Where every number comes from, how the AI score is calculated, and what we're still missing. Data updated 2026-10-26.
Any significant layoff announced or filed between 2026-01-01 and 2026-10-26, whether or not AI is involved. That includes every 2026 state WARN notice in the states we cover, every SEC Form 8-K Item 2.05 (exit or restructuring) filing we can find, and major company-wide announcements confirmed by the company or by major outlets. In total that comes to 1,509 events and 453,424 counted workers.
WARN notices are filed per site. We group them into one company-level event: notices from the same employer within about 60 days are combined, and notices that match a company-wide announcement are attached to it. Every site-level row is still available in the WARN locations CSV (2,424 rows).
Other trackers are used only as leads. Nothing is copied from them; every lead is checked against a primary or major source before it goes in. Several leads turned out to be 2024 or 2025 announcements, or headcounts nobody actually confirmed, and those were left out.
Please read this part: the AI score estimates how likely a careful reader of the public record would conclude that AI or automation was a stated or clearly documented contributing cause of the layoff. It is not a measured probability that AI caused the job losses. Companies rarely say why they cut, and several studies, including the Stanford Digital Economy Lab's below, find AI showing up mostly as reduced hiring rather than layoffs. So the score describes the evidence, not the cause. Each row shows three things: the score, an 80% interval, an evidence grade, and how much each signal contributed.
| Signal | What it measures | Fitted weight |
|---|---|---|
| A | Attribution strength: the strongest AI or automation link in the record (see table below) | +4.92 |
| AP | Bonus applied when that link comes from a primary source (the company itself or an SEC filing) | +1.58 |
| E | Executives tie AI to efficiency or productivity | +0.00 |
| R | AI exposure of the affected roles (mean percentile of AIOE, GPT exposure β and Anthropic observed exposure) | +0.03 |
| I | Industry AI exposure (AIIE percentile by NAICS) | +0.66 |
| C | Major AI investment announced by the company within 180 days (1) or earlier in 2026 (0.5) | +0.00 |
| X | Contrary evidence (subtracted) | +1.70 |
Score = 100 × logistic(-3.30 + Σ weight × signal). Contrary evidence is subtracted. The weights are fit by sign-constrained, L2-regularized logistic regression on 223 hand-labeled events (below). Every weight is required to push in the expected direction. The fit gave concurrent AI investment a weight of 0: in our labeled set, a big AI capex announcement didn't help predict whether a layoff was tied to AI. We keep the signal in the model and show it, but right now it doesn't change any score. Role and industry exposure carry small weights. That's on purpose: exposure measures what AI could do in a job, not what a company actually did. On its own, exposure can only lift a no-evidence WARN notice from about 2 to about 8.
| A | Evidence tag | Meaning |
|---|---|---|
| 1.0 | ai_explicit_replacement | ai explicit replacement |
| 0.85 | ai_cited_reason | ai cited reason |
| 0.7 | fund_ai | fund ai |
| 0.7 | ai_reskill_replace | ai reskill replace |
| 0.65 | ai_replacing_work_reported | ai replacing work reported |
| 0.6 | ai_disruption | ai disruption |
| 0.6 | warn_reason_automation | warn reason automation |
| 0.45 | exec_ai_efficiency | exec ai efficiency |
| 0.4 | sec_ai_language | sec ai language |
| 0.35 | sec_periodic_ai_language | sec periodic ai language |
| 0.35 | automation_mention | automation mention |
| 0.3 | press_ai_link | press ai link |
| 0.2 | automation_physical_mention | automation physical mention |
Contrary evidence (X): the company denies AI played a role (+0.6); a closure, bankruptcy or seasonal end (+0.4); each other co-cited cause such as tariffs, lost demand, a merger or relocation (+0.15). X is capped at 1.
| Grade | Meaning | Events |
|---|---|---|
| A | AI link stated by the company or in an SEC filing | 38 |
| B | Executive AI statements, AI language in a filing, or credible reporting that AI is taking over the work | 57 |
| C | Press or analyst inference, or a generic automation mention, only | 15 |
| D | No AI-specific evidence; score comes only from role/industry exposure minus contrary evidence | 1,399 |
The score reflects how strong the evidence is. The grade reflects what kind of evidence it is. Two events can both score 40, one because filing language mentions AI (grade B) and one on press inference alone (grade C). Treat any grade D event as "nothing public ties this to AI", not as "AI was not involved".
The interval shown with each score (for example "72–81") is the 10th to 90th percentile of the score across 300 bootstrap refits of the weights. It captures model uncertainty only. It can't capture information nobody has published, which is what the evidence grade is for.
We hand-labeled 223 events on a 0–3 scale. 3 = the company explicitly ties the cuts to AI. 2 = AI is stated or clearly documented as a contributing cause, for example in a filing, an executive statement, or reporting that AI now does the work. 1 = press speculation or a generic automation mention only. 0 = no AI evidence, or the company denied it. A label of 2 or higher counts as "AI-linked". The set has 122× label 0, 27× label 1, 53× label 2, 21× label 3 and contains 74 AI-linked events: every curated news event, plus random samples of SEC and WARN events. Scores were checked with 5-fold cross-validation repeated 10 times, so each event is scored by a model that never saw its label.
| Metric | v2 rubric | v3 model (cross-validated) |
|---|---|---|
| AUC (ranking) | 0.993 | 0.991 |
| Brier score (lower is better) | 0.05 | 0.045 |
| Accuracy at the 50 cut-off | 0.937 | 0.937 |
| Spearman correlation with the 0–3 label | 0.873 | 0.856 |
| Weighted κ, bands vs labels | 0.724 | 0.793 |
| Score bin | Events | Mean score | Share labeled AI-linked |
|---|---|---|---|
| 0-15 | 121 | 5 | 0% |
| 15-30 | 15 | 23 | 13% |
| 30-50 | 17 | 41 | 41% |
| 50-70 | 15 | 60 | 73% |
| 70-100 | 55 | 89 | 98% |
Caveat: the labels were assigned by our team from the same records the signals are built from. High agreement shows the model is consistent with a careful reading of the sources. It does not show the model knows the true cause. The v2 rubric already agreed closely with the labels. v3's improvements are calibrated bands (weighted κ 0.724 → 0.793), an interval, an evidence grade, and a per-signal breakdown. The old rubric score is still in the data as rubric_v2_score.
| Band | Score | Events | Workers |
|---|---|---|---|
| High | 70+ | 55 | 69,342 |
| Mid-high | 50+ | 15 | 25,000 |
| Mid | 30+ | 22 | 10,050 |
| Low | 15+ | 16 | 15,860 |
| Minimal | 0+ | 1,401 | 333,172 |
The AI Layoff Index for a period is Σ (counted workers × AI score / 100). It's the expected number of workers whose layoff a careful reader would tie to AI, given the public evidence. It is an estimate for tracked events, not a measure of AI job losses across the economy.
WARN data loaded for these states:
| State | Site notices | Workers | Source |
|---|---|---|---|
| VA | 59 | 7,722 | official source |
| UT | 13 | 1,861 | official source |
| OR | 76 | 2,646 | official source |
| CA | 1,183 | 62,569 | official source |
| TX | 109 | 17,370 | official source |
| NJ | 114 | 14,186 | official source |
| CO | 46 | 5,501 | official source |
| PA (last good copy; latest fetch failed) | 59 | 9,748 | official source |
| IL | 133 | 18,422 | official source |
| WA | 129 | 27,943 | official source |
| FL | 215 | 15,749 | official source |
| IN | 28 | 4,208 | official source |
| MD | 112 | 8,047 | official source |
| AL (last good copy; latest fetch failed) | 28 | 2,458 | official source |
| WI | 53 | 3,733 | official source |
| NC | 68 | 6,789 | official source |
Not yet covered by WARN: New York (its public dashboard sits behind a bot check), Ohio (page not found), Michigan, Massachusetts, Tennessee and Missouri (they block automated access), Georgia (site down when we checked), Arizona and Kansas (no public export), Minnesota (only monthly PDFs, which we can't parse reliably), and every other state not in the table above. We don't get around blocks. Large layoffs in those states are still included when there's an announcement, SEC filing or major news report. Illinois is partial: updates to notices first filed in 2025 are left out. Searching SEC full text for Item 2.05 finds fewer filings than we expected, so some 8-K restructurings may be missing. To catch more, we also search 10-Q and 10-K filings for workforce-reduction language and attach what we find to the matching events.
Benchmark: Challenger, Gray & Christmas counted 573,195 announced U.S. job cuts from January through September 2026, and 120,136 where employers cited AI. Challenger counts announcements, including ones without WARN filings, and doesn't publish a company-level list, so the gap between our numbers and theirs is mostly events we haven't documented with a source yet.
The data is rebuilt every week. A script downloads the latest WARN files and SEC filings and re-scores every event. New news events are added by hand only after they're checked against sources. Any new WARN or SEC entry that mentions AI is flagged for a person to review. New curated events get a hand label, and the model is refit and re-checked each time (calibrate.py). A changelog records what was added and what was re-scored.
We don't estimate headcounts, we don't publish anonymous claims without a public source, and we don't call any one layoff "caused by AI". The score describes the evidence and nothing more.
See something wrong? Tell us. We fix errors and re-score as soon as they're confirmed.
