Autonomy Index
A weekly read on how far AI has actually taken over human work — scored across fifteen sectors by how much genuinely runs without a person in the loop, rather than by what models can do in a demo. Then what that shift could mean, and what is worth doing about it.
no research data change since 2026-07-29
Global AI Autonomy index
ⓘReading the 60% score: it is the unweighted mean of the 15 sector scores. The gap between GDP-weighted (63%) and employment-weighted (55%) ≈ the gap between "AI has captured the high-value work" and "AI has captured most people's jobs." Watch it widen.
Path to society-wide AGI
Hover a phase for the short version, tap for the full one.
Phase definitions are fixed in the generator, not the weekly state — they are the measurement ruler, so they do not move with a rescore. Capability/autonomy anchors follow DeepMind's "Levels of AGI" (Morris et al., 2023).
Sector autonomy — C · A · B · R breakdown
Capability · Autonomy · Breadth · Reliability, each 0–5. Score = mean × 20. The signal that matters most is A or R rising while C stays flat — trust giving way ahead of capability.
Bubble size is breadth of real deployment. Sectors drifting up faster than they drift right are the early warning. Numbers match the table below — click either a bubble or a row for the full assessment.
Scores are Autonomy Index estimates against the definitions in the task spec, moved only on dated, citable evidence. Trend sparkline shows this sector's score across every state file to date (2 so far).
Forecast — time to full AGI
Index trajectory from 60% today to the Phase 3 (70%) and Phase 4 (85%) thresholds. "Full AGI" = society-wide autonomous operation, not a lab announcement — so this stacks a deployment lag on top of the capability forecast. Bands, not dates.
Capability clock from early-2026 forecaster medians for transformative AI; deployment lag is Autonomy Index's. Every revision Jan–Apr 2026 moved sooner, which is why the Fast band stays on the board.
Update log
Substantive changes only — sector re-scores, index moves, phase or trigger changes, and major external events. Cosmetic edits to the page are not tracked. Each run's evidence is grouped; tap a group to read it.
- 2026-08-24Weekly rescore: no material change this week; scores held. All 15 sectors carried forward unchanged; index unweighted 60% (Phase 2)
- Stanford Digital Economy Lab, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence", rev. 2026-08-12, ADP payroll data through June 2026: employment of 22-25 year olds in AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers, a divergence widening steadily since August 2025, operating through reduced hiring rather than increased separations, and concentrated where AI substitutes rather than complements. The same paper's first stated fact is "no evidence of widespread, economy-wide job displacement"
- Challenger, Gray & Christmas July 2026 report (2026-08-06): AI the leading stated reason for job cuts for a fifth consecutive month (10,970 in July, 33% of the month; 112,713 YTD, 24%) - but YTD cuts total 477,033, down 41% from 806,383 a year earlier, July's 33,429 is the lowest monthly total in two years, and YTD hiring plans of 107,500 are up 25% YoY, led by aerospace/defence, technology and automotive. Treated as genuine counter-evidence, not noise. Together these show entry-level compression at exactly the predicted first-appearance site without an aggregate downturn, so displacement_inflection holds at ENTERING rather than escalating
- EU Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the OJ 2026-07-24, in force 2026-07-27: Annex III stand-alone high-risk obligations deferred from 2026-08-02 to 2027-12-02, Annex I product-embedded to 2028-08-02, public-authority systems to 2030-08-02; Article 14 human-oversight duties unamended in substance and Article 50 transparency duties took effect 2026-08-02 as originally scheduled. This defers when the brake bites rather than removing supervision now, so no A or R move - but it is the flagged trust-before-capability pattern and is on watch across EU-exposed sectors (healthcare, autonomous transport, public administration, education, legal)
- no citable step-change found, last verified 2026-08-24: scientific research (self-driving labs still operating at level 2-3 of 5, humans set goals and handle exceptions - consistent with A2/B2); public administration (agentic pilots and a Gartner "80% of governments by 2028" forecast only, no autonomous statutory decision-making); customer service (median enterprise tier-1 deflection ~41%, top quartile ~59% - consistent with B4/R3); financial markets (FCA Mills Review published July 2026 is regulator attention, and institutions remain on bounded-autonomy models rather than autonomous strategy-setting). Healthcare diagnostics, logistics/autonomous transport and military ISR were checked in the daily window to 2026-08-24 with no first-of-kind event. Autonomy cascade remains NOT OBSERVED. TWO ITEMS
- with re-verification dates (see unverified[]): China's agent framework (re-verify by 2026-09-27) and the vendor-reported agentic benchmark jumps (re-verify by 2026-09-21)
- the EUR-Lex ELI page for Reg (EU) 2026/1744 returned empty on direct fetch and consilium.europa.eu returned HTTP 403, so the regulation was confirmed against an independent law-firm briefing rather than the OJ text itself; the Stanford PDF would not parse, and its figures were taken from the publisher's own publication page
- 2026-07-28Baseline established. Global index 60% — Phase 2 (Delegation). Highest: financial markets 85%. Lowest: public administration 45%. Displacement-inflection trigger ENTERING; autonomy cascade not observed. No sector re-scored yet.
- 2026-07-13Event: 200+ economists, incl. 16 Nobel laureates and the chief economists of OpenAI and Anthropic, publish the "We Must Act Now" statement (organised by Anton Korinek), warning AI could transform the economy on a far shorter timeframe than the Industrial Revolution. Corroborates the central-to-fast forecast; no score change.
Transition risk — what could go wrong, and how badly
The downside set, placed by assessed likelihood and severity. Position is editorial judgement reviewed on the weekly run — not a measured value. Tap any item for why it sits where it does.
Likelihood →
Placement derived from the transition-risk sources below (Acemoglu & Restrepo on task displacement; Korinek & Stiglitz on AI and income distribution; Caprettini & Voth on Swing-riot unrest), plus AI-safety work on oversight failure.
Trigger watch — the surprises to catch early
Everything else on this page is a gradient. These two are discrete — the things that would change the picture abruptly.
States come from the weekly state file. The meter renders the state word; it is not a separate score.
What it could mean for society
The two columns are not ranked or scored — there is no measurement behind a claim that one of these is twice another. Tap any item for what it means, what it would do, and the caveat.
▲ Upside
▼ Downside
Oversight risk (separate axis): confidently-wrong models acting unchecked · systems optimising conflicting goals · correlated failures from everyone running similar models · lost human expertise to catch any of it.
*The upside is conditional on the gains being distributed; the downside is largely what happens if they are not.
Framing draws on the economics of automation and inequality (Acemoglu & Restrepo on task displacement; Korinek & Stiglitz on AI and income distribution) and AI-safety work on oversight failure. Autonomy Index synthesis.
Transition risk — the dangerous gap
The risk is not the end state — it is the lag between two things moving at different speeds. Displacement moves at the speed of deployment. Redistribution moves at the speed of politics. Upheaval concentrates in the window where the first has outrun the second, and the three readings below are how far apart they currently are.
The first two move with the weekly rescore. The third has not moved since this monitor began — that distance is the gap, and it is why the risk above reads elevated.
What history suggests▶
In Britain's Swing riots (1830–32), parishes exposed to labour-replacing threshing machines rioted at roughly 20% against 13% elsewhere — concentrated where displacement was fastest and no alternative income existed. The wider industrial shift carried "Engels' pause": around fifty years in which output rose but real wages did not. New jobs did appear, but adjustment took a generation and the displaced were mostly not the people who got them.
Why this transition may differ▶
Three factors raise the risk: it is broad (most cognitive sectors at once, not one trade), simultaneous and global rather than a slow region-by-region wave, and fast — years rather than the decades industrialisation took. The historical safety valve was migrating into new sectors, which needs both time and somewhere to go.
This is broadly the view behind the July 2026 "We Must Act Now" statement (200+ economists, 16 Nobel laureates). It is not settled: a substantial camp still expects decade-scale diffusion as reliability and adoption frictions bite, which is why the Slow scenario stays on the board.
How income gets redistributed — the unsettled question▶
UBI trials in Kenya and Finland are encouraging on behaviour — recipients did not stop working — but they were poverty-scale, not full income replacement, and never tested funding a population once wage-tax bases collapse. And should a former executive receive more than a former cleaner?
Once labour is redundant, income based on labour has no anchor. Durable settlements therefore turn on ownership of capital, not wages for work — which is why who owns the models and compute is the variable to watch.
Sources: Caprettini & Voth, "Rage Against the Machines" (AER: Insights, 2020); Allen on "Engels' pause"; GiveDirectly Kenya and Finland basic-income trials; Alaska Permanent Fund and Norway's sovereign-wealth fund; Korinek et al. "We Must Act Now" (2026).
Where in the world — AI preparedness
Shaded by the IMF's AI Preparedness Index — a composite of digital infrastructure, human capital and labour policy, innovation and economic integration, and regulation. The scores are the IMF's; the reading of them below is this monitor's.
167 of the world's economies carry a score. Hover or tap any country for its reading; grey countries are not in the index.
Preparedness is not safety. The best-prepared economies are also the most exposed, because both rise with the share of work that is digital and cognitive. A high score means arriving first with the means to respond; whether those means get used is political.
Coarse framing, not relocation advice — legal status, family, cost, language and existing ties matter more than any index, and a benign transition makes most of it moot.
Scores: IMF AI Preparedness Index (DataMapper indicator AI_PI, 2025 vintage), retrieved 2026-08-25 and stored in the weekly state file. Background: IMF "Mind the Gap" (2025) and Gen-AI: AI and the Future of Work (2024). Country geometry: Natural Earth 110m (public domain).
Adaptation — response scaled to the signals
Contingencies matched to what the monitor sees, not predictions. Most of the early moves are no-regret — worth doing even if the transition stays benign. The point is to buy optionality early and only escalate as specific triggers fire, so you act proportionately instead of reacting to a surprise. The "move to land and water" idea sits at the far end of this ladder — a contingency you could scale into, not a starting step.
Three scenarios — a continuum
These shade into each other, and the monitor does not claim to know which one arrives — Turbulent is treated as the central case. The marker is a different thing: it shows which rung of the response ladder the trigger states currently justify. Select a scenario to see the response it would call for.
"Own AI-stack equity" means broad exposure to the firms that own the models and compute — the personal version of an ownership dividend. Context, not financial advice.
Provenance: the three scenarios and the response ladder are Autonomy Index's own synthesis — they are not drawn from an external framework, and the "current read" is an editorial judgement rather than a computed output. What is external is the underlying evidence: the no-regret / optionality framing comes from resilience and disaster-preparedness research, and the escalation tiers are keyed to the displacement-inflection and autonomy-cascade trigger states on the Risks tab, which do move with the weekly rescore.
Work — what is exposed and what is durable
The question people actually ask. Ordered left to right from durable to exposed, and every judgement is tied to a C/A/B/R score on the Status tab — so it moves when the evidence moves rather than when the mood does. Tap any card for the reasoning.
The meta-move beats picking a "safe" job: position to direct AI rather than compete with it, and hold capital.
"Safe from AI" is the wrong frame. Almost no job is fully automated and almost none is untouched — what changes is the mix of tasks inside a role.
Durability judgements are Autonomy Index's, keyed to the sector scores on the Status tab. Entry-level compression figures from Stanford Digital Economy Lab (ADP payroll data through June 2026).
Daily needs and where you live
Resources
The four things that decide how well a household absorbs disruption. Tap any for what to actually do.
Settlement gradient
There is no best answer — each ring wins in a different scenario, which is exactly the point. Tap a ring.
The pattern across all three: connected local systems beat both full dependence and full self-sufficiency.
Community-resilience and food-security research; settlement gradient is Autonomy Index's synthesis.
Estimates for review, not measured values. Weekly re-scoring + event-driven alerts. Last updated 2026-08-24. · Corrections and questions: contact@autonomyindex.org