The Stakes
That Cannot
Be Replayed
A diagnostic, not a survival guide.
The first essay argued that in the AI epoch, value moves to the edges of the work. Routine execution gets cheap. Even the architectural craft of building AI systems gets cheap, eventually. What stays human, I said, is the work at the boundaries: choosing what is actually worth doing, and discerning whether it actually got done.
Then I reread it and noticed something I had tucked into a single paragraph. I had admitted there that the architectural layer wouldn’t stay durable either. AI systems will learn to run their own workflows, design their own constraints and evaluate their own output. I had written a piece whose central claim about human durability quietly undercut itself. The middle layer I had been treating as solid ground was borrowed leverage on a timer.
Pulling on that thread produced this piece. I want to name what actually stays durable, why it does, and what most readers should do with that information. And I want to do it carefully, because the most comforting version of this argument, that human judgment is special and AI can’t replace it, is exactly what gets sold in every consulting deck and AI-coexistence white paper. It is wrong in a very specific way, and believing it will cost people their bearings.
This essay is a diagnostic: a tool for telling real, irreversible stakes from theatrical ones. It is not a survival guide, a political program or a battle plan. The work of organizing to defend these stakes is a separate project, and I’m developing it with people who organize for a living.
If you came for a survival map, this isn’t it. If you came to learn how to tell the real thing from a convincing imitation of it, read on.
What I Got Wrong
The first essay treated boundary work as one thing: choosing what to build and judging whether it worked, both safely out of the machine’s reach.
It isn’t one thing, and most of it isn’t safe.
Almost any boundary work that can be measured, scored, A/B tested, audited or written up as a procedure will eventually be absorbed. AI systems are formidably good at optimizing once someone writes down what success looks like. They are also good at producing output that merely looks like success when nobody checks too hard; economists have called this Goodhart’s law since the 1970s, and AI researchers rediscovered it as specification gaming. And they are exceptionally good at being cited by institutions as the “objective” reason a decision was made.
That third skill is where the real displacement happens. AI doesn’t have to become smart enough to take the human’s seat. The institution only has to wrap enough procedure around the AI’s output that the human’s seat becomes ceremonial.
The mechanism is older than the current wave of AI. Since the mid-1990s, American mortgage lending has run on automated underwriting systems. The score comes out of the system, and a human signs the file. Where underwriters have been displaced, it has mostly not been because a machine replicated their nuanced judgment. It has been because the institution decided it could rely on the score and keep a person around to sign off. The human is still there. The human’s judgment is not.
This is the compression mechanism we actually face, and most professional knowledge work is more exposed to it than we like to admit. If your job rests on judgment an institution can wrap in a process, your judgment is on a clock. AI capability doesn’t set that clock. The institution does, when it decides the procedural wrapping is good enough.
What Cannot Be Wrapped
Some judgment can’t be wrapped, because the person making it has something at stake that no institution can insure, indemnify, replace or roll back.
A surgeon who loses a patient on the table doesn’t get rolled back to the morning.
A parent who shapes a child badly doesn’t get to rerun the childhood.
A founder who burns a community’s trust doesn’t get a fresh ledger when the funding round closes.
A combat leader who orders the wrong assault doesn’t get to reload the save and try the other flank.
What these share isn’t some mystical essence of the human spirit. They share a person bearing a loss that no rewind, insurance payout, process update or replacement hire can restore. The stakes are irreversible. The person doing the work knows it, the people affected know it, and everyone organizes around that shared knowledge.
This is the thing that doesn’t compress. Irreversible stakes have three parts: someone bears a real loss, the loss can’t be replayed, and a community outside that person recognizes it. That structure is the one foundation I can find that gives judgment real weight. Every other kind of judgment is something an institution can ultimately rewind, replace or absorb, and under enough economic pressure, institutions absorb everything they can.
A note on scope: the finer mechanics, what irreversibility means structurally and what counts as a community of consequence, belong to a companion piece. Here the working definition is the plain one: someone bears a loss that can’t be restored, and the people affected know it.
The Map of What Compresses
I find it useful to sort cognitive work into four kinds of decisions, not four kinds of jobs. Nearly everyone reading this has a role that mixes several of them. Compression hits the decision, not the title on your résumé.
Pre-instrumented work is judgment that looks essential only because nobody has built the measurements yet. The success criteria are still implicit. Once someone defines what good output looks like, the “judgment” turns out to have been optimization nobody knew how to specify. I suspect much of what passes for specialist craft today sits here, and the measurements are arriving now.
Instrumented work has measurable success criteria and a feedback loop that closes cleanly: code that compiles, contracts that match a template, designs that meet a brief, analyses that match the data. It is already compressing fast. The human role here is supervisory, and supervision is the next thing institutions will wrap.
Wrapped work is judgment that an institution has surrounded with so much procedure that the human is reduced to a stamp. Think underwriting, lending, claims adjudication, routine legal drafting, standard clinical recommendations. The human is nominally accountable, but the institution can dilute, transfer or absorb that accountability through model updates, vendor contracts and audit formalities. The sociologist Madeleine Clare Elish has a name for the person left in that seat: a moral crumple zone, the human part of an automated system that absorbs the blame when it fails. I think this is where most professional displacement is happening now, and the wrapping is tightening.
Stakes-bearing work is the decision whose maker bears irreversible loss for getting it wrong, whose community recognizes the maker as one of its own, and whose consequences can’t be insured or replayed: founding commitments, raising particular children, defending particular traditions, building particular institutions. This is the durable layer.
The reason to think in decisions rather than jobs is that almost every role is a mix. A surgeon does stakes-bearing work in the operating room and wrapped work on the discharge paperwork. A founder does stakes-bearing work in early commitments and wrapped work in compliance filings. A claims processor signing an algorithmic denial is doing wrapped work; the same processor objecting on the record to a denial they were told to enter is doing something much closer to stakes-bearing work.
The unit of analysis is the decision, not the line on the résumé.
So the diagnostic question isn’t “Is my job safe?” It is: which parts of my work bear irreversible stakes, and which parts are wrapped procedure that an institution will eventually absorb?
Most of us can be honest about this for about ten minutes, and then we’re defensive for the rest of the afternoon. The defensive answer sounds like “My judgment is irreplaceable because I see things the model can’t.” The honest answer sounds like “The parts of my job that touch real stakes are smaller than my title suggests. The rest is wrapped, and the wrapping is getting tighter.”
The Simulation Problem
The hard part isn’t only that wrapped work is replacing stakes-bearing work. It’s that there will be enormous economic pressure to manufacture things that look like stakes-bearing work and aren’t.
Status markets reward the appearance of stakes. Communities, credentials, traditions, oaths and pledges are the public signals of stakes-bearing membership, and every one of those signals can be produced without the reality underneath it.
A “community of practice” that fits the formal description of a stakes-bearing group, but keeps every consequence inside its own walls and never answers to anyone outside its membership, is a sealed echo chamber that validates itself.
A “founder” who pledges everything to a project that the surrounding structure has fully insured against failure is performing.
A “tradition” hollowed out into a brand asset can’t bear the weight the brand claims for it.
I expect this to get worse. As compression eats more wrapped work and the displaced population grows, demand for high-status, stakes-bearing roles should climb steeply. Supply will lag. Irreversible stakes are unusual in modern life, and many institutions, insurance and limited liability among them, exist precisely to absorb them. But supply isn’t fixed forever. Stakes are partly made by institutions, and a much richer society could choose to create more roles in which people bear real consequences. Whether it will, and how large the human share of an AI economy can grow, is a question of economics rather than diagnosis. Part III takes it up.
The gap between demand for legitimacy and the supply of real stakes-bearing work will be filled by imitations of varying sophistication. Some will be sincere mistakes. Some will be deliberate grifts. Most will be somewhere in between.
A diagnostic can’t fix that. Diagnostics don’t build political coalitions, stop the imitation economy or protect anyone from displacement. What a diagnostic can do is give people who care a way to tell the difference.
How to Tell the Difference
Put these four questions to any role, community, institution or claim that says it bears stakes. They aren’t perfectly precise, but they’re sharp enough to catch most imitations.
1 Who actually loses, and what can’t be restored?
Is the loss absorbed by an organization or borne by a specific person?
It’s an imitation if… The institution loses some money but is insured, or the model simply gets retrained
It’s real if… A specific person permanently loses standing, freedom, legacy or their place in the community
2 Did the person who’s accountable have real authority?
Accountability counts only when it attaches to the person whose judgment could actually change the outcome.
It’s an imitation if… Someone is blamed for a result they had no power to change, such as a pilot held liable for software he couldn’t override. That person is absorbing risk the institution offloaded
It’s real if… The authority to decide and the consequence of deciding rest on the same person
3 Does the affected community recognize this person?
Not credentials or fluency in the jargon: would the community still claim them if all the procedural scaffolding were disclosed?
It’s an imitation if… The recognition wouldn’t survive full disclosure. The legitimacy is borrowed
It’s real if… The community claims them regardless of the scaffolding
4 Does the work answer to a world that didn’t consent to judge it?
When the community judges badly, does it lose markets, fail visibly or get refuted by reality? Or does it just expel dissenters?
It’s an imitation if… The actor can buy, fake, curate or corner the verdicts it claims to answer to. It’s a sealed echo chamber
It’s real if… The resistance comes from outside and can’t be bribed
These questions aren’t magic. Actors with strong enough incentives will learn to beat them. But they will beat them in ways that are themselves revealing, if anyone is watching.
Which leads to the last point.
The Standard Has to Live Somewhere
Any test that hands out legitimacy in a market this large will eventually be attacked, softened and captured by the people it threatens most.
That pressure has shaped credentialing regimes throughout modern history. Economists have a name for where it leads, regulatory capture, which George Stigler described in 1971: the regulated end up steering the regulator. Professional certifications, fiduciary standards, press credentials and food labels like “organic” have all been accused of drifting from rigorous standard toward incumbent-friendly logo, some with more justice than others. There’s no reason to expect a test for stakes-bearing work to escape the same pressure. A sharper test invites more capture pressure, not less.
The protection can’t be more clauses or more rounds of self-policing. The protection has to be public application. A diagnostic that matters has to be applied to high-status cases, with the verdicts and the reasoning published in full, so the track record stands or falls in public. And the choice of which cases to examine can’t be left only to the people most invested in how the framework looks. That’s the only honest mechanism I know for keeping this kind of work from rotting in place.
Eventually, a diagnostic that matters will need an independent institutional home, one built to resist the very forces it diagnoses. That home doesn’t exist yet, and building it is not the subject of this piece.
A Real Limit, Named
This framework carries an obvious political risk, and I want to meet it head-on.
The argument is about durability, not desert. When I say stakes-bearing work can’t be compressed, I’m making a cold claim about what AI and institutions can and can’t absorb. I’m not claiming that the people who currently hold stakes-bearing roles deserve them.
Many of them got there through accidents of class, geography, family and timing. Nothing here conflicts with critiques of inherited privilege. If anything, it sharpens them. If displacement runs the way I’ve described, access to stakes-bearing work becomes the scarcest resource in the economy, and who gets to gatekeep it by inheritance becomes a far more urgent question, not a less urgent one.
If you finish this piece believing the people now holding the durable roles have been shown to be morally superior, you’ve misread it. Being structurally durable doesn’t make anyone morally legitimate. Those are two different claims, and this piece makes only the first.
What This Is For
If your job is mostly wrapped work, this diagnostic is mostly bad news. The wrapping will tighten and your role will thin out. Skill and political positioning can postpone that, but they can’t stop it, and pretending otherwise costs you the time you have to prepare.
If your job is mostly stakes-bearing work, the news is also bad, in a different way. The imitation economy is going to crowd into your space. People performing membership will approach your community. Your standards will be attacked as elitist by people with a direct financial interest in softening them. The work itself won’t get easier, and defending it will get much harder.
If your job is mostly pre-instrumented or instrumented work, the task is to look closely at your day. Ask which of your decisions actually bear irreversible stakes, and whether you’re willing to organize your career around those. Most people who answer honestly will be surprised. Some of what feels durable will turn out to be wrapped work in disguise. Some of what feels marginal will turn out to carry the weight of your future.
This diagnostic doesn’t tell anyone what to do. It tells them what they’re looking at. Action comes after seeing, and seeing is what this piece is for.
The first essay said value moves to the boundaries. This one says most of those boundaries are already wrapped, the wrapping will tighten, and the unwrapped layer is far smaller than the demand for it will pretend.
The work that survives this transition is work where someone bears a loss that can’t be restored, and where that same person had the real authority to decide.
Everything else is on a clock.
Developed through adversarial review in the Crucible, my multi-model review process, with a council of four AI systems: Cartographer (Claude Opus 4.7), Adversary (Grok), Engineer (Codex/GPT-5) and World Builder (Gemini). Director and integrator: John Fite.
First published May 2026 · Revised October 2026