Why this exists
The public conversation about AI and jobs runs on two broken inputs: alarmist headlines that treat any layoff at a tech company as proof that robots are taking over, and dismissive takes that wave away a real, measurable signal. Neither is honest. Our job is the honest middle. We count what can be counted, we label what is inferred, and we never present a guess as a fact. If a number is uncertain, we show you the uncertainty. If a company said AI cut jobs but the evidence is thin, we tell you it is thin.
The distinction we never collapse
There is a difference between a company saying AI caused a layoff and AI actually having caused that layoff. We treat these as two separate facts. Industry surveys in 2025 found that a large share of companies that framed layoffs as AI-driven were really cutting for financial reasons, a pattern known as "AI-washing" (citing AI because investors reward it). Intuit is a live example in our own register: the company said the cut was to focus on AI, while its chief executive said on CNBC the same day that "None of it had to do with AI." Even Challenger, Gray and Christmas, whose monthly figure is the most-cited AI-layoff statistic in the press, states plainly that it is difficult to say how big an impact AI is having on layoffs specifically. So we record what a company claimed in one field, and how strong the independent evidence is in another. You can always filter to the strongest claims only.
The biggest reason our number is smaller than the ones you have seen
In 2026 to date, the role figures reported for the cuts the press connected to AI add up to several times the roles in events where the company's own filing, memo or on-record statement actually named AI as a cause (today's figures, recomputed live, are on the front page and the how-many page). Most of the rest come from employers who named something else, or who denied AI outright.
This is not a subtle filter, and until v1.13 we were not applying it consistently. Some of the largest numbers in circulation fail on the employer's own words. Microsoft's July 2026 memo states the figure itself, "Today we are eliminating around 4,800 roles", and in the same memo says "I also want to be direct that the roles eliminated today are not being replaced by AI." Amazon's memo announcing 16,000 role reductions does not contain the word AI at all, and attributes the cuts to "reducing layers, increasing ownership, and removing bureaucracy". British American Tobacco's own release names AI zero times and announces a reduction of about 5,500 roles plus a further 3,500 that move to outsourcing partners, meaning those jobs continue under a different employer (the widely-repeated 9,000 is the sum of the two, a number BAT states nowhere).
A second group fails differently: the employer did name AI, sometimes powerfully, but never said how many people. Oracle's 10-K states that "the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce" and books an AI-driven restructuring plan of up to $2.1 billion, yet quantifies it only in dollars. The widely-repeated 21,000 is a year-over-year subtraction of two reported workforce totals, computed outside the company, and Oracle was hiring into AI while that total fell. Accenture, Cognizant and Citigroup are the same shape. Meta was too, until its quarterly filing stated the figure itself (about 8,000 roles), and it now counts.
We keep every one of these events in the database with the disclosed figure visible, marked "reported, not counted", because the scale of what is happening is real and hiding it would be its own distortion. What we will not do is add a company's cuts to a total labelled AI-attributed when the company did not attribute them to AI. The gap between the two figures is published on the front page, because that gap is the finding.
What we publish
We publish three things, in order of how grounded they are:
- An event database. Individual, dated, source-cited layoff events where AI was named as a factor, each classified by evidence tier and attribution strength. Every event links to its source.
- The AI Layoffs Index. A single 0 to 100 monthly number that summarizes how much downward pressure observed-and-attributable AI is exerting on US employment, relative to AI's own short history. It is not a body count. It is a normalized trend indicator, like a market index, always published with a confidence band.
- Context layers. Occupational AI-exposure, firm-adoption rates, and the macroeconomic baseline, so you can see the forces around the signal.
How we collect events
We do not invent events and we do not trust a single tracker. Our pipeline has a spine and an attribution overlay.
The event spine today is our own hand-curated register: every candidate event is entered by hand and its execution status set from its own primary record, with the dense Challenger, Gray and Christmas AI-cited monthly series (Tier 0) as the aggregate trend spine beside it. We deliberately do not lean on any single automated feed. Two structured public backbones are on the roadmap to widen and cross-check that spine, both catalogued as planned rather than live: layoffs.fyi (free to use with attribution) for clean per-company event rows, and the official state WARN Act notices consolidated by Stanford's Big Local News (public-domain government records of executed mass layoffs). WARN will matter enormously because it is the antidote to the platform's biggest credibility risk: confusing an announced cut with an executed one. Companies announce cuts that never fully happen, or that are quietly reversed. Klarna publicly reversed part of its AI-driven customer-service push, rehiring human agents after its CEO said the AI-first approach had lowered quality. But the reverse decay is just as real, and v1.11 caught it: Klarna's own Form 20-F for 2025 does not retract that claim, it escalates it, raising the AI assistant to "the work equivalent of over 850 full-time agents", reporting headcount falling from 4,352 to 2,831 across three years explicitly "by leveraging AI", and saying it expects the decline to continue. So a reversal in one function sat alongside continued AI-attributed reduction overall. We record that contradiction rather than resolve it in our own favour. WARN will tell us what actually happened. As those WARN and event-tracker integrations come online we will cross-check executed cuts against multiple WARN aggregators (Big Local News and WARNTracker) and independent event trackers (layoffs.fyi and TrueUp) rather than trusting any single feed; until then, execution status is set by hand from each event's own primary record.
The attribution overlay is where AI causation is attached to a spine event, and it is the heart of our credibility. For each candidate event we ask: did the company's own primary record cite AI? We search SEC EDGAR full-text filings (10-K, 10-Q, 8-K) and, where available, transcribed earnings calls, for the company's own language tying AI to the headcount reduction. SEC filings are public-domain, legally binding, and audited, which makes them the strongest possible attribution, and no existing tracker does this verification rigorously. We use jobloss.ai and the AIAAIC incident repository only as discovery leads, then re-verify each event against its own source (the company's filing or statement where one is published, otherwise named reporting) and store that source ourselves.
An employer's public statement about AI and the size of its workforce, made when no discrete cut is announced (a projection, a memo about future headcount, a reversal), is recorded as a workforce statement: it shows what the employer said, it never carries a counted figure, and it is never presented as a layoff. Each company page answers "did it lay off employees because of AI?" from what the employer said about each of its events, so a workforce statement, a reduction reached through attrition, a denial or conflicting statements are reported as what they are rather than folded into a yes (since v1.31).
Each cut also records which jobs it hit, in the employer's words where it gave them and otherwise from the reporting it cites (marked as such), grouped into broad job families such as software engineering, customer service or management layers. A family is recorded only for roles a source says were cut: the areas where a company says it is applying AI, the processes it is redesigning, the teams it is still hiring for and the way it describes its own business are not cuts. "Support" counts as customer service only when it faces customers; internal support functions count as corporate functions. Where a company's own filing books its severance by function (research and development, sales and marketing, general and administrative), each line counts as the matching family. When no source says which roles went, the cut is recorded as not specified, and it counts in no family (since v1.32).
A cut enters the register only when the employer, or credible reporting, ties that cut to AI: as AI doing the work of the people cut, as money moved from their roles into AI, or as AI changing the business that employed them. Automation with machines, a cut at a company whose product is AI, or AI mentioned only as the backdrop of an industry does not make a cut AI-linked by itself. When a later review finds no such link for an event already in the register, the event is removed and the corrections log says why (since v1.33: UPS's 2025 and 2026 network cuts left the register on this rule).
The evidence-tier system
Every event carries an evidence tier: Tier 1 (the company's own filing or on-record statement, full weight), Tier 2 (credible journalism quoting a named company source, weight 0.8), and Tier 3 (a single secondary tracker or vague forward-looking language, weight 0.5, flagged as inferred). Tier 0 covers aggregate signals like the Challenger figure and the Stanford correlation, which calibrate the index but never appear as discrete events.
Orthogonally, every event gets an attribution sub-flag: Explicit (the company itself declared it AI-related), Blamed (a source named AI but the company did not), or Mixed (AI cited alongside other material factors). In the index, Explicit counts fully and Mixed counts at 0.4, so partial AI attribution is down-weighted, never hidden. Blamed is different, and has been since v1.13: a Blamed event is shown as context and never counted, because prong one of our standard requires the employer's own source to name AI, so a Blamed event is a failed claim rather than a discounted one. Its disclosed figure stays visible and drillable as reported-not-counted and feeds the disclosed-versus-attributed gap, but it never enters an attributed total (the legacy 0.6 weight is retained in the schema only so old corrections stay legible, and no counted event carries it). Since v1.30 the register also records cuts that other trackers or the press present as AI layoffs when the employer named a different cause or said AI was not the reason: each is Blamed, shows the claim and why it fails, and never counts, so a reader who meets the claim elsewhere can check it here. The claim has to rest on a source that can be opened and that presents the cut itself as caused by AI: a tracker's label with no such source, or coverage that only mentions the employer's AI plans beside the cut, is not recorded (since v1.31 written here; the v1.30 completeness check drew this line). When the employer's own statements conflict, a statement by the employer that AI is not a reason for the cut keeps its figure out of every counted total, even where another of its statements names AI; the event records both statements. A statement that AI is not doing the work of the people cut, made while the employer names AI as a reason (for example, freeing money to invest in AI), is not such a denial (since v1.31). For an employer that named no AI reason at all, a statement that AI did not replace the people cut is recorded as a denial too. A reporter's paraphrase of the employer's reasons ('citing advances in AI') is not the employer's own words where the employer's own statement on the cut names no AI. This lets you ask for "Tier 1 and Explicit", the most defensible claims, rather than treating every event as equally certain.
A count also has to come from the employer. A role figure enters a counted total only when the company itself put the number on the record: in a filing, a release or a letter it publishes, through a named executive speaking publicly, or through its spokesperson, including a number the press reports as the company's own statement (since v1.31 written here). A percentage counts only when the company states it as a single value and also publishes the workforce it applies to, taken as its most recent published headcount; a percentage given as a bound or as nearly a value is not converted. A range the company states counts at its lower end, and a planned cut it states as its headcount before and after counts as the difference. A number whose origin is the reporting itself (unnamed sources, an internal memo the company has not confirmed, or reporters' arithmetic from a percentage or a severance charge) stays visible as reported, not counted, and so does a net change in total headcount, an "up to" ceiling, or a count that includes contractors or roles moved to other teams or locations.
A third axis weights whether the cut was actually carried out, because announced is not executed: a completed reduction counts fully, a partially-executed one at 0.7, a merely-announced (often multi-year) plan at 0.5, and a reversed cut at 0.15. So a company that says it will cut 30,000 jobs over five years never lands in the index like 30,000 people already let go, and a cut that was rolled back fades toward zero.
How the index is computed
The AI Layoffs Index blends three components, each standardized against its own history, then combined with fixed weights and mapped to a stable 0 to 100 scale. One honesty note up front: the three do not move on the same timescale. The Direct momentum leg and the Macro weekly-claims composite update on every refresh, while four of the five Exposure legs are slow academic and survey releases. So the fixed weights govern each pillar's influence per unit of standardized movement, not a fixed share of the final number: Direct dominates short-term movement when AI-cited cutting is far above its own history, and Exposure anchors a slower mid-range level rather than swinging the headline. We consider re-standardizing the pillars onto a single common spread, but that would alter the headline and the deliberate design, so we would do it in the open and version it, not quietly.
Direct AI-attributed cuts (45%) has two aligned views. The concrete level beside the index is our own register's confidence-weighted attributed headcount: every US event scaled by its evidence tier (Tier 1 times 1.0, Tier 2 times 0.8, Tier 3 times 0.5), its attribution factor (Explicit times 1.0 and Mixed times 0.4; a Blamed event carries no counted headcount and is never summed in, so its legacy 0.6 factor is unreachable), and its execution status (executed times 1.0, partial times 0.7, announced times 0.5, reversed times 0.15), summed over a trailing window. The component's trend comes from the dense Challenger AI-cited monthly series, computed as two unit-consistent legs that we combine 60/40. The level leg asks how far the current pace sits above AI's short history: it standardizes the trailing-three-month AI-cited pace, annualized, against a fixed, documented log-scale anchor (calibrated to the 2023, 2024, and 2025 annual totals and then held constant), so it compares an annual figure with annual figures on a scale that does not drift as new years accrue. An earlier build re-estimated that scale from just the three prior-year points each time, which with three points on an exponential series was degenerate (the estimated spread was really the trend's own slope, and it would lurch every year-end); the fixed anchor removes that fragility while preserving the intended behavior that a rising series reads elevated on the level leg. The momentum leg asks whether cutting is accelerating or cooling right now: it standardizes the latest month-over-month log change against the recent monthly changes. Keeping the two on their own units matters. An earlier version spread each prior year across twelve flat monthly-equivalent points and mixed them into a single z-series measured against an expanding-from-inception window, which quietly inflated every recent month and muffled real moves; separating the level from the momentum fixes that at the root, and we show the raw month-over-month change beside the score so a sharp move is never smoothed away. Challenger counts every announced AI-cited cut (most without a per-company headcount), so it is far larger than our verified register, which is a curated, primary-checked subset; we show both so the gap is visible, and the agreement between them sets the component's confidence.
AI-exposure-weighted labor shift (35%) is the realized-pressure signal that does not depend on any company's self-narrative. It is a transparent weighted blend of five measures: the Stanford Canaries early-career employment decline (around 16% for workers aged 22 to 25 in the most exposed occupations, the heaviest leg), the Anthropic Economic Index share of AI usage that automates rather than augments, the Census BTOS firm AI-adoption rate, the occupational exposure breadth from Eloundou et al.'s task-based high-exposure share (with the Felten AIOE measure and the Yale Budget Lab harmonization serving as cross-checks that show how much the exposure measures disagree, not as co-inputs to the scored number), and a live realized leg. Each of the first four is mapped onto the pressure scale by a documented endpoint range, a stated "no effect" level and a stated "severe" level, rather than a slope tuned to land in a chosen band, so the mapping is a falsifiable claim you can argue with. The fifth, live leg is what makes this pillar move between academic releases: it is the six-month change in Information-sector payroll employment (the sector the Census reports leads AI adoption) measured relative to total nonfarm employment, so only the exposed sector's underperformance versus the broad market counts. Using a relative measure is deliberate, because the broad-market move is the Macro pillar's job, so the two never double-count; today the Information sector is shedding jobs while the overall market grows, which reads as elevated realized pressure. Its confidence falls as the measures disagree. Around these components we triangulate against independent real-data sources that do not rely on any company's self-narrative: aggregated payroll (the ADP National Employment Report, the substrate under the Stanford Canaries series, whose live private-employment level we show beside the macro signals), live job-postings demand (the Indeed Hiring Lab AI tracker, whose US AI-mention share we fetch live and show as the demand-side counterweight, not weighted into the score), the official JOLTS layoffs flow in the AI-leading Information sector and (added in v1.7) in Professional and Business Services, the number-two AI-adopting sector (both fetched live and shown as year-over-year figures beside the employment leg, never scored so they cannot double-count), the household-survey unemployment rate for the Information sector and that sector's job openings (added in v1.9, an independent household-survey cross-check on whether exposed-sector workers are landing in unemployment and an early demand-side read where a hiring freeze shows up before layoffs do, both context only), a sharper AI-core employment read (Computer Systems Design and Related Services, the most software-native slice of the economy, shown relative to the broad market as context), the intensive margin of average weekly hours in the Information sector relative to the whole economy (added in v1.14) and the compensation margin of Information-sector hourly earnings relative to total-private earnings (added in v1.15, the one signal category the pillar previously lacked, shown as context because a widening wage premium is ambiguous for a displacement thesis), transaction-observed AI spend joined to headcount outcomes (Ramp and Revelio Labs, whose keyless Public Labor Statistics mass-layoff feed we also catalogue), and Stanford's AI Economic Indicators platform.
Macro-adjusted baseline (20%) is the business-cycle subtractor. Rather than lean on any single series, it standardizes a composite of four cyclical signals from the St. Louis Fed's keyless feed, initial jobless claims, continued claims, unemployment, and job openings, each oriented so that cyclical weakness pulls the AI-attributable residual down, and weighted so no one noisy weekly print can swing the pillar (initial claims 40 percent, continued claims 20 percent, unemployment 25 percent, openings 15 percent). We estimate how much of the labor weakness the general cycle explains and enter only the residual; prime-age participation, a live ADP private-payroll level, the whole-economy layoffs and quits rates, temporary-help employment, and (added in v1.16) a model-based recession probability, the pillar's first non-labor coincident read, are shown alongside for context, never scored. This runs live with no API key to configure. It is treated as a live composite only when at least two independent signal FAMILIES are present (and at least sixty percent of the intended signal weight), where initial and continued claims count as one family because both are the same Department of Labor unemployment-insurance series: if only the correlated claims pair survives, it degrades to neutral at reduced confidence rather than pass off a one-family read as a multi-signal composite. Every live series also passes a freshness gate, so a frozen or discontinued feed (a real risk: one ADP series we once used was silently frozen for years) is dropped rather than scored as current, and only then does the term fall back to the latest published reference values and hold near neutral.
We blend the three z-scores (0.45, 0.35, 0.20) and map the result through a fixed logistic to 0 to 100. To be exact about the aggregation, because being exact is the point: each pillar sub-score is first turned back into a log-odds z by the exact inverse of that same logistic (the sub-score read as a percentage, then its logit), the three z's are combined by a weighted mean whose weights sum to exactly 1.0 (0.45 plus 0.35 plus 0.20), and that mean is mapped back through the logistic. So the headline is a logarithmic opinion pool of the three pillars, a weighted consensus taken in odds space, not a plain average of the three 0-to-100 sub-scores, and the logistic's steepness is fixed at 1 precisely so it is the exact inverse of the per-pillar encoding rather than a free knob we could turn to move the number. (Where the confidence band needs a spread in points it uses a separate, plain weighted average of the 0-to-100 sub-scores to center the dispersion, which differs from the headline by a small curvature gap; the two are deliberately distinct and we name both here so neither is mistaken for the other.) An index of 50 is the logistic's neutral point, a blended z of zero, and the named-band cutoffs are the symmetric counterparts of that neutral in the same log-odds space: 30, 50 and 70 map to a blended z of about minus 0.85, zero and plus 0.85, so Below average, Elevated and Extreme sit evenly around the AI-era neutral rather than at arbitrary round numbers. It is important to be precise about what that neutral does and does not mean here: because the Direct level leg measures today's AI-cited pace against a FIXED 2023-to-2025 AI-era anchor rather than a rolling average, and today's pace sits far above that anchor (about 2.3 standard deviations, since the current run-rate is roughly twenty times the AI-era annual baseline), the index reads well above 50 by construction. So 50 is not "the current normal": it is the AI-era-baseline neutral, and a reading well above that neutral point (in the named Elevated or Extreme bands) reflects that today's pace sits far above the 2023 to 2024 baseline. The actionable signal is the month-to-month movement and where the value sits within the above-neutral range, not its bare distance from 50. The scale is relative to AI's short history (tracking began in 2023), not to total unemployment, and we say so on the chart. The confidence band widens honestly when the data is thin (it is a transparency device, not a statistical confidence interval). Beside the index we always show the underlying confidence-weighted attributed-cut level, a concrete headcount with its range, so the page never floats free of real numbers.
How complete is each component, honestly
We weight the three components by how much we can actually stand behind, and we would rather tell you where a number is thin than dress it up.
Direct AI-attributed cuts (45%) is the most complete. It rests on our own event register, which we re-verify against each event's source, with the dense Challenger AI-cited monthly series as the trend spine. The register is a curated subset and will always trail the true total of AI-named cuts (we add events continuously as we verify them, and expanded it substantially in this version), but every number in it is checked, and the Challenger spine gives the component a real monthly trend. This is why it carries the most weight.
AI-exposure-weighted labor shift (35%) is mostly current, now partly live. Four of its measures (the Stanford early-career employment decline, the Anthropic automation-versus-augmentation share, the Census BTOS firm-adoption rate, and the occupational exposure breadth from Eloundou et al., cross-checked against the Yale Budget Lab harmonization) are periodic academic and survey releases, real and re-checked against their sources at every audit (the date of the latest is at the top of this page), but not a live monthly feed, so they read a level rather than a fast trend and disagree on magnitude. The fifth measure, added in this version, is live: the AI-exposed Information sector's employment relative to the broad market, refreshed on every update, so the pillar is no longer frozen between releases. We still hold its confidence deliberately moderate and let the disagreement among measures widen the band rather than firm up the headline.
Macro-adjusted baseline (20%) is live and multi-signal. It reads the St. Louis Fed's keyless data feed and standardizes a composite of four cyclical series (initial and continued jobless claims, unemployment, and job openings) rather than a single print, so one noisy weekly claims number can no longer swing the pillar, and the index is genuinely cycle-adjusted on every update with no API key to configure. Only if that feed is unreachable does it fall back to the latest published reference values and hold near neutral at reduced confidence. We show its inputs either way, including prime-age participation and a live ADP private-payroll level, so the business cycle is always visible beside the AI signal.
The honest summary: the index leans hardest on the component we can verify event by event, and the two context components lower the overall confidence (and widen the band) rather than inflate the number. As we keep expanding the register and the live feeds stay connected, completeness rises and the band tightens.
How we express uncertainty
We never publish a bare point. Every value carries a confidence band built from two honest inputs: how thin the data is (how high-tier and explicit the recent register is, the agreement among the exposure measures, the reliability of the macro composite) and how much the three pillars actually disagree with each other. That second input matters: when the direct, exposure, and macro sub-scores diverge, that spread is real epistemic uncertainty, so it widens the band directly rather than being averaged away. Thin or conflicting data widens the band automatically. A short confidence note accompanies every value, naming the dominant uncertainty. The band is a disclosed transparency device, not a statistical confidence interval, and we say so.
One caveat we state plainly, because it cuts against the reassurance a tight band offers: the three pillars are not statistically independent instruments, so their agreement is not fully independent corroboration. Direct and Exposure, which together carry eighty percent of the weight, share the same technology and Information labor cycle: our AI-cited register and the Challenger spine are Information-heavy, and the Exposure pillar's live realized leg is Information-sector employment itself, measured relative to the broad market. So an Information-sector contraction can push both pillars up whether or not the cause is AI. Because the disagreement term measures only the spread between the pillars, it is blind to this shared, common-mode movement: the three can agree and still be wrong together, and the band would not widen to reflect it. A low spread is therefore a floor on the uncertainty we disclose, not a ceiling on the uncertainty that exists. In the other direction, the Macro pillar is built to move opposite to Direct's cyclical part, so in a broad downturn it widens the spread by construction, which is the pillar doing its job rather than new doubt. The dispersion band is honest as a floor, but it is not a covariance-aware interval. Folding an explicit, assumed pillar-correlation structure into the width would move the published band, so we would ship that in the open and version it rather than quietly, the same discipline we apply to every other band-moving change.
We also surface the age of each input rather than hide it. The two oldest scored inputs are shown with their true vintage: the task-exposure breadth shares are 2023 estimates (the oldest number in the index) and the early-career decline is a measurement on data through September 2025, so each carries its own as-of date and a stale marker beside it rather than the refresh date of a companion series. Today we surface data age beside each input rather than fold it into the band. We have also designed, but not applied, a confidence adjustment that would lower this pillar's confidence as its scored inputs age: measured from the published data vintage rather than the calendar (so the band can move only with a versioned change), with a staleness horizon for each input matched to how fast that construct moves in the world (firm AI adoption fastest, structural task exposure slowest), and added beside the cross-measure disagreement term rather than replacing it. It could only widen the band, never narrow it, and it would leave the headline untouched: at the July-2026 vintage it would lower the Exposure pillar's confidence by about one hundredth, which rounds away at the aggregate. Because it changes how we disclose uncertainty, it would ship only as a deliberate, versioned methodology change, and it has not been applied.
We reconstruct history only where a pillar has a look-ahead-free, scale-stable record. The direct-attribution trend qualifies: its Challenger spine is essentially unrevised and is scored on a fixed pace scale, so a past value is comparable to today's. Macro does not: it is a cycle residual standardized against its own recent trailing window, so it measures change rather than a level and has no fixed-scale value to plot back through time, and its denominator shifts with local volatility. Reconstructing it (or the exposure pillar's academic anchors) from today's revised government data would also paint the past with numbers that were not known then and would look more precise than the index was in real time. So on the index-over-time chart only the direct trend is a real series; exposure and macro are held at their current reading and marked as held, and we would rather show an honestly held line than a convincing reconstructed one. The single exposure component that IS comparable across time, the live Information-sector-versus-market employment differential, is drawn on its own labeled track on a fixed scale, carrying the revised-data note, and is never relabeled as the exposure pillar's full history.
Every number is drillable
Nothing on this site is a number you have to take on faith. The headline index opens into its three components; each component opens into the concrete inputs that produce it; each input traces to the events or the cited statistic underneath it; and each of those links to its source. A percentage is never the end of the line, it is the top of a path down to the atomic record.
What we deliberately do not do
We do not republish Challenger's dataset or jobloss.ai's dataset wholesale; both are cited as statistics or used as leads, and our own database is hand-curated and verified against openly-licensed primaries. We do not present an exposure score as a confirmed job loss; exposure is pressure, not proof. We do not count non-US workers in a US figure (so Salesforce's global support reduction and Atlassian's Australian cuts are shown for context but excluded from the US sum), and we flag announced-but-not-executed and reversed cuts (Commonwealth Bank's AI-attributed call-center cut was reversed; Klarna's AI-for-support push was partly reversed). And we honor the most important finding in the research: as of mid-2026 there is no economy-wide AI jobs apocalypse in the aggregate occupational mix, but there is a real, localized signal for early-career workers in highly-exposed occupations. We report that signal at its true size.
How we handle corrections, in both directions
A correction policy that runs in one direction is not a correction policy, it is an editorial preference, so ours is written down here rather than left as a habit.
The rule: a finding is applied in the same review that discovered it if and only if three things are true. The primary source is in hand. The quote that carries the finding is written into the record. And the finding has survived a documented attempt to break it. Whether the correction raises or lowers the published number is not one of the three. Direction is never a reason to act and never a reason to wait.
If a finding fails that bar it is deferred, and the deferral is published with the reason, the expected direction, the expected size, and the version it is due in. The only admissible reason for a deferral is evidentiary: the source is not in hand, or the finding depends on data we have not yet verified.
This replaces an unwritten rule that ran until v1.12, under which a correction lowering our own number shipped immediately while one raising it waited a review. We audited our own correction log to see what that habit had actually done, and the answer is specific. Corrections to upstream data were never affected and moved freely in both directions. But among findings this index generated itself, eight of eight downward corrections shipped in the review that found them, and zero of five upward ones did, with delays of one to three reviews. In twelve reviews not a single downward finding was ever held for the verification a delay supposedly buys, including v1.11's decision to cut its own attributed level by 79% on judgment, in one review, with no delay at all.
The reason we are not keeping that rule, even disclosed, is that this index already applies its conservatism before it knows which way a result will go: the three-part evidence standard, the tier and attribution multipliers, the execution factors, counting only Challenger's explicit AI line, and excluding every global figure from the US sum. Applying a second conservatism after the sign of the result is known double-counts the first one and cannot be audited by a reader, because nobody can tell how much of the published number is measurement and how much is thumb. A guard that fires on one sign is not a guard against motivated reasoning.
The Klarna correction in v1.12 is what made this concrete. It was deferred in v1.11 explicitly because acting would raise the index. That reason was factually wrong: all three Klarna entries are non-US, and the model filters to US events before computing anything, so the correction could not have moved the headline in either direction. The guard had fired on the sign of the finding without anyone checking whether it had a consequence.
Corrections and versioning
Government and academic sources revise their numbers, and so do we. Every change to a published figure, a register entry or the method is listed in the corrections log at the end of this page, newest first, with the version it belongs to and, when the headline moved, by how much.
The methodology version changes when the method changes: how the index is computed, weighted or bounded, or the standard an event must meet to be counted. New data alone, such as a new Challenger month or a jobs report, does not change the version; it advances the data vintage, the newest reference month the index carries. The version and the data vintage appear on every page.
Two corrections show the range. In v1.11 we re-read the register's largest entries against the companies' own filings and cut our own 2026 US attributed total by 79 percent, from 10,205 to 2,155 confidence-weighted roles. In v1.10, verified monthly Challenger figures for late 2025 corrected a small-sample distortion in the momentum measure and raised the headline by about seven points. Both are in the log, as is every change since the first version.