A guide for policymakers and researchers: what the research finds about AI's effects on work, and how well the evidence backs each policy aimed at them. Every grade links to its sources and was independently fact-checked.
Gradesworksdoesn't workunproven: too little evidence to call
Issues are grouped by the stage of working life they touch: preparing for work, getting and losing work, on the job, and across the economy. Each tracked channel collects the studies that estimate AI's effect, labeled by type: peer-reviewed, working paper, government data, think tank or industry research. Dashed tiles mark issues that matter but aren't tracked yet, so the list doesn't pass for complete.
Each channel ends with a line on who captures the gains and who bears the costs, firms or workers, and which workers. It draws only on the studies in that channel.
A policy can be well evidenced on one issue and untested on another, so the grade sits on each policy-issue pair, not on the policy. Each grade answers one question: is there evidence it improves that outcome for workers? Every grade lists the sources it rests on. The rule:
| Strong | At least 2 causal studies (randomized, or a credible quasi-experiment) in different settings find it improves this outcome, and no credible causal study finds the opposite. A positive meta-analysis of causal studies counts. |
|---|---|
| Moderate | 1 causal study, several consistent non-causal studies, or causal evidence from a closely analogous earlier program finds it improves this outcome. |
| Weak | Credible evidence finds it doesn't improve this outcome, even if it helps something else. |
| Thin | Evidence only on the mechanism, a loosely related program or a different outcome, or mixed results. |
| Untested | No relevant evidence. |
The map groups these into three: works (strong or moderate), doesn't work (weak) and unproven (thin or untested).
Each issue lists 9 to 13 studies that estimate AI's effect on that outcome, or earlier automation's for comparison, using data or a stated design. Newer studies and stronger designs come first. Commentary is left out unless it's the source of a key number. Industry and advocacy research is included and labeled, and a Stake note flags any study whose authors or funders have a stake in the answer. Each study is tagged by design (experiment, quasi-experiment, descriptive, survey, model or review), with its data, country and period. Working papers change, so the version checked is noted.
Use "Cite this site" in the footer, or "Cite" on any study. Every study list exports to a spreadsheet, BibTeX or RIS.
Each policy opens as a one-page brief: the bottom line, where it would live in law, cost, live bills, the strongest case against, drafting notes and, for unproven ideas, how to build evidence into the bill. Each grade is tagged by what it rests on: AI studies, earlier programs, models only, or nothing yet. Bill statuses are checked as of the date in the footer.
Where a policy has been written into law, a full test record traces the chain of assumptions it relies on and marks each link:
Records grade two things separately: whether the policy works as designed, and whether it changes outcomes for workers.
Statute text, agency data, government audits and peer-reviewed research come first, then law-firm and think-tank analysis for status updates. Advocacy analyses are labeled as advocacy, whichever side produced them.
Claims from worker advocates and from employers get the same treatment. I spent 2022 to 2025 in Amazon's public policy team working on several of these issues, so industry claims are tested here at least as hard as anyone else's.
Updated as new research comes out and bills move. The update date is shown at the top of each page.