The number arrived the way most numbers do now — stripped of its qualifier and polished into a warning. Somewhere between the IMF's January report and my own community, “40% of jobs are exposed to AI” had quietly become “AI will disrupt 40% of jobs.” A member of The Silent Node — a woman I have known since the 2020 DeFi Summer, a former penetration tester who now audits model pipelines — posted the headline at two in the morning with a single line beneath it: “Read the footnote they deleted.” There was no footnote. That was the point.
I have spent twenty-three years watching what happens when a precise finding is repeated by people who never read it. In 2017, at the height of the ICO boom, I audited the smart contract logic for a data-provenance startup. The team wanted to ship to mainnet to catch the market, and I refused to sign off, because the encryption protecting user metadata did not meet the standard the marketing already promised. Five critical vulnerabilities. None of them hypothetical. The founders shipped anyway, and the metadata leaked, and the users who trusted them paid for a headline they never read. The “40%” is a different kind of leak — a leak of meaning. And meaning, in a system that rewards noise, is always the first casualty.
What the IMF actually wrote is more careful than what it became. In its January 2024 work on generative AI and the future of work, and in the accompanying commentary from its managing director, the fund placed roughly 40% of global employment in a state of “exposure” to AI. The number scales by income: about 60% in advanced economies, about 40% in emerging markets, about 26% in low-income countries. The word that carries all the weight is exposure, and exposure means something narrow and specific. It means the tasks inside a job overlap with what AI can do. It does not mean the job disappears. It does not even mean the job changes. It means the job is now in the room where the question is being asked.
And the IMF asked the question with a split answer. Within the exposed jobs of advanced economies, the fund estimated that roughly half may be enhanced — the human absorbed into a faster loop, the work augmented — while roughly half may see labor demand fall. One number, two destinies, and the headline kept only the number.
There is a second figure in the same report that almost no one repeats, and it may be the more important one. The IMF maintained its expectation for global growth near 3%. Not 5%, not 4% — the same mid-single-digit ceiling that has anchored the world's potential output for a decade. Read the two numbers together and a different story appears: a world that may reshape nearly half its exposed labor while its measured growth stays flat. That is not a story about a productivity revolution. That is a story about where the productivity goes.
I care about this because I work in a field that claims to solve exactly this problem — who gets to verify a claim, who gets to capture a gain, and who is left holding a headline. The IMF is a centralized oracle. Its data is authoritative and its process is opaque, and at the point of citation it becomes unauditable. That is the precise failure mode the first blockchains were built to address, and it is the precise failure mode we keep reproducing. A finding leaves a research department, passes through a wire service, lands in a feed, and arrives in a community with its qualifier sanded off. The chain of custody is broken, and no one is held to account, because there was never a ledger.
Consider what “exposure” actually measures and what it hides. Exposure is a statement about tasks, and tasks are the atoms of a job, but a job is not the sum of its tasks. A job is a bundle of tasks plus judgment plus accountability plus the trust that someone will be answerable when the bundle fails. The IMF's method, like every method in this space, is honest about its own limits — it measures overlap, not outcome. The media, and the market, and my own community, are less honest. We take the overlap and we call it a verdict. The distance between “exposed” and “disrupted” is the entire distance between a measurement and a prophecy, and we cross it casually, because prophecy sells and measurement does not.
I have watched this slippage in my own work. In a smart contract audit, an informational finding can become a critical finding in a founder's pitch deck, and a critical finding can become a “minor optimization” in a fundraise memo. The finding does not change. The framing does. And framing, at scale, becomes policy. When a legislator reads that “AI will disrupt 40% of jobs,” she drafts a different bill than she would if she read that “40% of tasks overlap with AI capability, and half of those may be augmented.” One sentence funds retraining. The other funds a wall.
Consider the half the headline drops: the augmented half. In the exposed jobs of advanced economies, the IMF's own split says about half may be enhanced. I have felt this in my own work. The tools I use to read a smart contract have changed; I do not write less code, I read more of it, faster, and I catch things I would have missed. But augmentation is not a free lunch, and this is the part the optimists skip. Augmentation raises the floor of competence, which means it also raises the bar of what counts as competent. When everyone can draft a contract, the value moves to the person who can tell whether the draft is dangerous. When everyone can generate a test suite, the value moves to the person who knows which test matters. Augmentation does not eliminate judgment; it concentrates the value of judgment into fewer hands, which is the opposite of what its cheerleaders promise. I have seen this pattern before, in the audits I have run: the tooling improved, and the number of auditors who could be trusted with the tooling did not.
The distribution problem underneath the number is the one I cannot stop thinking about, because it is the problem I audit every week in a different costume. If global growth holds near 3% while AI reshapes the exposed labor market, then the productivity gains are not showing up as broad-based growth. They are showing up somewhere narrower — in margins, in valuations, in the balance sheets of whoever owns the models and the compute and the distribution. That is the same concentration pattern I have traced through a hundred token distributions. A protocol launches, a small set of insiders and early capital capture the supply, the narrative says the value is “for the community,” and the ledger says otherwise. A world that grows at 3% while reshaping 40% of its exposed labor is a world where the gains accrue to whoever owns the models — which is the same concentration pattern I audit in token allocations, only denominated in jobs instead of tokens.
I have watched three generations of orderbook DEXs promise to dethrone the centralized exchanges, and I have watched every one of them meet the same wall. A market maker will not leave a resting quote on a public chain where it can be picked off in the time it takes to sign the next block. Latency is not a bug in market structure; latency is the structure. The reason this matters here is that the same logic governs where AI's gains settle. Capital flows to the venue where it can capture the spread with the least friction, and the least friction is almost never the most open venue. The productivity surplus will pool in the places that can move fastest and capture most reliably — and those places are, by construction, the most concentrated ones. Openness is a virtue. Openness is not a competitive advantage in the capture of a spread.
Now look at the geography, because the geography is where the comforting reading of the data does the most damage. Low-income countries show the lowest exposure — about 26% — and it is tempting to read that as a kind of safety. It is not safety. It is exclusion wearing the costume of safety. A low exposure rate in a low-income economy is a measurement of how little of that economy has entered the white-collar, digitizable, AI-addressable layer of work at all. These economies are not shielded from AI. They are simply not yet in the room. A low exposure rate is not a shield; it is a mirror of exclusion, and the danger it hides is not replacement but abandonment — the risk of being left outside the room where the value is being redefined.
This is where the decentralization narrative I have spent my life inside has to be honest with itself. We say permissionless, and we mean it, and yet we have built, in the space of a few years, more execution layers than we have users to fill them — dozens of rollups sharing the same finite liquidity, each one insisting it is the future, each one slicing the same small base of real activity into thinner fragments. That is not scaling. That is fragmentation wearing the language of abundance. And it is the exact pattern the AI-employment data is warning about at civilizational scale: a capability that concentrates rather than diffuses, dressed in the vocabulary of universal access.
So what does the IMF propose? Education and tax. Retraining on one side, redistribution on the other — capability building and outcome correction, the two classic instruments of a policy response. Both are necessary and both are incomplete, and the incompleteness is structural rather than technical. The IMF is a Bretton Woods institution; it can warn and it can recommend, and it cannot compel a single member state to do anything. Its power is the power of a number repeated until it becomes a norm. Which means its most powerful instrument is also its most fragile one: the accuracy of the number itself.
And here the report is silent in a way that matters. It does not name the EU AI Act. It does not name the executive orders or the model registrations or the disclosure regimes that were, by early 2024, the actual machinery of AI governance. It speaks of education and tax in the abstract and never descends to the mechanism. This is the standard shape of international-organization narrative — a problem stated fully and a solution stated emptily — and it is worth noticing, because the emptiness is not an oversight. It is the price of consensus among members who do not agree on what should be done.
There is a deeper silence, and I want to name it precisely, because it is the one my own industry is best positioned to fill and worst positioned to be trusted with. The IMF frames the response as education and tax. Notice where the responsibility lands: on the worker to retrain, and on the state to redistribute. Neither frame asks the model owners to internalize the cost of the transition they are causing. That is not an accident of phrasing. It is a choice of framing that moves the burden away from capital and toward labor and the public purse, and it is the oldest move in the book. The test of any AI policy is not whether it mentions retraining; it is whether it names who pays for the retraining, and the IMF's framing quietly answers that question in favor of the parties least able to refuse. In 2024 I drafted, with a European legal firm and two senior developers, a framework for ethical staking governance, and the exercise taught me that the hardest part of compliance is never the rule — it is the incentive that makes the rule survivable.
This is where the crypto question stops being decorative and becomes load-bearing. If the transition is real, it has to be financed, and financing requires identification — of who was displaced, of who is entitled, of who contributed what. Every one of those is an identity problem. And every identity system is a surveillance system until someone proves otherwise.
Which brings me to what I have actually been building. By 2026, as autonomous agents began transacting on-chain without a human at the keyboard, the question of who — or what — was behind a wallet became existential for governance. A DAO that cannot distinguish a person from a swarm of agents is a DAO that can be captured by whoever runs the most bots. So a small team of five researchers and I built a system we called Verifiable Humanhood: a way to prove that a unique human is present without exposing which human it is, using zero-knowledge proofs to separate the fact of personhood from the data of the person. The point was never to surveil. The point was to make contribution legible without making the contributor transparent.
That distinction is the whole ballgame, and it is exactly the distinction the employment debate is missing. To compensate a displaced worker, you must be able to verify the claim of displacement. To fund retraining, you must be able to route the funds to a verified person. To design an AI capital-gains tax or a robot tax or a negative income tax — the three mechanisms the IMF gestures at without naming — you must have an identity layer that can distinguish a human entitlement from an automated claim. The IMF names the need. It does not name the tool. And the tool it does not name is the one my field has been quietly building and quietly getting wrong.
I want to be careful here, because the last time this field tried to solve a governance problem with an enforcement tool, the tool became the crime. When sanctions fell on a privacy protocol and its developers, the message the industry received was not “govern the use.” The message was “writing code can be the offense.” Code is law, but conscience is the interpreter, and a system that indicts the interpreter has stopped governing and started punishing. If the response to an ungovernable technology is to indict the people who wrote it, then the response to an unauditable employment shock will eventually be to mandate an unauditable identity — and we will have traded one invisible system of control for another.
I spent three months in 2022 in near-total withdrawal after the collapses of that year, and I did not emerge with a better prediction. I emerged with a better question. I had watched trusted institutions fail from the inside, and I had watched the idealists who believed in them break, and I stopped asking which protocol would win and started asking which claim could be checked. Solitude is the only auditor that never sleeps. It is also the only one that will tell you, without flattery, that you have been repeating a number you never verified.
There is a market reading of all this that the sideways tape makes visible. The IMF holds growth near 3%, which is to say it does not price an AI productivity boom into its baseline. The equity and venture markets, meanwhile, price one with great enthusiasm. That gap — between the baseline that does not assume the boom and the valuation that fully assumes it — is the same divergence I watch on a flat screen in a quiet market: fundamentals and expectations trading at different prices, with the truth arriving later. In a consolidation, the honest question is not which narrative wins but which asset has to be repriced when the narrative is checked. The IMF did not check it. It only declined to assume it. That refusal, quiet and unglamorous, may be the most valuable data point in the whole report.
Here is the contrarian turn, and I want to state it plainly because it will make me few friends on either side. The instinctive fix for the identity gap — a universal, verifiable, biometric-anchored proof of personhood — is also the most dangerous instrument this decade could build, and the AI-employment debate is about to hand it a justification it did not have before. Frame 40% of jobs as exposed, frame the response as compensation, and you have manufactured a reason to enroll every human on earth into a verified registry of who they are and what they are owed. The people who build that registry will tell you it is privacy-preserving. Some of them will even mean it. But a system that can prove you are a person can also prove you are not one, and a system that can route your compensation can also freeze it. The loudest voice in the room will be the one promising to protect workers from AI, and the loudest voice is rarely the most aligned.
So what is the pragmatic test? Not whether a framework sounds ethical, but whether it survives contact with a bad actor. Ask of any AI-employment policy: who can forge a claim inside it, who can capture it, and who can weaponize it against the people it claims to serve. If the answer is “no one, because the ledger is auditable,” it may be worth building. If the answer is “trust us,” it is a promise, not a system, and promises are exactly what the “40%” already taught us not to trust.
The takeaway is not that AI will not change work. It will, and it already is. The takeaway is that the number everyone is repeating is a measurement wearing the mask of a prophecy, and the institutions repeating it have no mechanism to verify what they say. Before we build a global identity layer to compensate the exposed, we should build the audit layer that can prove who was exposed, who paid, and where the surplus went. The chain of custody comes first. The ledger comes first. Everything else — the tax, the retraining, the compensation, the dignity of the displaced — is downstream of a single question we have not yet answered: when a claim about human work is made, who checks it, and who is accountable when it is wrong?