Methodology

Each ledger entry grades the claim that AI caused the job cuts - not whether the layoff happened, and not the company's conduct. Grades are assigned by a human editor from primary sources (company statements, filings, regulator notices), with reputable reporting as corroboration. Every grade carries a written rationale, its evidence links and the date it was graded.

Rules we apply

• Top third of exposure means Anthropic observed exposure of at least 4.9% (the 67th percentile across 756 occupations). Because 54% of occupations score zero, the top third starts low.
• Headcounts are what the company stated or, failing that, credible reporting; we mark undisclosed numbers rather than estimate them, and note ranges.
• AI mechanism tags separate different stories: AI said to do the work; AI among several efficiency reasons; AI cut demand (e.g. search traffic); Cuts fund AI investment; Reorganising around AI strategy; Company denies AI link. A demand shock (e.g. AI search reducing publisher traffic) is AI-related but is not AI doing the work.
• Rehired since records reports of companies re-hiring for roles they said AI replaced.
• Status. Every entry on the site has been reviewed and approved by the editor before publication. Grades are re-dated when changed.
• Linking an event to occupations is our editorial mapping of the roles reported, not a company statement.

How events are found

An automated pipeline produces candidates only; nothing is published automatically. Candidates come from SEC EDGAR full-text search of 8-K filings (including Item 2.05 restructuring disclosures mentioning AI), California and Texas WARN notices, and news headlines found through the open GDELT Project DOC 2.0 API. For news we keep only the headline, link, date and publisher, link to the original articles, and never republish their text. An editor then checks each candidate against primary sources and approves it before it appears here.

We use Anthropic's observed exposure measure (Massenkoff & McCrory, Labor market impacts of AI, March 2026), published in the Anthropic Economic Index. For each O*NET task that a language model could theoretically make substantially faster (the Eloundou et al. β measure), it asks whether the task appears in significant work-related Claude usage (Aug and Nov 2025 samples). Fully automated use is weighted fully, assistive use half, and task scores are averaged to occupations weighted by time spent.

• It is a measure of current use of one AI system, not a forecast and not a probability of job loss. Other AI tools are not observed.
• Scores are for US SOC occupations; we apply them to UK roles via the mapping below.
• Anthropic's own results: no systematic rise in unemployment for highly exposed workers since late 2022; tentative evidence that hiring of 22-25 year-olds into exposed occupations has slowed.
• Task-level lists on occupation pages use Anthropic's task penetration figures, matched to O*NET 31.0 task statements by exact text.

• UK, by occupation: ONS "Online job advert estimates" (new adverts from Textkernel), table 3, monthly by 4-digit SOC 2020, not seasonally adjusted. Latest publication 22 October 2026, covering Jan 2018 - Sep 2026. We show year-on-year change over the last six months, comparing only months published in both years. Notes from ONS: Note text Some shorthand is used in this workbook, [x] = Data are unavailable. Textkernel has updated their methodologies and taxonomies throughout the timeseries back to 2017. This results in changes to classification of locations, which can produce some step-changes in the data. Data for October 2025 to February 2026, and June to September 2026, have been partially imputed. This is because we were unable to obtain accurate information from a major source for this period. Data for March 2026 has been suppressed because of issues with obtaining job advert data from a major source. Please note that the sum of English regions does not equal the England total. This is to do with the way Textkernel assigns geographies to adverts. To assign job adverts to a country, adverts are first assigned to a local authority level and then assigned to English region or country. However, if a local authority or region is not detailed on the advert then it can be assigned directly to country level if that is known.
• UK and US, by category: Indeed Hiring Lab job postings index (seasonally adjusted, 7-day average, 1 Feb 2020 = 100), aggregate and by job category, to 16 Oct 2026. Categories are broader than occupations. Year-on-year compares with 52 weeks earlier; "this week" compares with 7 days earlier.
• Layoff notices: California EDD WARN report (to 21 Oct 2026) and Texas Workforce Commission WARN notices via the Texas Open Data Portal (to 23 Jun 2026; dataset last updated 7 Jul 2026). The TWC website blocks automated downloads, and the open-data copy lags TWC's own list by several months, so recent Texas notices are missing. WARN covers all causes and is shown for scale only; it is never counted as AI-caused.

• Announced layoffs are a small and biased sample of AI's labour-market effect; most adjustment shows up as slower hiring, which is why we show postings alongside.
• Job adverts and postings are not vacancies or employment. ONS figures are official statistics in development with imputed and suppressed months; Indeed indices reflect Indeed's own postings.
• Exposure scores reflect one AI provider's usage, in the US task framework, at two points in 2025.
• The SOC mapping is editorial; low-confidence rows are flagged.
• Grades are judgements. We publish the rationale and evidence so you can disagree with them.