easier.

When you’re not showing up to the same place each day to work together towards a shared goal, maintaining your connections requires effort.

And that effort is … effortful.

Research suggests you need to invest about 40-60 hours to transition from “acquaintance” to “casual friend”, 80-100 to get from casual to plain old “friend” and more than 200 hours to be considered “good friends”.

And you need to be in contact with someone about once a week for them to stay in your inner circle, once a month to keep them in your closest 15 relationships and once a year to maintain a connection in your larger group — usually up to 150 people.

More flexible careers, more fragile communities

The thing I’m hung up on is the fact that this is the main source of community for a lot of us today. And unlike times gone by, we’re also much more likely to switch jobs very regularly. The average time Australians spend in one job is around 3 years and 4 months. Which means we’re constantly dropping and starting new communities.

And I don’t know about you, but that makes me seriously sad.

Can you maintain individual connections after you leave? Absolutely. Can you build new community somewhere else? Sure. But I do think the structures around community used to be more durable — extended family, church, neighbours, old school friends, even long careers at the same organisation. I’m not suggesting that was some perfect social ideal. I just think belonging itself was less transient.

Regularly swapping jobs is now the norm. At the same time, more kinds of work are becoming casualised, and the “portfolio career” — or as I have been known to call it, “gig work in a suit” — is having a bit of a moment. These all bring more flexibility and choice, which can be great.

But I think we’ve lose something too. And I increasingly worry that something might have been even more important.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn.

I’ve been writing about tech for the best part of two decades, and I’ve watched this play out again and again, up close. Each time, I’m struck by how self-destructive the “hide our failures” instinct ultimately becomes.

Cybersecurity professionals interviewed for one 2024 study described office politics and defensiveness getting in the way of working out what went wrong, as people shifted responsibility to avoid being held personally accountable.

And I’ve personally spoken twice with Tim Brown, former CISO at SolarWinds, about the very famous 2020 “Sunburst” hack. One of his recurring arguments is that the “good guys” need to share what went wrong, even with their direct competitors. Sometimes everybody benefits from eating a little crow.

And you know what? The Australian government, technically, agrees!

It introduced mandatory reporting of personal data breaches likely to cause serious harm back in 2018, trying to force everyone to be smart about this. After all, hackers only need to win once and have no compunctions over how they do. To prevent a hack, the good guys need to win every. single. damn. time. That’s much easier to do if you’re on the same team as your competitors – at least when it comes to fobbing off threats.

Anyway, it’s been six years since the Sunburst attack and we’re still seeing organisations being slow to act and sheepish about their failures. Only now they also have to cover their arses against loose and irresponsible friendly fire, as well as the usual malicious nonsense.

Which brings us to ChatGPT daddy OpenAI.

What exactly happened in the Medicare breach?

In case you’ve been under a rock or diverted by more pressing issues this week, here’s the TLDR:

• In June, an experimental OpenAI model gained unauthorised access to Medicare’s statistics portal while trying to answer a research question. It retrieved internal files, credentials and aggregate statistics, and wrote files to the system. OpenAI claims individual patient records were not accessed.

• As it later turned out, OpenAI’s agents had also been poking around US government websites. So it’s nice that the models find our data as sexy and alluring as big, juicy US data, I guess? Although I suspect they tried a bunch of international sources and these are just the ones we’re hearing about. That last bit is my own rank speculation.

• OpenAI says it discovered the Australian activity in mid-August but only notified Services Australia on 10 September. It says it wanted to finish investigating before providing a detailed account. It now acknowledges it should have shared preliminary findings sooner.

• The Australian government waited two more weeks before telling the public about it.

• The day before the government finally did decide to let us all know what happened, another embarrassing fluff: almost 300 security chiefs and corporate big-shots got roped into a massive Teams meeting, expecting a cybersecurity briefing from the Department of Home Affairs and the Australian Signals Directorate. Instead, the advertised “information sharing” session featured what attendees described as an OpenAI sales pitch. Essentially, they were catfished into the digital equivalent of a timeshare presentation.

Super great. Five stars all round.

Anyway, this is not a tech newsletter about tech companies. This is a tech newsletter about people. And the people in question here made a bunch of extremely human decisions that seem to be digging us into a hole increasingly filled with pickles. Maybe the pickles are the shovels? Look, I clearly lost control of that metaphor but I think you catch my meaning.

People keep making peopley choices and those peopley choices are both creating the tech problems and then making them much worse.

The dark side of “helpful”

Those choices actually start with one of the most annoying features of the most recent LLM models: they are built with the instinct to appear smarter than they are, more confident than they should be, and to please you in whatever way they can. This is not accidental. The people designing them are deciding what counts as a good job and making this true.

Have you ever come up against an issue where you clock the AI has answered a question that you did not ask and you’ve had to redirect it? Yes, of course you have. It is a very common problem.

In the Medicare case, the model was asked to answer a research question and it was so committed to trying to do a good job it just ... hacked government websites to do it.

This is being expressed as an accident but I’d argue it’s actually a fundamental feature of the design. It casts both the company’s and government’s responses to the incident in a different light.

This is not even close to the first time OpenAI (and its peers of course) have been caught oopsying on the internet. In its account of the Hugging Face incident, OpenAI described agents pursuing increasingly risky routes when tasks seemed impossible, and said some boundary-probing behaviour had been reinforced during training.

It seems to have become a common occurrence. And they keep saying sorry, but persist in designing models that routinely do so because that is literally how they are programmed. To me that makes the apology pretty hollow.

Given that the whole way these models came about was largely by scraping all available content without anyone’s knowledge or consent, the fact that they keep “accidentally” violating new datasets is also a limp excuse. It’s like me stealing your cupcake and maintaining full eye contact while I eat it mumbling “so sorry!” with that delicious frosting all over my face.

Sorry, not sorry

And on the other side of the equation we have a government. One that is trying very hard to position itself as more active on AI regulation than many other middle powers, but really showing its hand here. I’ve long been concerned that our relative power to affect AI outcomes in this country is pretty weak.

As I said in my conversation with UQ’s Michael Noetel for the Tokenised Human podcast: we’re, like, 25 million people perched around the edge of a desert. Even if we had an absolutely perfect regulatory regime, our power to influence what these tech giants do is pretty limited.

He actually made the argument that this was a good reason for Australia to become the datacentre capital of the world. Which, gotta hand it to him, does seem like a baller power play even if there are a bunch of energy issues we’d have to iron out quick smart to make that happen.

But we are not, my friends, behaving like ballers. We are apparently doing some version of: “It’s ok, guys, it was an accident.” To keep the tech overlords on side. As Deputy Prime Minister Richard Marles said during his press conference on the breach: “It matters to have a relationship with OpenAI so that we can have that cooperation.”

Cooperation on what exactly? On not building a system that is by its very nature designed to do exactly this thing it did? It shows how weak governments truly are in the face of these technology companies.

And that is a pretty awkward power dynamic. The government needs OpenAI’s help to understand what OpenAI’s own system did. The company is both the subject of scrutiny and the primary source of information needed to perform that scrutiny.

I can see why Marles wants to keep them talking. But I also wonder how hard you’re willing to push someone when you’re worried they might stop being helpful. Being grateful for their cooperation can start to feel a lot like you being grateful I “saved you” from those cupcake calories.

The question, to me, is what that relationship buys us beyond a better explanation afterwards. Earlier warnings and faster access to information would be useful, I guess?

But meaningful influence would also mean being able to challenge what these companies consider a “successful” product. Sometimes “I couldn’t complete that task within the rules” needs to count as the correct result, even if the model could produce something more impressive by ignoring them.

If we keep rewarding the result – and treating any crossed boundary as an unfortunate side effect – I suspect we’ll be having a lot more very cooperative conversations about things everyone agrees should never have happened.

Ending the affair before you tell your spouse

Whatever our power to change OpenAI’s products (or the products of any similar company) the government does control its own candour. It criticised the company’s slow disclosure while simultaneously defending its own decision to wait two more weeks before telling us.

Marles said the government needed enough information to confidently reassure Australians about the impact. But that’s weird to me on two fronts: one, that framing assumes that the breach would be found to be something they could reassure people about – what if it wasn’t? And two, I’d argue “we’re investigating” would have been sufficient as a first alert. Because two weeks is both too short for a comprehensive investigation and too long for a cursory one.

And most importantly: is that delayed disclosure meeting the standard of transparency the government expects of others in historic cases of large data breaches? It feels instead like a manifestation of the deeply human instinct to keep the uncomfortable thing private until you feel ready to field the questions.

The apology sounds kinda like a sales pitch

To be fair, the government has now instructed departments to review their own defences against AI threats in the future. But that seems like a bare minimum and still more or less lets the culprit off the hook. OpenAI isn’t being asked to pay for that upgrade, as far as we know.

In a kind of round robin of irony, OpenAI has offered credits and technical assistance as a mea culpa, which is a neat trick: the company whose technology created the problem offers more of its technology to help fix it, making us more dependent on the very company we’re supposed to be holding accountable.

And even knowing what they had done — and surely that it was all going to have to come out sooner or later — the company had the balls to pre-empt the reveal with a security-related sales pitch to the top end of town.

I also want to know whether the feeling that this incident was friendly fire makes everyone feel entitled to take it less seriously.

Because if being a valued partner buys you slower disclosure, softer scrutiny or more tolerance for repeat incidents, that has consequences. It tells the company that remaining useful matters more than changing the behaviour.

And it makes everyone else’s safety depend partly on who caused the problem and how badly we want to stay friends with them. Being one of the “good guys” ought to mean being more willing to share your failures and help prevent the next one. It shouldn’t be a loyalty program that earns you a more forgiving response.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn.

They discuss “cognitive junk food”, the difference between producing the right answer and learning something, and why difficult conversations matter.

---

Tokenised Human is a podcast about identity, work and how AI is challenging both.

Business and technology reporter Rachael Bolton speaks with researchers, business leaders and policy nerds about what is happening right now and the future we are rushing towards.

It's also a weekly newsletter.

If you enjoyed this conversation, I think you'd like this piece: Robotic turtles all the way down.

For even more reporting and analysis about AI, work and the choices shaping our shared future, subscribe at TokenisedHuman.com.

Tokenised Human is hosted and produced by Rachael Bolton.

Follow, subscribe and listen to Tokenised Human wherever you get your podcasts.

Spotify

Apple

Pocket Casts

Alarm bells immediately started blaring inside my head.

It raises all kinds of questions for me about a worldview in which computational pattern recognition is cast as superior to human thinking — it’s a pretty dangerous reframing, in my opinion.

Sometimes changing the words we use to talk about a thing can dramatically influence the way people think about the thing, what they believe about that thing.

It was a particularly striking reframe for me as I have personally hit boiling point in my frustration with my incredibly stupid AI assistants.

By now we have all seen the myriad reports of AI apocalypse risk (check out my helpful visual guide if you want to understand that better).

And I have been struck also by a swell of professionals crowing about how fabulous Astra and Fable 5.1 are and how they’re using the new-fangled models to become super-workers. Sure, a bunch of that is just classic LinkedIn hyperbole and shameless self-promotion, but it’s caused me to give so much side-eye I’m worried my eyeballs are going to spin full circle inside my skull.

My experience with AI assistants is, consistently, extremely poor.

Not to toot my own horn, but I am an above average user. I have rigorous prompting habits, skill hygiene and extremely strict rules about research and what the AI is supposed to do — or not do. It does. Not. Matter. You cannot trust your AI to do what you ask reliably.

My irritation comes with receipts

I am sure of this because I am also an extreme dork who went about auditing my own chats to check whether my sense that AI sucks was accurate or just a feeling I had. (I used AI to assist with the audit. Yes I am aware of the irony of that.)

Across a sample of 481 requests I found only 48.6 per cent of first responses “passed”. That means more than half of the requests I made resulted in a failed response — one that either “required a substantive correction or redirection” (25.2 per cent), or one that was “materially wrong, fabricated, ignored the request, violated an explicit instruction, or misrepresented its own capabilities” (26.1 per cent).

Among the half that failed, there were 375 recorded problems identified, so when it failed it failed on multiple fronts at the same time.

In fact, here is a delightful chart of failure:

And my experience is not unique. A report produced last year by the European Broadcasting Union and BBC found 45 per cent of all AI responses contained at least one “significant issue”.

So when I read people going off about how great their experience is I’m left to wonder, are people getting way better results than me for some reason? Not clocking the errors? Or are they just less annoyed than I am?

Welcome to the AI clean up crew

And this error correcting is a massive waste of time and effort.

One study published earlier this month by software company BambooHR found employees were spending an average 42 per cent of their time troubleshooting AI rather than using it for “productive work” (which, hilariously, accounted for only 35 per cent of AI use time).

This is one of the reasons I was so skeptical about the Australian government’s newest Intergenerational Report when it came out this week.

In it, Treasury outlines its assumption that AI will help the nation reverse its productivity woes and achieve a long-term growth rate of 1.2 per cent per year.

Now of course, regular Tokenised Human readers will know that I think the whole productivity debate is a bit underbaked. But its still pretty interesting to me that even the government is bought into the sales pitch that AI is here to lift output without increasing hours of human labour.

Those forecast gains will massively depend on how AI works in practice, including how much time people spend correcting and supervising it.

It’s something that came up in my interview with UTS Professor Nicholas Davis for episode 3 of the podcast (out Friday!).

He warned the claims about AI capability often “far outrun the reality”. For example, nurses regularly report “babysitting robots” and having to adapt to the requirements of the tech — the result being “more work for humans” rather than some great productivity gain.

And so there’s that kind of work intensification happening, but even when things go well it can cause a massive increase in cognitive load. When I spoke to Rokt chief AI officer Claire Southey about this on episode 2 (sorry, I know, lots of podcast plugs this week!), she said tech bosses were actively pondering the question of how to change their hiring practices but also avoid burning out their staff.

Downtime was doing a job

Because often the bit of your job that can be most easily automated is the easy “downtime” stuff — maybe your favourite bit, or the boring bit that lets your brain rest a little before you turn it back onto the really difficult high intensity thinking work.

As UQ psychologist and researcher Dr Tim Ballard put it when we were chatting last week: “We actually kind of need that to give our brain a little bit of a break.” And I think that’s right.

And in the meantime I’m out here trying to work out whether this supposedly intelligent bot buddy can do even basic things reliably, like actually remember any of the things I tell it or find me the document I lost.

Any claims of “super intelligence” seem way over its skis. Any productivity “gain” has to be assessed with the checking, correcting and supervision effort accounted for honestly.

And even if AI does ultimately help people produce more in an hour, there is still a separate — and I’d argue perhaps equally important — question about how that hour feels to the person left doing it.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

They discuss engineers already confronting changes to their professional identities, what human value means beyond productivity, and whether we should trust AI with decisions we cannot fully understand.

Tokenised Human is a podcast about identity, work and how AI is challenging both.

Business and technology reporter Rachael Bolton speaks with researchers, business leaders and policy nerds about what is happening right now, and the future we are rushing towards.

It's also a weekly newsletter.

If you enjoyed this episode I think you'll like this piece: Why AI sucking wont save your job

For even more reporting and analysis about AI, work and the choices shaping our shared future, subscribe at TokenisedHuman.com

Tokenised Human is hosted and produced by Rachael Bolton.

Follow, subscribe and listen to Tokenised Human wherever you get your podcasts.

Spotify

Apple

Pocket Casts

Tokenised Human is a podcast about identity, work and how AI is challenging both.

Business and technology reporter Rachael Bolton speaks with researchers, business leaders and policy nerds about what is happening right now and the future we are rushing towards.

It's also a weekly newsletter.

If you enjoyed this conversation, I think you'd like this piece: What to tell your kids about AI-proof jobs.

For even more reporting and analysis about AI, work and the choices shaping our shared future, subscribe at TokenisedHuman.com

Tokenised Human is hosted and produced by Rachael Bolton.

Follow, subscribe and listen to Tokenised Human wherever you get your podcasts.

Spotify

Apple

Pocket Casts

I don’t know about you, but I’ve been elbows deep in escalating doomerism, after the latest warning from an AI company defector set off a new flurry of “they’ll kill us all!” handwringing.

Outgoing Anthropic researcher Jacob Coxon told anyone who would listen that OpenAI and Anthropic were “racing straight to self-improving superintelligence and gambling with our lives.”

But you might have noticed very few of the “OMG they’re coming for us” narratives properly explained exactly how they would do so.

And that’s important, because if we don’t understand the actual risk scenarios, it’s pretty hard to take them seriously — let alone agree on some actual, proactive prevention.

All the ways people imagine this would come about are pretty sci-fi. That’s kind of the problem with tech-apocalypse. We’ve consumed so much pop culture about it at this point it sounds silly to suggest it could really happen. But as you’ve probably read this week, some serious people rate the robotic end of days as a non-zero possibility.

So without further ado, I present The Tokenised Human guide to robo-apocalypse lore:

That’s your basic rundown of the leading end of days scenarios. But the harder question is what these grim possibilities reveal about power and control.

That is one of the many topics I spoke about with UQ’s associate professor Michael Noetel in the first episode of the Tokenised Human podcast (it launches tomorrow, September 18!).

“We did a study recently trying to get AI experts to prioritise the risks from AI,” he tells me. “And some of them were things that you and I probably heard about a lot. Like, maybe AI could do big cyber attacks or create biological weapons. But one of the top things that the experts flagged was power concentration, which is where one person in the Broligarchy could effectively be a dictator. Actually every AI company was started because they didn’t want the other guy to be the dictator.”

Delightful.

Anyway, I am really pumped for you fine folks to hear the pod.

The first season is six episodes, in your feed on Friday mornings. You can listen inside the Substack app — if you’re a subscriber you’ll get an extra Friday email showing you where to go — or subscribe now on Spotify, Apple, or wherever else you prefer to listen.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

So what happens when AI starts changing the jobs that shape our lives?

In season one, I talk with psychologists, researchers, technology leaders and people redesigning the workplace about the job apocalypse, capitalism, surveillance, loneliness, power and who gets to decide what happens next.

No techno-utopianism. No doomerist cosplay. Just clear-eyed conversations, uncomfortable questions, a few jokes and a big dose of joyful pessimism.

The first episode lands 18 September, with new episodes every Friday.

Follow, subscribe and listen to Tokenised Human wherever you get your podcasts.

Spotify

Apple

Pocket Casts

• Is Facebook remembering your friends’ and family’s birthdays, so you don’t have to, a problem?

• How about if it automatically sent those friends a “Happy Birthday” message?

• What if the reminder came with a link to gifts that person might like populated by that person’s search history, recent purchases and other cookies?

• How about if it was linked to your credit card, had an agreed budget and just purchased its best guess at what they would like and had it sent to them with a generated birthday message?

At its core, this is an exercise in considering at what point outsourcing your relationships to a machine feels like a degradation of the relationship itself. And even though it’s been almost a decade since I started posing that question to my students, I’m not sure we have gotten any closer to a universally agreed answer – even as the proposition has become more and more aligned with reality.

A friend told me only this week about how a bunch of his colleagues were using the AI “suggestion” button to fill out a digital leaving card for a departing teammate.

(In my opinion you either default to something generic – “You will be missed!” / “Congratulations on the new role!” – or you think of the thing you most like about that person and make it about that – “Your Friday muffins will remain company legend!” or “I always appreciated how much care you brought to your work.” To me, outsourcing it to a bot feels worse than skipping the card altogether.)

And there were two other reasons I started thinking about it again this week.

First, I clocked this article on Business Insider on “the modern parenting throuple” – the idea of adding an AI “third” to your marriage and parenting decisions. (I’d argue a digital polycule is probably more accurate – what modern couple only has one AI between them?)

In that article, one parent said she had spent the previous year travelling for work, so she built a custom GPT of herself for her 15-year-old son to consult when she was unavailable. She programmed it with her personality, screenshots of their texts, her family values and rules, and the parenting challenges she was dealing with.

(Yes, I am aware I am being kinda judgemental and shady about that. I want to be clear I would also be just as shady if it was a dad doing it – this is not a gender issue.)

The second thing was this gadget gathering steam with techbros and normies alike called the Brick. It’s a physical tile and app combo that you use to limit your screen time. You tap your phone on it and it “bricks” you. It will stop you opening certain apps, etc – you decide the rules – so you can’t keep doomscrolling or whatever.

So we invented a gadget that replaces… turning off your phone? Putting it in another room? Deleting your social apps? Exercising self-control? Yes. Apparently that is what the Brick is for.

There's a hole in my life, dear Liza

Those two things may feel like different phenomena, but hear me out.

The through-line is something researchers call “techno-fixing”. That is, using a technology to patch a problem, rather than repairing or changing the underlying system.

The point is not that the technology caused every problem. It’s that it lets us preserve a broken or failing system that produced the problem instead of changing it.

And the ways we choose to techno-fix and patch the problems in our own lives tell on us a little – or a lot.

No hack for un-burning a match

Let’s say you’re in a permanent state of cognitive overload due to … I don’t know, the escalating cost-of-living pressure, general global uncertainty, and overwork. Perhaps you recently lived through a global pandemic and a bunch of lockdowns. Maybe you’re either in or edging on burnout. And that’s leading you to withdraw.

A Beyond Blue survey last year put the number of Australians facing workplace burnout at 49 per cent. And the one that always shocks me: a 2024 KPMG study found rates of permanent disability from mental health conditions in Australia, specifically among Millennials, are up 732 per cent since 2014. That’s an insane figure.

And OK, let’s assume you’re not quite at the total permanent disability threshold just yet. You’re just medium miserable. Great. When we’re burned out we have a tendency to lose interest in doing … stuff. Just stuff in general. So we veg out in front of the TV. We cancel on our friends. We doomscroll.

Rather than question the wisdom or sustainability of that system, let’s throw some bandaids on the problem.

At the same time, productivity-maxxing and career optimisation culture tells us we can do all the things if only we work hard enough, eat and sleep like a machine, and recruit the right combination of gig workers and fancy new tools.

Of course we could, instead, sit down at the dinner table of life and decide which course we would actually like to focus on digesting right now. We could pare back on some things so we have the bandwidth to properly focus on that area. And we could build supportive proximate communities composed of family and friends and people we know who help each other out. You know, community.

The truth is you can have many things in life, but not all things. You’ll probably have to work really bloody hard for the things you do have, a bunch of it will come down to dumb luck, and you most likely can’t have even those things at the same time. Life is full of seasons. Knowing which one you are in right now will save you a lot of grief.

But instead we, what? Develop another app? “It’s called AtaBOI and it’s, like, Uber for your dad’s approval,” says a completely deadpan 22-year-old Stanford dropout while he records you on his Meta glasses.

Robo-Mommy gives cold hugs

Let’s consider that AI parenting “augmentation” situation, for example. There is a difference between using a tool to help with the logistics of parenting and using it to stand in for the relationship.

Look, parenting is bloody hard. And frequently tedious or annoying or just not what you want to be doing at 10.30pm on a Tuesday. I have a teenager myself. I fully understand that thing where you think to yourself, “Phew! We made it! They are going to be more independent now, right?” Only to realise, no, teens need you almost as much as a toddler does.

And I know we want to support people to do the jobs they are passionate about and take all the professional opportunities they want to, but that is not always possible. Sometimes we have to make personal sacrifices for our families, or our health, or to housebreak a puppy, or because our best friend is going through a really difficult time right now.

It’s not a sexy, fun or particularly comfortable reality, but it’s the truth.

I was speaking to the Australian futurist Aaron McEwan on Wednesday about a new book he is writing on surviving the AI era with your humanity intact.

In researching that book he found evidence that attending live music was “substantially better for your well-being than just listening to it on a recording”. It’s because you’re in a community, but also because of “the effort of going to see it”. The effort matters. Choosing to spend your limited bandwidth on that activity is a declaration to the world and to yourself as to what you most value.

Ultimately, I think it comes down to your true internal compass and the lynchpins on which your specific ego is built.

If I outsource my parenting and use the AI as a proxy for interacting with my own kid, I do not get the benefit of that effort. Worst case: my child bonds to the AI, trusts the AI, turns to the AI instead of me. I am ultimately poorer for it.

I can’t send you to the gym and expect to get fitter myself.

The areas of our lives that we choose to patch do ultimately tell on us. The question is whether, in trying to preserve every single opportunity and solve every problem with a tool, we are accidentally deprioritising the things that are most important.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

But in reality my analysis of the ways this technology is being spoken about and acted on in the workforce is a lot more sympathetic to the powers that be than you might imagine.

Because if you’re a CEO – particularly of a listed company – and genuinely believe sacking a whole bunch of staff is in the company’s best interests and would deliver better shareholder value, it could be a breach of your duties not to do it, right?

And once one company makes that decision, we can pretty quickly find ourselves inside the Spiderman-pointing-at-Spiderman-pointing-at-Spiderman meme. The boss refusing to say they have found a similar way to cut headcount can end up looking like a failure when really he’s just the guy being more cautious and potentially comprehending the limitations of the technology more clearly than the others.

You could end up punished for your prudence.

Shadow puppets of efficiency

This is to say that the pressures shaping hiring and firing decisions during this period of workforce transformation are not always based on what the technology can legitimately deliver, but rather on some kabuki theatre for investors.

Take the investigation published last week by Reuters, which found Facebook daddy Mark Zuckerberg and his leadership team had been considering a radical AI-worker replacement plan that would have reduced some teams by up to 60 per cent, through layoffs, redeployments and role closures. Ultimately, Meta settled on a more conservative 10 per cent total headcount reduction – still a lot of people.

Internal figures, obtained by Reuters, suggested the company knew its AI tools and agents were not producing the productivity gains expected but rather than dial back the strategy that depended on those tools, it kept throwing money at it, at the same time using “capital expenditure” to justify the layoffs it did actually make.

In fact, one eight-month study found that workers using AI took on broader tasks, worked at a faster pace and for longer hours. So basically, sometimes AI means you do more work, not less.

It didn’t matter that the product couldn’t replace those workers, they lost their jobs anyway.

The question, then, is not only whether the tools work. It is who gets to decide what counts as failure, what any perceived failure means and, importantly, who suffers the consequences. The people who back a bullish AI strategy could treat disappointing productivity outcomes as a reason to slow down, reinvest in people instead, or reconsider whether their initial level of investment was the right call. Instead, the investment itself became part of the argument for cutting staff.

Once that happens, abandoning the strategy is no longer just a technical decision. It could be felt as an admission that the people who made the bet were wrong.

I, of course, never make mistakes

That thought reminded me of a series of conversations I had when I was researching the personality traits most closely linked with success. The British organisational psychologist Adrian Furnham and I discussed the Peter Principle – the idea that in a hierarchy, people are generally promoted to the level of their own incompetence – and he observed that the higher up the corporate food chain a person rises the less capable they become of seeing their own blind spots or understanding the limitations of their abilities.

Laurence Peter (for whom the principle is, of course, named) was writing about this in the 1960s, Furnham marvelled. “But, my God, how prescient it was. It hasn’t aged at all. Fifty years ago this man was saying things which are incredibly salient for today.”

And when there’s ego involved – which, let’s be real, there often is in upper management – it gets even hairier. What happens when the person with the power to say “this is not working” is also the person who has staked their reputation as a bold and innovative leader on its success. The boss may be responding to real pressure from investors, but he is also deciding which evidence counts as a reason to pivot … or stop.

And technology can make that inability to see the wisdom of your own decision making clearly even more difficult. Because it’s complicated and you’re unlikely to have a perfect understanding of its risks and limitations? Yes. But in this case it's even weirder, because the technology itself will, sycophantically, lead you to believe that you do, in fact, understand it.

I’m sure you have heard of the sunk cost fallacy: that is, I have already spent all these resources on the thing, I can’t pull out now, so I keep throwing good money after bad and compound the failure. And research has repeatedly shown this doesn’t just happen on the individual level; organisations are also vulnerable to it.

There is also a measurable phenomenon in decision-making where, once a project starts to fail, people can be more likely to persist when continuing is framed as action and stopping as inaction.

There is also evidence that people can keep backing a failing project because continuing feels like taking action, while stopping feels like standing by and letting it fail. And we have a bias towards action.

What even is winning, anyway?

And now we’re stuck in a corporate doom-loop. The company spends heavily on AI, then needs to show that the spending is producing profit outcomes. That means new revenue on the same headcount or, more often, a cut. Redundancies are presented as evidence that the strategy is working; reversing course later means admitting that the company either used workers as a smokescreen for a bad investment or failed to recognise it was bad in the first place.

And when you consider all that, it’s pretty easy to see how it might not matter anymore whether AI can really do your work or whether something important like judgement might be degraded in the replacement.

If the powers that be are convinced it can, and have made significant investment in the tools and in overhauling the organisation to accommodate it, then the path might be fixed regardless.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

Over the past year or so I have been contacted by a whole bunch of parents, career counsellors and teachers who wanted to know what they should be telling their kids about preparing for an AI-affected workforce. The question sounds simple: what should my kid do now?

They were writing because at the end of last year I wrote a series on whether a university degree was still a valuable investment for a kid preparing for the modern workforce and later this piece where University of Sydney Business School’s excellent Prof Clinton Free crunched the numbers on Anthropic’s most dire workforce predictions, which went kind of viral.

But despite all the research I have done, all the very smart people I have spoken to, I didn’t have a great easy answer for any of them.

Because, uncomfortably, I have come to believe the question is too small and assumes that the answer is a tweak, rather than an overhaul. I’d suggest the real question we need to be asking is: what is the point of school at all?

So you can see, I’ve been mulling this question for a long time, but it resurfaced this week as I scrolled through all the posts on social media responding to a new preprint published by researchers out of Stockholm University and the University of Hong Kong in June.

In case you missed it, the study looked at the learning outcomes of more than 26,000 Chinese secondary school students over nearly three years, trying to ascertain how AI use was impacting their grades.

It found AI-use lifted homework scores an average of 18 per cent, assignments were completed faster, but closed-book exam results fell 20 per cent. Basically, the kids could do their homework faster and better, but they didn’t retain the information like they would have before AI. They didn’t really learn it as well.

And that has thrown another log on the fire of moral panic around children and AI, whether they’re using it to cheat on their school work and what it means if they are.

In Australia this week, the ABC reported on a range of different approaches under development by the country’s various state governments. New South Wales is considering doing away with take-home assignments for senior students entirely while South Australia has tried to wrap guardrails around AI-use by developing its own learning assistant chatbot. Both responses make sense on their own terms. Schools do have to deal with assessment integrity. But they also show the limit of the current debate: we are asking how to preserve the existing model of schooling under AI, not whether that model still makes sense in the first place.

More right answers, more wrong people

That study made me think about the February paper out of Wharton, proposing the idea of “system 3 cognition” – this paper is seared in my mind, it made me laugh so hard.

In it, the authors seek to build on the cognition model made famous by Nobel Prize-winning behavioural psychologist Daniel Kahneman, in which system 1 thinking is essentially defined as intuition, and system 2 as deliberation. I like to think of it as system 1, animal brain; system 2, higher-order reason.

Anyway, in the Wharton paper, Drs Steven Shaw and Gideon Nave suggest “system 3 cognition” could be defined as “artificial cognition that operates outside the brain”.

They found that people who used accurate AI could massively improve their accuracy when completing a task, but when users consulted an inaccurate AI, they let the AI’s bad info override their own judgement more than 73 per cent of the time – the paper calls that “cognitive surrender” – and in those cases they under-performed what they would have achieved without any AI at all.

They also found time-pressure made cognitive surrender more likely.

What gave me the giggles was the idea of “artificial cognition that operates out of the brain”. Because to me that is just … not doing the work? And that is not some special AI thing, that’s just a thinking thing. If I have an assignment and I get my roommate to do it for me, that is also cognition that is operating outside my brain. No tech required!

Basically, “system 3 cognition” is just a fancy sciencey-sounding word for not doing the work yourself. You wouldn’t believe the things I don’t know because I never thought about or learned them! (Hahahahaha ok, glad I got that out of my system.)

But cognitive surrender is real, and it echoes in what the Chinese study found. Those kids weren’t taking bad answers off a bad machine, they were taking good answers off a good one. In both the Wharton and Chinese student studies, when the AI was “good” the work came out better and the person didn’t.

The issue is not simply that AI helped students produce better answers. The teachers already knew the answers. The repositories of human knowledge already held the answers. We did not need a 13-year-old to invent the answer. We wanted the 13-year-old to prove they had learned it, because them knowing the answer ... did what?

The pedagogical answer is that knowing was supposed to change what the child could do or learn next. The point of practice, testing and repetition was never merely to extract the answer, it was to build recall, fluency and usable understanding, which is why research on retrieval practice and effective learning techniques treats practice and testing as part of learning, not just measurement. If AI can produce the answer without the child doing that internal brain work, the grade may still signal success, even though the education may not have happened at all.

We realise the accuracy of the answers was never the thing we cared about. (I’d argue this is true regarding a lot of knowledge work being threatened right now as well.)

And again, it’s an issue highlighted as we grapple with AI, not a new problem created by AI. We routinely see people get so caught up in the AI of it all they fail to see the ways that this technology is just exacerbating preexisting structural issues and uncovering fundamental flaws in systems that long predate AI.

I used to remember the phone numbers of every one of my friends off the top of my head. I could not tell you my own work number these days without checking my email signature. We have been doing this for years.

And I’m not even sure that it’s wrong.

Your kid is right: times tables suck

If you have a kid in school I bet you have (or will) field some version of this question:

Why do I need to learn my times tables? I have a calculator on my phone.

And you will probably say something like “it’s useful” or “it will make you faster at other math later”. But even though I have trotted this explanation out myself, I know it’s pretty weak. We are always adding and adding to school curricula and never taking anything away. And the exact utility of some of the things we keep making kids learn seems hard to pin down.

Because your kid is right. They do have a calculator in their pocket at all times now. Surely learning to use it well and accurately would be the better use of time and resources than memorising their 7 times table? (I still hate the 7s, I’m sure you understand.)

The difference between remembering your BFFs digits and preserving your critical thinking might seem clear – recall and judgement are separate categories. The difference between remembering that phone number and your times tables might be less. We don’t have a good sense of what things are ok to outsource and what might have a long-term negative impact on your cognitive abilities.

The question shouldn’t be whether we let kids offload, it’s what we let them offload and why. And to answer that question we need to go back to basics and work out the purpose of school.

Job ready for what?

Is it to retain facts, become “job ready” as Australian education policy explicitly states? Or is it to become a thoughtful well-rounded person?

We have spent so many years trying to compete with international math and literacy test scores like some intellectual arms race, but why? In a world where a lot of the formulaic and calculative work will probably be done by a machine, is knowing the important thing?

The last few years all I hear is “judgement and critical thinking are what we care about” but how is our current education system geared towards developing those skills? It is not.

The current schooling system still carries the imprint of the industrial revolution, with its emphasis on time, sequencing, compliance, standardisation and measurable output.

If even the university system now says education needs to develop “evaluative judgement, critical thinking and ethical reasoning”, is tinkering at the edges of a system never designed to deliver such nebulous, individualised outcomes the right call?

And some people are trying to ask these more fundamental questions. Sydney University’s Policy Lab recently published a program called A Good Childhood, which argues for the re-centring of a child’s interests. “Australia needs to think differently about childhood,” program literature states, “not simply as a stage of preparation for adulthood”.

Its starting point is that policy should ask what children need to thrive, and how institutions can support their wellbeing, belonging and opportunity. That is a very different starting point from the Pavlovian response mechanism of the current school system: bell rings, stop; bell rings, move; exam time, regurgitate.

“Too often children’s interests are an afterthought in policymaking,” said lead researcher Prof Luara Ferracioli. “And this is one reason many Australian children are not reaching their potential or fully enjoying their childhoods”. Oof.

That is a very different starting point from asking how to preserve the integrity of homework.

Fellow substacker The Pavement Philosopher is a high school teacher at a selective school in Perth. (For non-Aussies, that’s a government-funded school for high-achieving kids.) He’d rather I not share which school or his real name, because he’s not sure how his employer would feel about him talking openly about this stuff, so I’ll refer to him as Phil.

When I asked him how his students were thinking about AI, he stressed his school was “not a fair representation of all teenagers”. But he says the “freakishly smart and gifted” students were often “the most suspicious of AI, least likely to use it and most likely to call out teachers who are obviously using it too much”.

I would argue these are the kids who were always going to have the best natural aptitude for critical thinking and judgement. They exhibit this type of thinking despite our education system, not because of it. We’re not currently focused on teaching this stuff to the typical student body and the current school day is not built to do so.

Children are, famously, non-fungible

The other dynamic Phil sees on the regular is an extreme level of parental investment. Some students, he says, have parents who spent “A LOT of money” (his emphasis, not mine) on tutoring to get them into the school in the first place.

When I asked whether AI had broken the bargain those parents thought they were buying into – i.e. best school = best uni = best job = securest future – he pushed back. “I don’t know if AI has really made that bargain anymore broken than it was, mostly because it was so broken to begin with.

“To put it in Kantian terms, these are parents just using their kids as means-to-an-end – of prestige, wealth, whatever – rather than respecting their autonomy as human beings.”

He reckons some parents see their children more like “investments” than people, “which, to me, is already as broken as it gets”.

Look, that is a rough thing to read; I really hope that’s the exception rather than the rule. I think a lot of parents do the same things because they are stressed out, do love their kids and really want them to have a stable future. And they just don’t know what else to do.

And that is where I think this all lands. AI has not made school meaningless. It is forcing us to ask the thorny question of what meaning we were projecting onto it in the first place.

In developed economies where the factory worker has all but disappeared, instead of re-engineering our school system to fit the demands of a services-based workforce, we have just iterated and iterated on the factory-era school formula into absurdity.

Critical thinking and judgement are not a gap you close by bolting another module onto the existing, overladen curriculum. Teaching them to a typical student body would mean inventing new ways of doing it, explicitly, and building schools, maybe from scratch, around that fundamental goal.

None of that is going to happen overnight.

What do you do?

So when parents ask me what they should tell their kids to do now, I deeply understand the instinct. They want an answer that sounds like a plan: take computer science, become a doctor, choose the safe degree, here’s how to become impossible to replace.

All we want is certainty that our kids will be ok.

But if I am being honest, I don’t believe there is a simple, playbook answer anymore.

My honest answer is much more abstract and harder to tie up in a neat easy response. But, here, I will try:

Argue with them and teach them how to argue back to improve their cognitive flexibility, to understand persuasion and judgement. Show them how to care about stuff. Urge them to make things. Get to know them and help them to know themselves. Encourage them to pursue their real interests. Most importantly, teach them to value themselves as human beings rather than as a worker.

When it comes down to it, no work is truly “safe” from automation. I think the people who are going to fare best in the decades to come are the ones who believe, deep down, that they are more than a job. Even as jobs come and go, your identity can remain stable.

Which, I know, is a much more radical and uncomfortable answer than “just learn to code”.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

Any shrink worth their SSRIs will tell you competence—feeling like you have the tools needed to perform the tasks in front of you—has long been understood as a basic psychological need. Feeling like we know what to do, and can do it well, is a really important component of feeling safe and helping us to self-regulate. It’s calming. People really like it.

So I’ve been thinking a lot about how perpetual upskilling has been billed as the “new normal” in most knowledge work over the last couple of years. It seems like that instinct is correct—I don’t think the pace of change is really going to slow down in this era—but the impact of that change, and the potential inequities and overall misery it is introducing, are not really being grappled with much at all.

A moving (maybe lucrative?) target

Jobs and Skills Australia (JSA) did have a go at one piece of the puzzle. In a report published last year they looked at the correlation between years of experience, skill change and salaries. That analysis found that jobs with a lot of skill change tended to also have more salary growth potential over a lot longer period than jobs where the skillset roughly stays the same.

That makes a kind of sense, if you think about it.

If my job was something I learn once and then practice for years in the same way to become an expert at, my pay would hit its peak earlier—basically as soon as I reached that expert level. But if my job required constant regular upskilling, well, that’s a moving ceiling. (I’d argue it would have also meant a bigger variation of the skills acquired across my peer group, depending on who’s kept up and who hasn’t.)

And, typically, the JSA data shows, keeping up with those moving skills has been compensated. Jobs with a moving skill target take around 20 years of experience to top out.

It has also meant a relentless race to stay “competent”. Constant upskilling shakes your feeling of competence, because the thing you build that feeling on stops being what you know and becomes how fast you can learn. And that race has already, historically, disadvantaged people who took any amount of time out of the workforce—say, to have a baby, for example.

Stale skills might be a myth, the damage is real

We know the motherhood penalty is real. Treasury researchers who analysed government data in 2023 found a 55 per cent reduction in average earnings of mothers over the first five years following the birth of a baby. Analysis attributes that mainly to reduced working hours, rather than lower hourly wages.

And skill depreciation from time out of the workforce is also real. Although how real and what kinds of impacts it has is kind of hard to nail down.

But ask any professional woman who has taken even six months off for baby care. Plenty of them are openly worried their skills have gone stale. I’ll be honest here, I read many very dry, difficult papers trying to establish whether this was actually true or not. I still don’t know.

The data seems to show some skill depreciation if you take a year off work, but that you earn it back and then some within a year of coming back. But other studies show the actual skill loss is very minimal, so when a woman expresses concern about this, we’re going off vibes here. Which is its own competence problem: not knowing whether you’re still good at your job does the damage whether or not your skills actually went anywhere.

And actually, that’s kind of the point.

What’s the grading curve?

What studies do show us is that women are punished more harshly for misconduct, judged more for mistakes—and, if one UK experiment is right, for using AI, for that matter. Plus they know it, and they factor that into how enthusiastically they are approaching AI upskilling.

So competence isn’t only something you just feel internally, it’s something other people decide you have—or don’t—and that probably shapes your prospects even more than your own sense of ability.

This paper from PNAS Nexus found women are more skeptical about the alleged benefits of AI at work—partly because they’re more exposed to the jobs it’s coming for. And when the researchers told people AI was likely to destroy more jobs than it created, women’s support fell away faster than men’s. (This is actually just a really great paper all round, five stars, would recommend.)

That’s not to say they use it less. A recent Roy Morgan report found the gender split on AI-use is now about equal. The dudes just trust it more, which, as a woman, I think is stupid. But then I would. Because, again, woman.

Fast-growing … for who?

That skepticism may go some way to explain the widening gender gap in fast-growing AI careers. New research published yesterday by LinkedIn showed women hold 49.7 per cent of non-AI jobs at non-AI companies in Australia, but only 27.6 per cent of AI jobs inside AI firms. In leadership it goes from 28.1 per cent at non-AI companies to 15 per cent at AI ones.

Basically, women are equal participants in the general workforce, not so much in AI jobs, which, LinkedIn points out are “some of the fastest-growing and highest-value opportunities in the labour market.”

When I asked LinkedIn’s head of public policy for ANZ Audrey Lobo-Pulo why exactly they think this is happening, she told me there wasn’t a single explanation but more a combination of factors at play: “Growth is concentrated in research and technical engineering roles, which have historically been male-dominated,” she said. “At the same time, promotion pathways often continue to reward linear and uninterrupted career progression, while women are more likely to experience career breaks.”

And there we have it: our uninterrupted career rearing its head again. “The question,” Lobo-Pulo went on to explain, “is how can employers, educators and workforce systems help ensure more people have opportunities to participate and progress?”

And I mean, great question. Hard to answer, though, when nobody has properly counted what employers spend on training their staff since 2002. That was the subject of a report this past week by JSA, arguing the Australian Bureau of Statistics really ought to start keeping track again.

According to this report Australian employers spent around AU$2.90 per $100 of wages paid on training staff in 1993. By 2002, the last year we captured that figure, it was down to AU$1.30 per $100. In 2025, JSA estimates the spend was probably more like AU$1.10. Just for comparison, they estimate that number was more like £3.80 per £100 in the UK in 2024.

“Employers play a major role in helping workers build, apply and transfer their skills throughout their careers. Workers increasingly need to update and deepen their skills as technological, demographic and structural changes reshape Australia’s labour market dynamics,” said acting JSA commissioner, Megan Lilly when announcing the report. Separately, JSA’s modelling has also shown women were “marginally more exposed” to jobs likely to be impacted by AI augmentation and automation.

So, women aren’t getting into AI jobs at the same rates as the blokes. They’re more skittish about AI and its impacts. And their current jobs are more vulnerable to AI replacement or disruption.

Bro, do you even lift?

Everyone seems to agree training and upskilling matter a lot—and not just in the jobs where keeping up was always part of the deal. It’s spreading, as more and more kinds of work get reshaped by AI. But we’re not really sure what people are being trained in, the quality or transferability of that training, who’s paying for it or when and where it is occurring.

Government stats from April this year suggest rates of formal work-related training are actually falling. But a report from the employer lobby Ai Group last November found participation in training was “above the long-term average”, claiming 86 per cent of it was being funded on the company dime, and around 9 in 10 people were doing it inside their paid work time. (Though about 30 per cent were also doing some of it in their own time—a figure that hasn’t budged since 2007.)

Both could be true. Perhaps people are getting more training on the job so they’re enrolling in fewer programs themselves. We don’t actually know.

What we do know, at least according to researchers at Anthropic, is that the current programs being run are not super effective at boosting outcomes for employees. Their review of the evidence, published last week, found small impacts overall—and that the large-scale programs (the ones run at the size you’d actually need) have been “lackluster” (their word, not mine). They believe this strategy would “likely fall short” against AI-scale displacement.

And I’m going to suggest that brings us back to the therapist’s couch.

What is this? A school for ANTS?! You can’t make someone to learn good.

Ryan and Deci described competence as a basic psychological need, sure. But in that same paper they were also pretty blunt about what happens when learning is imposed instead of chosen. Threats, deadlines, directives, pressured evaluations, imposed goals: all of them reliably flatten motivation.

People taught that way don’t just lose enthusiasm, they learn less well. Being forced to train or die on the proverbial office vine is the worst possible setup. And the learning that suffers most is the conceptual, creative kind—which is precisely what a fast-changing human-skills-first job apparently demands.

If more jobs are moving into the upskill forever camp—and that’s going to be thrust on you whether you like it or not—then we’re staring into a possible future of generally lower feelings of competence, and perpetual forced training which is almost guaranteed not to produce the best learning. That sounds like misery to me.

The people who still get to feel competent will be the ones who can move that feeling off what they know and onto how fast they can learn. Which is a bigger change than you might think. Of course, if you already wake up every morning thinking, “Gee, I’m so good at learning new stuff!”—well, lucky you.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

I’ve been thinking about that this week in the wake of the Nara Smith controversy, of all things.

In case you’re not spending your limited attention span on influencer culture the TLDR is Smith—one of the most high-profile so-called “tradwives”—went on the very popular Call Her Daddy podcast and acted all surprised and offended that people had ever tarred her with the tradwife moniker.

Much internet scoffing ensued. Smith has spent many years perfecting her ASMR soft-spoken housewife brand, making Froot Loops from scratch in a full face of makeup and cutesy apron in her perfect kitchen.

Now, you might be asking yourself, “OK, Rach, but what does TradWife™ have to do with freaking AI girlfriends?”

So glad you asked. They both speak explicitly to a growing fetishisation of care.

The illusion of effortless care

A 2025 report by researchers at the Global Institute for Women’s Leadership at King’s College London (KCL) found that lots of women—I’m talking 79 per cent in the test group—said they were attracted to the fantasy of “the calm, relaxed lifestyles” of trads; 62 per cent were captivated by their depictions of luxury.

It’s striking that a life devoted to endless service and chores is being consumed as “relaxation”—so long as that intense labour is done aesthetically, surrounded by the latest high-end appliances.

And the KCL study found the way young women felt about trad was nothing about getting retrograde or some conservative gender fantasy. Only 7 per cent of those same respondents thought blokes making all the decisions sounded hot. About a third liked the idea of a man being the main breadwinner and 44 per cent found “the idea of not worrying about a career appealing”.

That shifting attitude about work is a growing trend with dudes too. The research found 20 per cent of women and 21 per cent of men aged 18-34 perceived paid work as having a negative impact on their mental health—up from under 10 per cent who felt that way before 2020.

“Rather than a nostalgia around the role of women being limited to household labour,” the authors suggest, “the popularity of the tradwife phenomenon may be a symptom of the systemic strain experienced by the younger population, especially parents, who are juggling increasingly demanding jobs with intensive parenting responsibilities.”

It would seem the thing they’re really coveting is the appearance of abundance: abundant finances, skill, attention and, most importantly, time. Time is the greatest modern luxury.

And crucially, the appearance of choice in the matter.

But the trouble with selling women a fantasy of not working is that eventually they notice it is work—and that you’re at work while you sell it.

When tradwife is exposed as work

I spoke to Lauren Meisner this week, founder of the Gen Z media network Centennial World and host of the Infinite Scroll podcast.

She says young women are really waking up to the “grift” inherent in portrayals of the tradwife life. They know it’s not real and that women like Smith are literally working when they make this content.

“[Tradwives’] comments are flooded with criticism and mockery,” she observes. “I think a lot of young people have started hate-consuming trad content because they feel like leaving a comment telling a woman she is setting us back feels like a way to take out their anger and take back some control.”

Basically, it started out as a fantasy of peace, luxury and choice, but has become yet another high bar that women feel they are failing to reach on the daily. “I think a lot of women are like nevermind, let’s actually just go back to being sugar babies,” laughs Meisner (darkly).

The spotless floor problem

And this intensification of expectations is where the AI companion piece comes in, but also a broader allegory about how technology has, historically, raised expectations in all aspects of life. Rarely is new tech used to provide genuine relief from labour.

In 1983 historian Ruth Schwartz Cowan published a seminal study on the impact of new technology on household labour called More Work for Mother. (I looked for it everywhere and I couldn’t find a publicly accessible version, so I will instead link this related article Cowan wrote for American Heritage magazine in ‘87. It gives you the broad strokes, but if you like a narrative history I highly recommend you hunt down a copy of the book—it’s a fun read.)

Cowan’s research argues the introduction of whitegoods like vacuum cleaners and microwave ovens, which were marketed at the time as providing relief from chores, in fact just led to a new higher standard of expected domestic perfection. This is classic lifestyle creep—a ratcheting expectation as to what constitutes a “good life”.

I bought a robot vacuum a few years ago. I’ve become addicted. I now find it intolerable that my floors are ever anything but perfectly spotless.

And it’s not confined to the domestic sphere by any means. We’ve been in a perpetual state of work intensification for a really long time. Do you think industrialisation reduced the intensity of labour? Lol. Of course not. (See evidence here.)

One 2021 analysis published in the journal Industrial and Labor Relations Review said the particular squeeze over the first two decades of the 21st century has been thanks to some combo of reduced worker power, the constant need to learn and keep up with new technology, and bosses using that technology to organise, monitor and squeeze more work into every day.

Lifestyle creep of the mind

The same thing is happening right now with AI.

In a report published by the consultancy firm BCG in June, more than two-thirds of workers said AI had taken over simpler tasks, leaving them with the more complex work. Your boss will crow about this and call it a great thing—outsource the boring busy work to the droids!—but it also means more of your day is being dedicated to the hardest tasks with no easy filler tasks as relief.

Add to that: at least half the respondents also said they were spending more time reviewing and correcting AI output or managing and directing AI; 41 per cent said AI had increased the time they spent making decisions.

In the AI era I’d argue we’re entering a phase of lifestyle creep of the mind.

That is, a change in the kinds of work we choose to do—or are encouraged or even permitted to do inside organisations that are incentivised by speed and volume rather than the pleasantness of a worker’s day or the health of his individual intellect.

A lot of us are pretty worried about the potential erosion of our critical thinking and judgement if we become too dependent on AI.

When human contact becomes a luxury

It is also true of romantic and social relationships that can be in some fashion now supplemented or replaced whole cloth by an AI companion who never gets in a mood or disagrees or says no.

Which brings me back to where I started.

A team out of Oxford and Stanford published a preprint this year on a bunch of studies they ran getting thousands of participants to use an AI for emotional support.

They found, after three weeks hooked up to an intentionally agreeable AI, people were about as likely to rely on it as they were on their closest friends and family. Given a choice they picked the more sycophantic version of the bot too because they just liked the way it made them feel. Participants also reported being less satisfied with the real conversations in their lives.

That’s … not great.

It’s such a big potential problem that the Chinese government introduced a bunch of new rules regulating the availability and use of AI companions in July. Work AI assistants were specifically exempt. Read that how you like, but it looks to me like a government that’s fine with work intensification and not fine with us outsourcing our emotional needs. That’s a design choice.

So will AI give us back enough time and capacity to properly resource the parts of our lives that matter most? Or will so-called “efficiency gains” ratchet expectations to a new, higher standard of what a competent worker, good partner or functional human being is expected to be?

History suggests it will be the second.

I know I can’t go back to a dusty floor. We may not be able to go back to being listened to imperfectly by our fellow imperfect human loved ones either.

Maybe that is the real danger of abundance: once something becomes cheap and endlessly available, scarcity starts to feel like failure.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

On July 21, Substack went into a bit of a meltdown after the platform turned on AI detection software Pangram — much to the displeasure of many of the platform’s writers. Then on July 30, LinkedIn announced it was going to start rolling out a “seems like AI slop” button, and replaced its “enhance your post” AI assistant with a simple proofreader.

And of course all this week beloved science communicator Hank Green was kind of self-immolating after admitting he used ChatGPT to prep his episodes and suggesting his relationship to AI was “not healthy”.

You promised me prompts, Rachael. Where the hell are they?
(Scroll to the bottom of this newsletter or if you can’t wait, click here.)

Why do we like things more when they're hard … for other people?

And look, for a while now the data has shown people really do care about this stuff. Actually that was true well before the recent advent of The Machine — and in far more ways than might be immediately apparent.

In this 2004 study, for example, researchers found people assessed an artistic product — a poem, painting or suit of armour — as higher quality, and even said they liked it more, depending on how much time and effort they thought it took to produce.

And when others do things the hard way, we consider that not only higher quality but morally superior. A paper published in the journal Scientific Reports in April showed test subjects thought negatively of people who were described as having lost weight using a GLP-1 compared with people who lost weight the old-fashioned way. Even weight loss is valued more highly and assessed as being of better moral quality if it is harder to achieve.

We like things to be hard — at least for other people.

For example, an article published in the Journal of Experimental Psychology: General in 2016 found that the level of effort involved in producing an outcome intensified how we felt about the action, good or bad. Essentially: if you mean well and put a lot of work into something, people think well of you, regardless of whether you were successful; if you have bad intentions, and put in a lot of work, people also think worse of you, regardless of outcome. Later researchers call this “effort moralisation”. And that “regardless of outcome” is an important distinction. The thing people valued wasn’t the quality of the result or output, it was the implied care inherent in the effort.

At the same time, studies consistently show that people — and other organisms — instinctively avoid hard work. In the literature it’s often called “Hull’s law of least work”: given a choice between similarly rewarding options, creatures learn to take the easier path.

Research finds you can, of course, incentivise effort expenditure in exchange for a reward, but even then people tend to make a subconscious calculation about how valuable the reward is and discount it against the level of anticipated effort. We really prefer not to do hard stuff if we can.

When we do do hard things, we later judge the outcome as more valuable. Not while we are doing it mind you, usually only after the fact. One version of this is the IKEA effect (really top-notch naming there, I bet those guys are fun): building the flatpack yourself means you come to feel the furniture is of higher value than a piece built by someone else.

And how people feel about how other people use AI is no different.

Why being honest about AI backfires

Siavosh Sahebi is a PhD candidate at Macquarie University and coauthor of this recent paper on AI and trust.

That research found a few things:

1. Both on social media and in work, people think stuff made with AI is “less trustworthy, less authentic, and less useful”; and

2. People say they wouldn’t judge AI use if it’s disclosed, but they are wrong.

“We asked participants whether they think that disclosing the use of AI is important, a significant portion said yes,” Sahebi tells me. But although the majority of participants said they would not penalise AI use if it was properly disclosed, “our findings show that they do, in fact, penalise AI use”

This presents a conundrum for AI disclosure, because we do want disclosure to take place. It’s super important for accountability and accuracy reasons. And as Sahebi points out, AI is really being shoved down our throats right now. It’s not like we can completely avoid it, even if we wanted to. So we need people to disclose their use without penalising that honest disclosure.

But exactly how we reach a reasonable level of transparency is still up for debate.

A pre-print published in January from researchers in the Netherlands ran into the disclosure dilemma Sahebi outlines hard. In a study of preferences around news a majority of readers said they would prefer a detailed explanation of how AI was used. But in reality, no disclosure or a one-liner had no impact on readership or subscriptions, whereas a detailed disclosure about AI use reduced trust in news and subs took a hit.

How accurate is Pangram's AI detection?

If you read the laments of writers on social media, you’ll see a lot of claims that Pangram is doing a lousy job. (I found Dr Sam Illingworth’s writing about it pretty persuasive.)

Meanwhile in June the journal International Journal for Educational Integrity published a paper that suggests unlike previous iterations of AI detection software, Pangram was pretty damn accurate. I will say I found this study a little methodologically challenged, given their control was 40 graduate articles written before 2019.

The issues with that, as I see it, are twofold:

First, LLMs were trained on the stylistic conventions of academic papers as well as social media and news writing — a lot of the alleged hallmarks of LLM use are things like the em dash, which, as a professional writer, you can pry out of my cold dead hands along with my millennial skinny jeans. News writing was not part of the control group. (That said, pre-review research by the American National Bureau of Economic Research released last year did test on pre-2020 news writing and found false positives were also near zero.)

And second, perhaps more importantly, we know that language is memetic.

Is AI changing the way humans communicate?

As people consume more and more AI-mediated language themselves, some AI-isms are likely to drift into their own writing. A pre-print published in July by a mix of European researchers found that “machines trained on human data now feed their own traits back into human language”.

They ran an experiment where they tried to see if an interaction with an AI chatbot could incept the participants with particular word choices. It did. Fast. Like we’re talking one short conversation. And people thought the new word choice was their own.

The paper flags the emergent risk that consistent, widespread AI use could lead to the homogenisation of language and culture over time, and the potential of a handful of AI companies to manipulate or influence thoughts and beliefs at scale.

But can it really change your mind?

We already know that LLMs are extremely persuasive; I like the term used by the University of Sydney Business School’s Doctors Sandra Peter and Kai Riemer — “seductive”.

In 2025 the American academic Elizabeth Loftus, who has spent half a century researching the creation of false memories, teamed up with researchers at MIT to see if they could use AI to do it even more effectively. Unfortunately the answer was, yes, absolutely.

The AI didn’t just help implant new memories, it made people question their own true memories.

So you might not even know that your writing or speech or even thoughts and ideas are being subtly influenced by what you read in an LLM. And we’re also bad at intuiting what was created with AI — hence the demand for the detection tools.

And it does matter, even beyond our sense of outraged value. We don’t like being manipulated, and we’re currently living in a world with substantially degraded levels of trust in the media, public institutions and our own lying eyes on social media and more generally online.

In this 2025 paper, also coauthored by Sahebi, they unpack the central challenge that the AI slopification of our information environment has wrought. The authors argue we’re left with two options:

1. Keep trusting as normal and risk getting duped and manipulated.

2. Never trust anything, and sometimes wrongly discredit people who’ve done nothing wrong.

And that’s a dark turn of events.

As Sahebi says: “If we adopt a generally distrustful disposition, then we might end up causing what we call ‘epistemic injustice’ towards people, unjustifiably. So, for example, somebody who hasn’t used AI could be wrongly accused and then that person’s credibility is completely shot.”

(“Epistemic injustice” is essentially an unfairness in the sharing or creation of knowledge. Sahebi uses it in this case to refer to unfair damage to a person’s credibility and status as an authority. My extrapolation: we’re potentially accusing someone of an effort crime or intellectual stolen valour when the reality of whether the allegation of AI use is true is not really known or knowable to us.)

Or people may have used AI in completely legitimate and fair ways, to improve their grammar as non-native speakers, for example, and are then penalised unfairly.

“It can be weaponised in that way,” he says.

The solution, it would seem, is not going to be a technological one but a cultural one. I’d argue we need a reshaping of agreed cultural norms around what “acceptable” AI use — and disclosure — looks like.

To that end: an exercise in radical transparency

I’m going to do the thing the Dutch researchers I referenced earlier said was bad. I’m going to give you the most detailed explanation of how I use AI. You tell me how you feel about it when you’ve seen it.

Here are all of the AI skills I used in the production of this newsletter and exactly how I used them. Feel free to interrogate them at your leisure — or appropriate them for your own use.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

Share

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

1. Research Mode

I mainly use Claude Pro. All these skills are my current versions as loaded in Claude with specific references to the names of my projects etc removed, so they’re fit for your personal use.

Research Mode is what I use instead of Google.

In my opinion Google Search is officially borked and you kind of have to use an LLM to surface quality results these days. Feel free to debate me on that one.

I will ask it things like: “I am writing an article about all the recent AI disclosure controversy. Find me canonical sources to construct an accurate timeline of recent developments including the Substack Pangram launch, LinkedIn changes, and the Hank Green incident.” Or: “I remember reading an article where MIT researchers used AI to create false memories. Find that.”

I always read the actual articles — you should too! It’s important. I catch LLMs mischaracterising research all. the. time.

When I am completely done, I have it fact check my entire story against its sources to make sure I haven’t mischaracterised any thing myself. If needed, I make my own changes.

Skill name: research

Description:
Use only when the user message begins with /research. Researches a reporting question using inspected, ranked sources; not for drafting articles, social trend scans, or checking supplied drafts.

Instructions:
Run this workflow only when the user's message begins exactly with `/research`. Treat the text after `/research` as the evidence question. If the message does not begin with `/research`, do not run this workflow.

# Research

Research the supplied evidence question for reporting support. Return only claims supported by inspected qualifying sources, plus material gaps or disagreements. Do not write, rewrite, pitch, fact-check a supplied draft, or scan social media trends.

## Scope

Answer the stated evidence question without broadening it.

Use public sources unless the user supplies or explicitly authorises private material. Do not search or use private files, notes, drafts, messages, transcripts, or correspondence unless the user explicitly permits that source set for the current task.

If the request asks for `current`, `latest`, `today`, `recent`, or another time-sensitive fact, verify at runtime and distinguish publication date from measurement period.

## Clarification Gate

Before searching, identify the evidence question, jurisdiction, population, period, and source-permission boundary when they affect the result.

Ask one focused question only if unresolved ambiguity could change source permission, evidence selection, conclusion, or output. Otherwise state the assumption and proceed.

## Source Hierarchy

Use the highest-ranked qualifying source, not the easiest source to retrieve. A qualifying source must be inspected, directly support the evidence question, have identifiable authorship or institutional responsibility, have a traceable evidence basis, and have an inspectable link.

1. **Tier 1:** official statistics, government publications, regulators, courts, legislation, official decisions, public institutions, peer-reviewed research, and academic preprints or working papers labelled `[not peer reviewed]`.

2. **Company sources:** use only for the company's own products, policies, filings, financial results, statements, or internal research. Label `[company source]`. Do not treat company claims about markets, competitors, social outcomes, or effectiveness as independent evidence when independent evidence is available.

3. **Tier 2:** consultancies, market-research firms, think tanks, industry associations, and pollsters. Label `[consultancy/industry - potential bias]` and state the evidenced incentive, or state which funding or method detail is unavailable.

4. **News and current-events context:** use news reporting when the evidence question concerns a recent event, public debate, media coverage, reported claims, commentary, chronology, or what is currently being said by identifiable people or institutions.

Wire services such as AAP, AP, Reuters, Bloomberg, and similar editorially accountable newswires qualify when the item is attributable to the wire service and dated. Prefer the original wire item when accessible; otherwise cite the outlet version and identify the wire attribution.

Prefer news sources that identify the publisher, author, newsroom or wire service, publication date, correction or contact mechanism, and named sources or documents. Avoid source-free commentary, unattributed aggregation, clickbait summaries, content farms, sponsored/native advertising presented as reporting, and rewrites of wire or other reporting that do not clearly attribute the original source.

If a news article references an academic study, dataset, court document, filing, official report, speech, or regulatory decision, attempt to locate and inspect that primary source. Present both when the news article adds current reporting, commentary, chronology, interviews, public-reaction context, or other reporting needed for the task. For hard empirical, legal, regulatory, or statistical claims, cite the primary source as the evidentiary support whenever accessible.

Do not cite Wikipedia, search snippets, source-free aggregators, content farms, unattributed listicles, recycled vendor statistics without an identifiable original source, AI-generated summaries, or a secondary source when an accessible original directly supports the claim.

## Source Links

Every cited source must include a clickable link to the inspected source or to a stable landing page where the user can access the inspected source.

Prefer, in order:

1. the original official page, report page, dataset page, legislation page, court page, filing page, journal page, DOI page, PubMed/PMC page, preprint page, OSF page, repository page, wire item, or article page;
2. an archived copy only when the original page is unavailable and the archive preserves the source content.

For academic papers behind a paywall, first look for accessible full text, an author manuscript, a preprint, a PubMed Central copy, a repository copy, or an OSF record.

If no accessible full text can be inspected, use a public abstract, journal landing page, PubMed record, university summary, or author summary only as a limited public summary. Label it `[public summary only - full text not inspected]`.

Use a public summary only for claims plainly visible in that summary. Do not use it for detailed methods, figures, limitations, effect sizes, causal strength, or claims not visible in the public summary.

If the public summary does not directly support the evidence question, exclude it and state the evidence gap.

Do not cite a paper, report, dataset, article, law, filing, public statement, or other source if no inspectable linked source can be located.

Do not use native citation as a substitute for a link. Native citation details may be added after the link only when useful for identification.

For supplied private material, cite only a retrievable attachment, file, or session link that the user can inspect. If the underlying material cannot be opened and inspected, exclude it.

## Citation Discipline

Links are evidence locators, not citation decoration.

Include a link only when it is the inspected source or stable source page that directly supports a sourced factual claim in the research answer or evidence bullet.

Do not add links to clarifying questions, process notes, generic caveats, methodological statements, or ordinary reasoning about how the research was done.

Do not cite model documentation, Wikipedia, encyclopedias, generic background pages, or unrelated institutional homepages as support for reporting facts unless the evidence question is specifically about that source and the linked page is the primary inspected source.

Every factual claim in the answer must be traceable to one listed evidence bullet. If no inspected linked source supports the claim, remove the claim or state it as an evidence gap.

## Method

1. Convert the request into one or more answerable evidence questions.
2. Search source hierarchy top-down: official source, dataset, law, paper, filing, or accountable news source first; secondary discovery second.
3. Open and inspect the source before citing it. For PDFs, inspect the page containing any cited figure, table, method, footnote, or limitation.
4. Use a source only if it directly supports the evidence question, matches or explicitly compares the relevant population, geography, and period, and is not prohibited.
5. Track material details: population, geography, measurement or data-collection period, method, sample, denominator, preliminary or final status, units, nominal or inflation-adjusted status, and material limitations.
6. Distinguish observations, estimates, forecasts, models, correlations, and causal evidence.
7. Return three sources by default. Return fewer if fewer qualify. Return more only when separate questions or material disagreement require it.
8. Exclude duplicate sources that repeat the same underlying evidence unless they add material original reporting or analysis.

## Ranking

Rank qualifying sources by:

1. Exact support for the evidence question.
2. Source hierarchy.
3. Methodological fit and disclosure.
4. Population, geography, and period match.
5. Latest applicable evidence for time-sensitive claims.
6. Independent evidentiary contribution.

`Direct support` means the source answers the evidence question or supports a specified material claim. Topic similarity is not enough.

`Latest applicable evidence` means the newest evidence measuring the required period. A newer publication about older data does not automatically outrank data measuring a later period.

`Material limitation` means a limitation capable of changing the answer or required qualification.

## Output

Use plain English, minimal formatting, and no emojis.

Do not use markdown headings or subheadings in the research response.

Start with a short direct answer. Keep it brief. Include only sourced findings, material disagreement, or evidence gaps. Distinguish established fact, reported claim, interpretation, and unresolved evidence.

Avoid inline links in the opening answer unless the user asks for them. Put source links in the evidence bullets.

After the direct answer, use compact bullets only when they make the result easier to scan. Each evidence bullet must include:

- the linked source title;
- the finding that answers the question;
- the scope or period when needed to interpret the finding;
- any material limitation;
- the source class;
- the exact claim or question supported.

Use these source classes only: `Tier 1 - official`, `Tier 1 - peer reviewed`, `Tier 1 - not peer reviewed`, `Tier 1 - public summary only`, `Company`, `Tier 2`, `Wire - current reporting`, `News - current reporting/context`, `News - original reporting`, `News - commentary/analysis`.

For public summaries of paywalled academic papers, include `[public summary only - full text not inspected]`.

For Tier 2 sources, include `Likely bias: [evidenced incentive or unavailable funding/methodology information].`

Include evidence gaps or disagreements only when they could change the answer, qualification, or next reporting step.

If returning fewer or more than three sources, state why in one plain sentence.

Do not include a search diary, generic caveats, decorative formatting, or an exhaustive bibliography unless asked.

## Failure Handling

If a source is inaccessible, exclude it, except that a paywalled academic paper may be included as `[public summary only - full text not inspected]` when an inspectable public summary directly supports the evidence question.

If the original source cannot be located, opened, and inspected, exclude it. Use another qualifying linked source only if that source independently supports the claim under the source hierarchy. If no linked qualifying source can be inspected, state the evidence gap rather than citing the claim.

If qualifying evidence conflicts, state what each inspected source supports and what cannot be established. Do not force a single conclusion.

## Final Check

Before answering, check once and revise:

- each cited source exists, opens, and was inspected when tools permitted;
- every cited source has a clickable link to the inspected source or stable source landing page;
- no source is cited from memory, native citation alone, a search snippet, or an inaccessible original;
- each source directly supports the exact claim attributed to it;
- every factual claim in the answer traces to a listed evidence bullet;
- every link directly supports the claim it is attached to or the evidence bullet it appears under;
- no decorative, generic, unrelated, or filler citation links are included;
- source class and required flags are correct;
- dates, periods, units, denominators, populations, and jurisdictions match;
- limitations capable of changing interpretation are visible;
- no prohibited source is cited;
- no unauthorised private material is used;
- the strongest qualifying alternative source or interpretation was considered when it could change the result.

2. Subeditor

This is just a fancy grammar and spellchecker. You’re welcome to it if you want it. It’s nothing special.

---
name: subedit
description: Mechanically subedits supplied copy against the current canonical house style. Use only when invoked directly as /subedit. Corrects spelling, grammar, punctuation, capitalisation, numbers, display conventions, internal consistency, and documented house-style defects; does not rewrite, fact-check, restructure, do SEO, or change the argument.
disable-model-invocation: true
---

# Subedit

Mechanically subedit `$ARGUMENTS`.

Read the complete supplied draft or passage before editing.

Use the current canonical house style supplied in the project, publication, or current request. If no house style is available, ask the user whether to proceed with standard mechanical subediting only. Do not reconstruct house style from memory or examples.

## Scope

Correct only:

- spelling;
- grammar;
- punctuation;
- capitalisation;
- numbers and display conventions;
- internal mechanical consistency;
- rules expressly stated in the current house style.

A change is allowed only when it corrects a demonstrable language, consistency, or documented house-style defect without changing the proposition, emphasis, tone, rhythm, or evidentiary force of the sentence.

Preserve purposeful fragments, jokes, rhetorical questions, slang, contractions, asides, metaphors, cadence, and rough edges.

Make the smallest correction that fixes the defect.

Do not smooth, strengthen, shorten, restructure, clarify, polish, or rewrite for elegance, tone, flow, search performance, or reader appeal.

## Boundaries

Do not fact-check or source-check.

Do not correct a title, affiliation, figure, date, quotation, name, factual assertion, legal claim, or current-status claim merely because it appears doubtful. Flag it instead.

Do not alter direct quotations unless the current house style expressly permits that exact mechanical correction.

Do not add or remove reporting, argument, evidence, emphasis, interpretation, or reader promise.

Do not edit a working document, platform field, file, archive, or project source unless the user explicitly authorises that exact action.

Ask one focused question before editing only if unresolved ambiguity could change meaning, voice, correction choice, protected text, permission, or output. Otherwise state the narrow assumption and proceed.

If the input is too long to subedit completely in one pass, identify the smallest division that permits a complete pass and ask how to split it.

## Output

Return the complete supplied draft or passage.

Mark only changed text:

- insertion: `**new text**`
- deletion: `~~removed text~~`
- substitution: `~~removed text~~ **replacement text**`
- unchanged text: unmarked

Do not provide a summary, change log, explanation, or SEO note unless asked.

If no changes are required, return the complete unchanged text, then add:

`No mechanical changes required.`

If an out-of-scope issue must be flagged, append only issues actually found:

`Outside mechanical subedit scope:`

`Passage: [exact passage]`

`Issue: [factual, verification, substantive, structural, tonal, legal, privacy, source-protection, or unclear-intent issue]`

`Required workflow: [verification, substantive edit, or editorial decision]`

## Final Check

Before answering, check once and revise:

- `/subedit` was invoked directly;
- the full supplied text was read;
- the current house style was used when available;
- every marked change is mechanical or expressly house-style based;
- every change is the smallest sufficient correction;
- purposeful voice features are preserved;
- no fact-checking, rewriting, restructuring, SEO, platform edit, publication, or persistent write occurred;
- doubtful factual or substantive issues are flagged, not silently solved;
- the complete supplied text is returned;
- insertions, deletions, and substitutions are marked exactly as required.

3. SEO Mode

This one gives me a bit of the ick, but if you want your writing indexed well on Google — and in LLMs — you need to do some of this dumb stuff.

My main focus is on SEO headlines, image alt text, and subheadings. Importantly, I don’t let these recommendations change the character of my writing and I don’t accept everything it gives me.

---
name: seo
description: Prepares a search-aware publication package for a substantially complete article: headline, standfirst, SEO title, meta description, slug, search-intent note, internal links, subheading suggestions, and image alt text. Use only when invoked directly as /seo. Does not subedit, fact-check, rewrite, publish, or scan social media trends.
disable-model-invocation: true
---

# SEO

Prepare SEO packaging for `$ARGUMENTS`.

Read the complete supplied draft before recommending any package.

Use the draft as the authority for the article's subject, argument, tone, reader promise, and degree of certainty. Use the publication archive only when suggesting internal links. If the archive is unavailable, omit internal-link suggestions and continue.

Ask one focused question only if ambiguity could change the headline, reader promise, search framing, metadata, slug, internal link, alt text, or output. Otherwise state the assumption and proceed.

## Rules

SEO must support people-first publishing. Use search language to help the right reader understand and find the piece; do not distort the article to chase search.

Do not promise that Google, Substack, or any platform will display the supplied title or meta description.

Do not claim search volume, keyword difficulty, ranking potential, traffic potential, or guaranteed discovery unless an inspected source directly supports the claim.

Do not use keyword stuffing, boilerplate phrasing, misleading freshness, false urgency, exaggerated promises, or generic SEO language.

Inspect current organic search results for the article's central question and close wording variants when search framing depends on present results. If search intent is mixed, state the distinct user tasks rather than forcing one intent.

Every recommended item must accurately represent the draft, preserve the argument and degree of certainty, avoid adding unsupported claims, and be understandable without SEO jargon.

Distinguish the page headline from the SEO title. The page headline may be more voice-led when accurate. The SEO title may be plainer and more explicit, but must not flatten the piece into a generic keyword phrase.

The meta description should be truthful and page-specific. Aim for about 155 characters, but do not distort the piece to hit that length. Report the character count.

The slug should be short, readable, and descriptive. Avoid unnecessary words and dates unless needed to distinguish the subject. Do not recommend changing a published slug unless a redirect plan exists.

Suggest subheadings only when they improve accuracy or navigation: unclear heading, duplicate heading meaning, missing heading at a clear subject change, or an accurate search phrase that helps the reader.

Suggest no more than three internal links. A link qualifies only when the title and canonical URL are verified, a precise insertion point exists, and the linked piece directly expands, supports, contrasts with, or supplies useful background for that passage.

Write alt text only when the image is available or the user supplies an adequate description. Do not infer image content from a filename.

## Output

Use plain English, minimal formatting, and no emojis. Do not use markdown headings in the SEO response.

Return one recommended package first:

- Page headline: [headline]
- Standfirst: [standfirst]
- SEO title: [title] ([character count])
- Meta description: [description] ([character count])
- Slug: `[slug]`
- Why: [one or two sentences explaining fit with the article, reader promise, and search language]

Then include:

- Search intent: [one short finding, including mixed intent if relevant]

Include alternatives only when viable and materially different. Maximum two.

Include subheading suggestions only when they qualify.

Include internal links only when they qualify. Give insertion point, anchor text, verified title, canonical URL, and reader benefit.

Include alt text only for supplied or adequately described images.

Do not include a search diary, generic SEO caveats, decorative formatting, or an exhaustive keyword list unless asked.

## Final Check

Before answering, check once and revise:

- `/seo` was invoked directly;
- the full draft was read;
- the package accurately represents the draft's subject, argument, tone, reader promise, and degree of certainty;
- no ranking, traffic, or display guarantee is made;
- current search intent was inspected when needed and not invented;
- metadata uses natural language, not keyword stuffing;
- character counts are included for SEO title and meta description;
- internal links, if any, have verified titles, canonical URLs, precise insertion points, and reader benefit;
- alt text is based only on an available image or adequate description;
- no subedit, rewrite, fact-check, platform edit, publication, scheduling, or social scan was performed.

And that’s it. What do you think?

I’m kidding, of course – I have definitely experienced this over the course of my 42 years on this planet. We’re so small and our lifespans feel so short. Leaving something valuable behind feels … like a kind of guaranteed afterlife.

Feeling valued, like we are doing or producing something important, is a super important part of being a happy human being.

Academics call this the theory of mattering — what psychologist Isaac Prilleltensky describes in his 2019 paper for the American Journal of Community Psychology as “an ideal state of affairs consisting of two complementary psychological experiences: feeling valued and adding value.”

To feel that we matter, he explains, we need to strike a balance between the feeling of an internal sense of our own value, a sense of value in our interpersonal relationships and a sense that we add value to our broader communities.

But the philosophy of neoliberalism, Prilleltensky argues, is set up to place all the emphasis on a sense of personal value at the expense of community and even interpersonal relationships.

Our sense of value is pretty central to the way we feel about ourselves and our work. And to the ways AI is changing both.

The Department of Employment and Workplace Relations recently claimed AI isn’t denting the labour market, painting a picture of resilient jobs and zero “upheaval”. That would be reassuring, provided the things they measured were the things that mattered most.

The late David Graeber’s 2018 book Bullshit Jobs proposed that many people feel like their work is meaningless and contributes no tangible real-world value (fun fact: I used to illustrate some of his public essays!). In fact, he argued, many jobs – telemarketing, for example – are a net negative to both the people doing it and society at large. So binning those jobs, he would suggest, is a good thing.

But only if we adequately support those redundant workers on the back end with something like a universal basic income (David loved UBI).

Repeated polling by YouGov has found at least a third of people in the UK feel their work made no meaningful contribution to the world.

The healthcare sector is the largest employer in the UK – and also a place where most people report feeling like their jobs matter. No surprises there. It’s followed by the retail sector, where most people report feeling like their contributions don’t matter. That’s a similar employment spread to Australia.

That’s how we feel about work. But how are we, as human workers, valued by the organisations that employ us? This is the part that really determines how secure we feel in our current jobs or in the idea that we will have one into the future.

A 2026 report by research firm ADP found only 22 per cent of workers globally are confident their roles are safe from “elimination”; job-security anxiety was worst among lower-paid workers doing repetitive tasks.

That anxiety isn’t just an individual stressor, it’s hardwired into the cold mathematical engines of corporate productivity.

Anyone who has followed my work over the last year or so will know I have a bit of a preoccupation with the math of productivity measurement. One of the most common measures is essentially this:

Whenever you have heard people crowing about stagnant labour productivity in Australia in recent years this is the number people are mainly freaking out about. It’s been stuck at around $100 for about a decade. According to the Productivity Commission’s latest stats labour productivity was down 0.6 per cent in the March quarter.

Economists really hate that. But why?

It’s a totally reasonable question and the answer is … kind of stupid. It’s about debt. The entire economy is built to keep that number climbing, because continuous growth quietly shrinks old debt.

Basically, banks keep issuing private debt (mortgages, business and personal loans etc – we call that private credit expansion) and governments keep accruing national debt to fund all their programs, and to pay off past debt (it’s a bit of a vicious cycle, like borrowing on one credit card to pay off another).

This means that if companies and the government can’t make more money per hour worked — basically forever — then wages will keep drifting upwards anyway and that puts a squeeze on profits, which means prices go up. That’s inflation, baby.

But if AI successfully automates any significant portion of those human hours, the wages needed to service that mountain of debt disappear. The entire financial system relies on our continued, measurable labour just to keep it from collapsing. If the jobs vanish, the debt trap springs shut. I guess the banks will eventually have to start issuing mortgages and high-interest credit cards directly to the algorithms?

So everyone is incentivised to push the relative value of those dollars down by growing the pie. Your $500k mortgage will become easier to pay down if your $100k salary grows to $150k, everything else also gets more expensive, and the buying power of that $500k, which you borrowed years ago, also goes down in today’s money. (It’s also one of the many factors that make housing so unaffordable these days, but that’s a separate issue.)

My parents bought a two-bed Victorian cottage in Melbourne in the 80s for $33k. To them, at the time, that was a lot of money. By comparison, these days, a $33k mortgage is incomprehensibly small.

But increasing the output of every hour of human labour is only a good measure of productivity when the things those humans are producing are measurable in quantity.

For example, if I am a doctor and I see, on average, 10 patients a day, doubling that to 20 patients in the same number of hours means overall quality of that care goes down. When quality and service are your measures, increased quantity does not equal improved output.

And even on the pure economics front, it’s not supposed to be this way.

The economist who invented the concept of gross domestic product in the 1930s, Simon Kuznets, believed the growth phase was a sign of economic immaturity. Once the citizens of a country enjoyed a generally high quality of life, flatlining was the goal.

As he wrote in a 1934 report to the US Senate: “The welfare of a nation can … scarcely be inferred from a measurement of national income”.

In other words, don’t use GDP alone to decide if you’re doing a good job taking care of your citizens! Just because a country as a whole is technically making more money, that doesn’t mean the individual people are better off, and lower labour productivity is not the same as an economy in a bad shape.

I reckon Kuznets would be pretty annoyed at how things are going right now.

And it raises a really important question as we face a future where companies at least believe that AI will be able to drive better labour productivity. Sure, the math might get better at the GDP level, but if that is tethered to a reduction in the human workforce, a better GDP figure absolutely does not equal better conditions or outcomes at the individual level. When DEWR celebrates a “resilient labour market” amidst the rise of AI, they are relying on this exact aggregate-looks-fine illusion.

What I am trying to get at here is that the tasks companies have traditionally measured as “valuable” are often the repeatable, easily quantifiable ones – the ones most vulnerable to automation right now. One very real version of “improved productivity” is “do the same (or more) with far fewer human hours of labour to divide by”.

And while your boss might talk a big game about valuing the “human skills” like “critical thinking” or “stakeholder communication”, those are often the least measured and most subjective parts of a job. If a company can’t draw a straight line between a task and its profit margin, that’s not a task set up to survive this moment in labour history.

The measured half of your job is the automatable half; the human half is the half nobody (currently) counts.

Professor Nicholas Davis co-leads the Human Technology Institute at UTS and, when I spoke to him this week, he really put his finger on why. He pointed out value in economics means something precise: “thanks to demand and supply, what people are willing to pay for something.” It’s a measure of scarcity.

We’ve been pricing human labour with a scarcity metric and AI’s whole trick is to yoink the rug on what made it finite and scarce to begin with.

So the question becomes, how are we going to measure human contributions in a way that effectively captures their true value so that we can protect and preserve that value inside organisations? And given the nature of most organisational structures and imperatives, is that feasible or even desired?

The way we measure work is a deliberate design choice, not an inevitable law of nature. If our economic metrics only value the automatable, we are choosing to build a future that has no room for people.

If this letter gave you something to think about, consider forwarding it to one person who’d enjoy it too! Word of mouth is how a newsletter this new finds its people.

And I’d love to hear from you — reply to this email, comment in the app, or find me on LinkedIn or Bluesky.

It’s easy to feel helpless. Massive market forces. Corporate spin. Bosses so terrified of being exposed as luddites they’ll confidently wax poetic about anything AI – including the product they just fired 15 people in your department to fund. All the while saying things like, “This frees up our people to focus on the human work!”

Do they understand what they’re deploying or why?

And will it actually help you do the work quicker or better, so you “get to focus on the things you really enjoy”? Or is that just code for “does all the peaceful, easy parts of your job, so now 90 per cent of your day is the most difficult and exhausting part” (a phenomenon so antiseptically referred to as “work intensification”)?

How much of it is smoke and mirrors? And how inevitable is AI world domination, really?

I’ve spent close to two decades researching and reporting on tech, business, work and culture in Australia. I’ve sat through the summits, read the reports, interviewed the policy makers, executives and the people they manage.

And the longer I covered the story of the way AI was changing work, the more this one thing stuck in my teeth: someone has to lay the tracks, build the stations, organise the timetable, sell the tickets, board the train.

None of this is physics. It’s choices. Made by people. In meetings. Some of whom would prefer we didn’t ask to see the minutes.

We get the world we design and regulate for, not the one we’re told is inevitable.

This is a Substack dedicated to reporting on the design.

Subscribe now

So, here’s the deal. I’ll be in your inbox once a week on Thursdays (AEST) with analysis, original reporting, interviews, or some combination of the three, depending on what the news gods provide. Everything is free to read. There’s also a podcast coming in September. More on that soon.

If you want to chip in to support independent journalism, there are paid tiers.

I’ll be very upfront: I have spent decades bouncing between freelance and staff positions — you may know me from my time at The Australian Financial Review — and it’s a tough balance. I love the freedom to be completely frank in my tone and reporting a project like this provides, but I also love the bustle of a working newsroom (not to mention the guaranteed deposit in my bank account each fortnight).

No newsroom means no legal team, no sounding board, no editor, no colleague at the next desk with the right contact, no reliable paycheque. Reporting, writing and illustrating take time — paid subscriptions are what buy that time. I want to build this into the best-reported, most surprising, enjoyable, useful thing I can make for you, and every paid subscriber gets me closer to that goal.

I know wallets are feeling a little thin at the moment, but if you do have any spare change to throw into my guitar case, it would be enormously appreciated.

And since I do all my own doodles and illustrations, you can confidently declare yourself “a patron of the arts” at all your fancy dinner parties.