Research: AI Labor Displacement and the Managed-Dependency Class

A research position. Sourced, present-all-sides. Named people and organizations carry primary citations; characterizations are attributed to whoever made them, never adopted as this page’s own voice. This is the socioeconomic seam of the series — where the capability that displaces (Evil Robots) meets the grid that manages the displaced (The Ratchet). Where a claim rests on a source, the source is cited; where an intent is at issue, it is attributed to whoever asserted it, and never stated in our voice.

The load-bearing observation is not that the robots are coming for every job. It is narrower and harder to dislodge: automation exposure is not evenly distributed. It concentrates by task and skill tier, and it lands hardest on the parts of the workforce — and the parts of the world — least able to absorb the shock. The people displaced do not evaporate. They become a population that has to be managed. The management system is the one this project already maps: the funding ratchet, the welfare rail, the surveillance layer. That is the whole argument compressed to a sentence — the capability makes the transfer necessary, and the rail makes the transfer a leash.

The exposure gradient

The exposure is graded, not universal, and the gradient is the whole point.

The IMF put a number on it. In a January 2024 staff note and an accompanying blog by managing director Kristalina Georgieva, the Fund estimated that roughly 40% of jobs worldwide are exposed to AI — about 60% in advanced economies, roughly 40% in emerging markets, and about 26% in low-income countries — and warned the effect “is likely to worsen inequality” (IMF blog, Georgieva, 14 Jan 2024, archived; CNBC on the 40% / inequality framing).

The developing-world twist answers the obvious objection — that the poorest countries, with the fewest exposed jobs, are insulated. They are not. In the IMF’s framing, low-income countries have lower direct exposure but “don’t have the infrastructure or skilled workforces to harness the benefits,” so over time AI widens global inequality: they get neither the substitution shock’s upside nor the productivity gains. Exposure without absorption capacity is the worst square on the board.

The ILO’s joint work with the World Bank sharpens the same finding. Its refined global index (Working Paper 140, 2025) estimates that about one in four jobs globally is exposed to generative AI — most likely “transformation, not replacement” — with exposure running 34% in high-income versus 11% in low-income countries, but with the warning that “in developing economies, disruption may materialize faster than productivity gains due to existing digital gaps” (ILO / World Bank Working Paper 140). The global south gets the harm before the benefit. That ordering — harm first, benefit maybe, later, if the infrastructure ever arrives — is the mechanism by which a productivity story becomes a redundancy story for a specific tier.

The mechanism — why displacement need not be replaced

The optimist’s rebuttal is historical: automation has always destroyed jobs and always created more. The rigorous version of the pessimist’s answer is that the second half of that sentence stopped being reliably true.

Daron Acemoglu and Pascual Restrepo, in “Tasks, Automation, and the Rise in U.S. Wage Inequality” (Econometrica, 2022), document a robust negative relationship between direct task-displacement and real wages across worker groups: automation reduces the labor share and can depress overall wages and employment. Their inflection point is the load-bearing finding. From 1947 to 1987, displacement of about 17% was offset by reinstatement of about 19% — the new tasks that historically rescued the displaced kept pace. From 1987 to 2016, displacement of about 16% was met by reinstatement of only about 10% — the rescue mechanism weakened by half. Layered on top is their notion of “so-so technologies”: cost-saving automation that displaces workers without raising productivity enough to justify new hiring, so firms neither expand employment nor raise pay (Econometrica via DOI; MIT News summary).

This is the engine behind “fewer roles for the lower tier,” stated without the crude shorthand: the reinstatement that once absorbed displaced workers has weakened, and a large share of deployed AI is exactly the so-so, labor-share-cutting kind. The displaced tier is not a moral category or an IQ threshold; it is the set of tasks most exposed to substitution and least served by whatever new tasks the technology creates.

The famous, contested headline

The number most people remember is contested, and it belongs in the record precisely so it can be handled honestly.

Carl Benedikt Frey and Michael Osborne (Oxford, 2013) estimated that 47% of US jobs were at high risk of automation within ten to twenty years. It is the figure that launched the modern debate — and it has been heavily criticized. The Information Technology and Innovation Foundation, reviewing the record in 2022 (“Oops: ‘The Predicted 47 Percent of Job Loss From AI Didn’t Happen’”), noted that the mass job loss did not arrive on schedule and that the correlation between the computerization-risk score and actual subsequent job loss was weak, on the order of 0.26 (ITIF critique). The 47% belongs on the page as the famous-but-contested figure, never as fact. The durable signal is the direction — task exposure graded by tier — not the headline number, which the intervening decade did not vindicate.

The management thesis — dependency as control

Here the analysis moves from documented fact to attributed thesis, and the line is kept bright.

The historian Yuval Noah Harari has argued that as AI absorbs jobs, a large class becomes economically irrelevant — a “useless class,” in his phrase — maintained by subsidies, algorithmic entertainment, and soft surveillance while a smaller, augmented elite concentrates power. That is Harari’s thesis, a contested one, attributed to him and not adopted here. The move that matters for this project is structural: the remedy Harari describes is itself a management system — income support plus pacifying content plus a legible population.

The remedy is not hypothetical, and it is proposed by the builders. Sam Altman has put money behind it: OpenResearch, the study organization he backed, ran a $1,000-a-month, three-year guaranteed-income trial paying out roughly $45 million to thousands of recipients across Illinois and Texas (CBS News on the Altman-backed OpenResearch study). Elon Musk and the philosopher Nick Bostrom have likewise, in their own public statements, floated Universal Basic Income as the answer to automation-driven displacement. The pattern is the tell: several of the people building the systems that displace also propose the dependency that would manage the displaced.

The synthesis this project makes — and marks as its own — is that a UBI-dependent, economically redundant class, paid through a programmable rail, entertained by recommender systems, and legible to the surveillance layer, is exactly the population the control grid is built to run. Dependency is the pawl that does not release: once income is a grant on a rail the recipient cannot exit, the recipient’s compliance surface approaches total. That reading is ours, attributed, and it depends entirely on the next distinction holding.

The pilots rebut the laziness trope

The honest evidence sharpens the thesis rather than sloganizing it, and it does so by splitting one loaded question into two separable ones.

Does a basic income create idle dependency? The trials mostly say no. The Altman-backed OpenResearch study found that recipients did not drop out of the workforce — they worked slightly fewer hours, spent more on basic needs and on helping others, and used more health care (including a rise in hospitalizations and emergency-department visits, read as deferred care finally sought). The effect was modest, not transformative, and there was no collapse into idleness (CBS News on the OpenResearch findings). Stockton’s SEED program — $500 a month for 24 months, unconditional — reported that recipients reached full-time employment at roughly twice the control rate, with better mental health and lower income volatility (University of Pennsylvania SEED evaluation). GiveDirectly’s Kenya program — the largest and longest, reaching roughly 23,000 people across 195 villages, some for twelve years — found the money went to productive assets like livestock rather than “temptation goods,” and improved food security and health, though it also “induced increased risk exposure” (GiveDirectly 2023 UBI results).

The “UBI makes people lazy dependents” trope is not supported by the trials. Say that plainly — do not defang the concern, and do not overclaim the harm. Which raises the second, sharper question.

The rail, not the money

So where is the control? Not in the transfer — in the rail it rides. Dependency-as-control is orthogonal to the laziness question. It is not about whether people work; it is about how the money is paid.

On a programmable central bank digital currency, money can carry conditions in code — expiry dates, vendor and zone restrictions, category limits, automatic deductions. This is not speculative. The documented proof-of-concept is China’s e-CNY subsidies distributed with 90-day expiration windows enforced by embedded code: money that deletes itself if not spent on schedule. Analysts warn that programmable money becomes “a single point of control” and a “surveillance tool” — “no spending outside designated zones, no purchases from unapproved vendors” — the conditionality being a feature of the rail, not a bug (CCN, “Why CBDCs aren’t just digital cash”; Finextra, “Programmable Money: CBDCs and the New Era of Policy-Driven Payments”).

Put the two findings together and the thesis stops being a slogan. A basic income can be benign — even beneficial, as the pilots show — and still become a control surface the moment it runs on a rail the payer can program and the recipient cannot exit. The displacement makes the transfer necessary; the rail makes it a leash. That is the precise, non-hysterical form of dependency-as-control, and it ties directly to the money layer mapped in CBDC research and The Ratchet’s “The Money” chapter.

Where it converges

Restated economically, this is the series’ own arc: Evil Robots capability → mass displacement → The Ratchet management grid → the human cost the closer, The Last Antibodies, tallies. The capability displaces the lower tier and, disproportionately, the developing world. The reinstatement that once rescued the displaced has weakened (Acemoglu-Restrepo). The proposed remedy — income support on a programmable rail, plus the algorithmic pacification that is Quiet Autocomplete’s domain, plus a surveillance layer that renders the recipient legible — is not a safety net so much as a management system, on the attributed reading above.

The convergence is not incidental. The same funders and intergovernmental bodies that set the “AI and the future of work” agenda are the ones underwriting the governance instruments and the UBI pilots, so the leverage graph and the displacement thesis describe one apparatus, not two. The moving parts — the exposure gradient, the weakened reinstatement, the programmable rail, the governance ratchet — are tracked side by side in the convergence table, which lays each mechanism against the others so the seam is visible in one view.

Related research

• CBDC — the programmable rail this thesis turns on; the money layer that converts a benign transfer into a conditional one.
• The AI Governance Ratchet — the instrument inventory that governs the capability doing the displacing.
• The Body Layer — the endpoint of the managed-dependency arc, where the control grid reaches the person rather than the paycheck.
• The China Parallel — the Body-Control Stack — the same programmable-rail logic (e-CNY, social-credit adjacency) shown in its most developed state form.
• Convergence Table — the live cross-map where displacement, the money rail, and the governance ratchet line up as one system.
• AI Governance Tracker — the implementation calendar for the governance half of the apparatus.