These days it feels like AI’s swallowing the world (some would say it already has). It’s everywhere. Impossible to ignore. Even my mom’s using ChatGPT, daily.

If you’re a white-collar worker though, specifically in tech, when you hear or think “AI” it probably means something different at an existential level. Chances are, over the last several months, it’s very quickly become your reliable work companion.

And, you know, there’s a version of the AI workplace that stands to liberate white-collar labor from drudgery and lets human attention flow toward what humans are truly good at.

What are humans truly good at, you ask? Definitely worth defining before we go any further.

Not chess. Not multiplication. Not memorizing code. Humans are excellent at the things that took our ancestors a hundred thousand plus years to evolve:

• the carpenter whose hands feel grain through the saw

• the trial lawyer pivoting an argument three seconds into a witness's answer

• the engineer who solves an electrical problem by thinking about plumbing

• the teacher who knows exactly when to explain and when to let the student struggle

• the negotiator who can read which silence at the table means yes and which means “this deal is dead”

Cognitive science has a name for this layer of competence. It's called tacit knowledge, after Michael Polanyi's observation that we know more than we can tell. It's the substrate of every kind of expertise that survives in the wild.

Ah, yes, please AI, take the chart-reading and the form-filling and the email-triage, and so on off my desk, so I can spend my day doing this so called “work humans actually do exceptionally well” and reach ultimate fulfillment in life, alongside my fellow human peers. I whole-heartedly welcome this. It sounds like the version of AI we should all be racing toward, right? A true enabler of human potential.

Well, the problem is this isn’t the version you and I are getting. Not really. The actual version we’re getting has its top sights set on strip-mining the worker for behavioral patterns, then baking those patterns into systems that are designed from the ground up to displace the worker. All being framed as “progress”.

We’re getting this version almost everywhere, and the only question worth asking with any seriousness now that the cat’s out of the bag is: Why?

Most current writing about AI starts by taking a stance on the technology (e.g. is it good for us, bad for us, transformative, overhyped, existential threat, cool toy) and then reasons forward from the posture. That’s the wrong place to begin, I think. AI has the same property as most powerful tools: it does what the system around it rewards it for doing. The intelligent move is to look deeper at the system.

The current system, in 2026, has converged on a deployment pattern with predictable properties, and those properties aren’t features of the technology. They’re features of the incentive structure that surrounds it.

In April, Reuters reported that Meta was installing software on U.S.-based employee computers to capture mouse movements, clicks, keystrokes, and occasional screen snapshots. The internal program’s called the Model Capability Initiative, a phrase engineered to convert workplace surveillance into a procurement line item, no doubt. Meta said the data wouldn’t be used for performance assessments. The goal, the company explained, was to train AI agents to perform everyday computer tasks.

Three weeks later, employees at multiple U.S. offices distributed flyers (on vending machines, in meeting rooms, atop toilet paper dispensers) asking colleagues whether they wanted to work at the “Employee Data Extraction Factory.” A parallel unionization drive had quietly begun in the UK with United Tech and Allied Workers.

I know what you’re probably thinking: “These workers were being a little extreme.” And you would be right, if only those flyers hadn’t landed roughly a week before Meta started to cut ~10% of its workforce.

I also have to admit that Meta’s defense was, on its face, accurate.

If you want software to act like a human using a computer, it makes sense to optimize for collecting any and all traces of humans using said computers. Synthetic data has limits. AI Agents would need the pause before the dropdown, the correction after the wrong click, the weird little ritual of opening a zillion tabs, forgetting why, returning to Slack, copying a sentence, pasting it somewhere random, deleting half of it… in order to truly approximate “cloning” the human-led work.

And what if I told you Meta isn’t the only company engaging in this practice?

In January, Wired and TechCrunch reported that OpenAI, working with the data-labeling firm Handshake AI, had begun asking contractors to upload concrete work products from their previous jobs under instructions (in OpenAI’s own brief to contractors) to “take existing pieces of long-term or complex work that you’ve done in your occupation and turn each into a task.” Different vector, same move.

Meta captures its own employees through behavioral software. OpenAI captures contractors through structured upload. Looks like a secondary market could emerge for the operational exhaust of dead companies involving Slack archives, Jira tickets, and email threads, etc., bought by AI labs, for training. It also looks like this pattern actually doesn't stop here.

In May, METR released its first Frontier Risk Report, a pilot assessment conducted in partnership with Anthropic, Google, Meta, and OpenAI to study misalignment risks from internal AI agents. The operational details state plainly that the same loop has already closed on the people building at these companies. Anthropic reported to METR that a substantial portion of its own code is now written by AI, with engineers and researchers shifting away from writing toward reviewing the output of coding agents. OpenAI and Google confirmed essentially the same restructuring across their R&D teams. The “training-set workers” aren't only on the periphery or outside of these companies. They're inside them, at the core.

Is this a taste of what’s to come for all white-collar work, not just software engineering?

Valid question.

I think the most interesting question isn’t whether the machine should be allowed to “want” training-set workers wholesale. The question is what kind of institutional arrangement should govern the transaction, and what we can learn from the arrangement we’re currently building.

Ultimately, I think this is the, if not one of the most relevant economic issues of the early-to-mid 21th century, and one your grandchildren are likely to read about in the history books.

Subscribe now

WHAT YOU ARE WHEN NOBODY’S LOOKING

Focusing on the deeper concern, you might be shocked to learn that this is nothing new. It's a reframing of older debates—Marx on alienation, Arendt on the human condition, Heidegger on craft and the standing-reserve—that asked what work does to who/what you are. This concept of the AI-ruled workplace asks a sharper version of the same question.

To take that seriously you have to think about where those capacities came from in the first place.

Human cognition didn’t fall from the sky.

The things you do well at work, from the things that make you worth hiring and onward, are products of evolutionary processes that operated on time scales measured in hundreds of millennia. Reading a room is a fitness adaptation. Modeling another person’s mental state is a fitness adaptation. Recognizing patterns in chaotic environments, recovering gracefully from interruption, extrapolating from incomplete evidence, sensing when a colleague’s “fine” actually means “not fine”; none of these arrived with your offer letter. They arrived through millions of years of selection pressure in social primate populations where being able to do these things meant your DNA continued, and not being able to do them meant it didn’t. The technical term for this layer of competence (i.e. Polanyi’s tacit knowledge) names a property of expertise that’s hardest to articulate, hardest to teach, and historically impossible to capture in any explicit form.

Until, now?

AI-powered behavioral capture software is the first technology in human history that seems capable of producing high-fidelity recordings of tacit knowledge in action. The mouse movements and click sequences and dropdown selections that Meta is harvesting aren’t trivia. At least I don’t think they are. They’re externalized traces of cognitive processes that evolution spent eons building, processes the workers performing them couldn’t fully describe even if they wanted to. The fact that the workers can’t describe the processes is precisely why the recordings are valuable. I suspect the machine won’t need the explanation. It just needs the trace.

Interesting patterns will emerge from this practice.

Lucky for you, you don’t have to travel to the future to see this in action. These AI systems are already “tacit enough” to displace workers en-masse. This is at the root of all those tech layoffs that keep making the news headlines (there’s another story here about executive malfeasance, but it’s out of scope). Any honest software engineer doing worthwhile work alongside AI can attest to just how tacit it already is. If this deployment is currently eliminating engineering jobs which require high-levels of tacit knowledge, you don’t think it’s likely to spread to other sectors?

Spoiler alert: it 100% will, and soon

The AIs are here, and they’re eating jobs for breakfast, so what’s next?

Under the current system, on the road to total implementation, you’ll start noticing the following net effects, and they’re stranger than the surface argument lets on.

Primary: Private-life consequences

Work captures so much of your cognitive output daily, and what’s left to be private becomes both more important and more constrained. Your retreat to “real life” gets narrower exactly as the capture intensifies. The version of you that exists outside the recordings, and the part that hasn’t been quantized into a feature vector becomes the place where you continue to be a person rather than a pattern.

That part is also, by definition, the part that isn’t getting remunerated for its evolutionary inheritance, but it’s shrinking.

I’m willing to bet a lot of white-collar folks using AI daily are already feeling this. You feel more tired at the end of the day, and it might be hard to explain why. You’re working longer hours even though you’re using AI.

One of the things that could bring solace is that there’s a version of Goodhart’s Law that could apply here, quietly. When the measure becomes a target (as it already has), it ceases to be a good measure. In other words, when a person’s competence becomes a data-set, the competence could start to degrade in exactly the dimensions that made it worth capturing in the first place.

The thing the labs most want to record could be the thing surveillance most likely reliably destroys.

Which would mean the current deployment is, on its own terms, partially self-defeating. If you can’t reliably capture tacit knowledge by surveilling people, because surveillance changes the behavior being surveilled, all you’re doing is optimizing for short-term shippability and a noisy training data-set.

Secondary: Atrophy & Decay (less obvious)

This is the part that won't show up in any quarterly review. As the deployment spreads, your job quietly changes shape. You stop doing the work and start supervising the thing that does the work. You review the PR instead of writing it. You approve the draft instead of drafting it.

You catch the model's mistakes instead of making, and learning from, your own.

Lots of people are already here as well. This sounds like a good place to be. It isn’t. And without a clear governance and operating model, this path will invariably lead you and/or your organization to collapse.

In 1983, a cognitive psychologist named Lisanne Bainbridge published a short paper called “Ironies of Automation” that’s been quietly predicting your future for forty years. The observation was simple and brutal: automate most of a task, leave the human the rest, and the human stops practicing the skill.

The skill decays.

Then (and this is the irony bit) you keep the human around precisely for the rare, hard cases the automation can’t handle, which are exactly the cases that demand the deep skill they’re no longer allowed to build. Failure mode.The jury’s still out on whether or not we’ll ever be able to map tacit knowledge 1:1 onto machines. Does it matter?

We do know humans can and do accumulate it, rep by rep, over hundreds/thousands of hours of actually doing the thing. Take away the reps and you don’t get a faster expert. You get a slower one, then a worse one, then none at all. The software engineer reviewing AI output all day is slowly forgetting how to write the code. The one who started coding with AI and never challenged him/herself to learn the fundamentals and practice without it will more than likely never learn in the first place.

We also know the models get better by watching you, and without discipline, you get worse by watching them. The machines are winning.

If put on a graph, it’s obvious to me these two curves are on a crash-course bound to overlap. And after they cross, the most basic question of your working life, “What am I here for?”, will suddenly start having a very uncomfortable answer.

Final: Evolutionary Irrelevance (more obvious)

The cognitive capacities that made you valuable to your tribe, that are, in the strictest possible sense, your inheritance from the entire human lineage, have successfully been extracted from you, and embedded in systems engineered to render you redundant.

There was no precedent for this in human history, but somehow we got it done.

Previous extractive technologies took your time, your physical labor, your attention. None of them took the thing selection pressure spent the longest building. Your edge over the model was your edge over the model right up until the model watched you long enough to copy you, fully.

I believe this is the most likely end-state given the current trajectory we’re on, and like I’ve already alluded to earlier, it makes sense to fight it. This is also the point at which I could introduce a myriad of “AI doomer arguments”, but I’ll refrain from doing so. At the end of the day, I choose to remain optimistic, because we have time and options.

In population-fitness terms, this would be a novel scenario: our own evolved competencies being weaponized against our fitness payoffs.

Of course, there’s a non-zero chance all of this turns out to be conjecture in the end. There’s unfortunately still a lot of mystery with respect to exactly what happens inside of LLMs. There are also ways in which this deployment could end up going totally asymptote. I’m not a time-traveler. Nor am I psychic. For the sake of humanity however, I hope we find a better way of deploying “tacit AIs” capable of supplanting all work across the world. One thing that’s become clearer in recent months is we don’t need AGI/ASI in order to get to mass worker displacement.

Subscribe now

CONTROL LOOP UPGRADE

The AIs are here, they’re eating jobs for breakfast, and it’s all ending with human work becoming irrelevant, so what’s the mechanism driving this transformation?

Cybernetics is the post-WWII discipline founded by Norbert Wiener, who gave the twentieth century its formal vocabulary for thinking about feedback. A control system has four parts:

• a sensor that detects the state of the world

• a target the system is trying to reach

• an error signal representing the gap

• a corrective action that closes the gap

The old workplace fit the diagram with the elegance of a stencil. Managers measured output. They compared it to quotas. They issued corrections. The corrections produced new output. The loop closed.

The new AI workplace bolts a third-order loop onto that inheritance, and this is where the philosophical, evolutionary, and cybernetic frames start to bleed into each other. The old workplace loop corrected the worker’s behavior toward a target. The new workplace loop captures the behavior, abstracts it into a model, and deploys the model into the world to perform the behavior the worker used to perform. The corrective action is no longer “do more of this.” The corrective action is “I, the model, will now do something like this, and you, the worker, are part of the input to the function that produced me.”

What’s easy to miss is that biological organisms are also cybernetic systems.

Homeostasis is feedback. Hunger is an error signal. Reaching for the glass of water is the corrective action. Every adaptive process evolution has produced is, in formal terms, a control loop. The cognitive capacities described in the previous section (tacit knowledge, situated judgment, embodied skill) are themselves outputs of an enormously long feedback process in which natural selection played the role of the corrective action. We are, organism and behavior alike, the most refined cybernetic system on the planet, and the refinement isn’t coincidental. It’s the product of a search algorithm that’s been running for roughly 3.7 billion years.

The technical name for a system that produces and reproduces itself by recursively integrating its own outputs into its inputs is: autopoietic.

Biological life is autopoietic.

The cell membrane synthesizes the chemistry that synthesizes the cell membrane.

What’s happening in the AI workplace is the first attempt to engineer an autopoietic loop in which the substrate being recursively integrated isn’t a chemical signal but the evolved cognitive output of a sapient organism.

Read that last sentence again, and take a deep breath.

This isn’t a metaphor, friend. It’s a description of the engineering problem the AI labs are attempting to solve. The substrate is you. The recursion is the training pipeline. The autopoiesis is the eventual closure of the loop, in which the model’s outputs become the work, the work becomes the data, the data trains the next model, and the human being who originally performed the work is, in the strictest cybernetic sense, no longer required for the loop to maintain itself.

The old control system wanted obedience. The new one wants imitation. The next one wants substitution. And the one after that wants the human out of the loop entirely.

I’m not talking about surveillance by itself, or more mass layoffs. The dangerous thing is the construction, in real time, of a feedback system that uses evolved human cognition as its training source and is engineered, on purpose, to make that cognition obsolete.

We’re all watching the construction of a system whose long-run telos is to make a particular kind of being—the embodied, adaptive, situated human worker—economically and cybernetically redundant.

You don’t have to be a Luddite to find that worth pausing on. Maybe even worth acting on.

Before we get to acting, though, there's a comforting story standing in the way: the one where AI is overhyped and underwhelming, so maybe none of this is really coming for you after all. That story is the most expensive lie in the whole arrangement. Taking it apart, and then asking what we actually do about any of this, is what Part II gets into.

More to come soon.

Subscribe now