California Built the First AI Job-Loss Tracker. Here Is What Employer Brand Leaders Should Read Into It

What the California AI-Unemployment Tracker actually found, where it agrees and disagrees with the other big AI labour studies, and how talent and employer brand teams can use it.

On 25 June 2026, California became the first state in the United States to launch a public tool that tracks AI-related job loss in close to real time. The California AI-Unemployment Tracker, known as CAIT, links unemployment insurance claims to occupational measures of AI exposure, and it ships with a research report that reaches a careful, two-part conclusion. Statewide, there is no surge in layoffs among AI-exposed workers. Underneath that calm headline, claims are running persistently higher for college-educated workers in high-exposure jobs, for workers in the San Francisco Bay Area, and in technology-heavy sectors.

SPONSORED

Your employees know the truth. Does your EVP? At Fathom we measure the "Credibility Gap" between your promise and their reality.

Close the Gap

For anyone who builds an employer brand, recruits, or manages talent, the tracker is worth more than its headline. It is a rare piece of administrative-data evidence in a debate that has been dominated by vendor forecasts and chief executive soundbites. It also models the kind of measured language that talent leaders will need when candidates, hiring managers, and boards ask the question that is now unavoidable: is AI taking our jobs.

In this breakdown:

• What CAIT is, who built it, and what it measures.
• The three findings that matter, with the numbers attached.
• How CAIT sits alongside the Stanford, Yale, and Anthropic research.
• The limitations the authors flag, and why they matter for how the tracker should be cited.
• Concrete implications for employer brand, talent attraction, and internal communication.

Who built CAIT, and what it tracks

CAIT is a partnership between the California Policy Lab, a nonpartisan research centre at the University of California, and the California Employment Development Department, with backing from the Governor's office. The accompanying report, Tracking AI-Related Job Loss Using Unemployment Insurance Claims Data in California, was prepared by a team led by Dr Ben Hyman and Professor Till von Wachter, and published in June 2026. The tracker will be updated monthly, and the underlying tabulated data are public.

The method is the interesting part. When a Californian files for unemployment insurance, they select an occupation code describing their pre-layoff job. CAIT attaches an AI-exposure score to that occupation using two established measures. The first, the potential exposure measure from Eloundou and colleagues, published in Science in 2024, asks whether a large language model could cut the time needed for an occupation's tasks by at least half. The second, the observed exposure measure from the Anthropic Economic Index, asks how often workers actually use Claude to complete those tasks. Workers are then sorted into low, moderate, and high exposure groups, with high defined as the top 25% of scores.

The two measures pull in different directions by design, and the authors are clear about why. The potential measure covers every occupation but cannot say whether AI is being used. The observed measure captures real usage but skews toward technology occupations, because Claude's user base does. Computer programmers, customer service representatives, and data entry keyers sit near the top of the exposure rankings. Heavy truck drivers and nursing assistants sit near the bottom.

One point matters for any reader tempted to over-read the observed numbers. The report states plainly that observed exposure is a relative measure of where Claude is used, not an absolute measure of AI penetration across the economy. If Claude's users cluster in software, the measure reflects that concentration.

Finding 1: No statewide surge in AI-exposed layoffs

The first finding is the one most likely to get lost in a headline, so it is worth stating cleanly. Across California, from the release of ChatGPT-3.5 in November 2022 through May 2026, initial unemployment claims from low, moderate, and high AI-exposure occupations moved roughly in parallel. There was no sharp divergence after generative AI reached the public.

The share of claims coming from AI-exposed occupations also held broadly steady. In the post-ChatGPT years of 2023 to 2025, 30.3% of initial California claims came from high-exposure occupations and 39.6% from moderate-exposure ones. The pre-pandemic benchmark from 2017 to 2019 was 27.8% high and 41.0% moderate. The composition barely moved.

SPONSORED

Employer branding now has a measurement standard. The Talent Gravity Standard is a six-driver framework for quantifying employer attractiveness and the gap between brand promise and employee experience.

Read the Standard

There is a useful secondary point buried in that statistic. From 2023 to 2025, 30.3% of all initial claims came from workers whose prior jobs were highly AI-exposed. That cuts against the assumption that AI-exposed workers, being more educated and more employable, would rarely file for unemployment. They do file, which is what makes UI claims a usable signal in the first place.

Finding 2: College graduates in exposed jobs are filing more

The calm aggregate picture hides a sharper one underneath. When the researchers split claims by education and exposure, a clear pattern appears. After ChatGPT-3.5's release, high-AI-exposure claims rose among more educated workers, peaked in July 2023, and stayed elevated.

The specific numbers give the finding its weight. Claims among workers with a bachelor's degree in high-exposure occupations rose by more than 50% from November 2022 to July 2023, climbing from roughly 13,000 to more than 22,000 claims per month. By May 2026, that group was still elevated at about 16,000 claims per month. Master's and PhD holders in high-exposure occupations followed the same shape, moving from a baseline near 13,000 in November 2022 to a range of 16,000 to 22,000 per month from mid-2023 onward. Similarly educated workers in low-exposure occupations did not show the same shift.

The authors are careful here, and talent leaders should borrow their caution. The pattern is real and persistent. The cause is not settled. Reduced hiring in information technology after the pandemic, restrictive monetary policy, and broader economic conditions could all contribute. The report's own event-study analysis treats these as descriptive accounting exercises, stopping short of causal estimates.

Finding 3: The Bay Area and tech sectors run hot

Geography and industry tell the third part of the story. High-AI-exposure claims in the San Francisco Bay Area rose by more than 50% after ChatGPT-3.5's release and stayed above the rest of the state. The average AI "content" of Bay Area claims, meaning the share of occupational tasks that AI could perform, rose steadily from before the launch, peaked in mid-2023, and remained higher through May 2026.

By industry, the Information and Professional Services sectors showed large increases in high-exposure claims in 2023. Professional Services, which contains many software development roles, stayed elevated relative to the state as a whole. The California Policy Lab's launch materials note a more recent wrinkle worth tracking: the Information sector has fallen back toward pre-generative-AI benchmark levels in late 2025, while Professional Services has remained higher. The Finance and Insurance industry consistently carries the largest share of high-exposure claims of any sector.

The authors close this section with the line that should anchor any responsible reading. They could not rule out the post-pandemic unwinding of the technology hiring boom as an alternative explanation for these patterns. They also note that it becomes harder to explain the persistence into 2025 and 2026 by COVID-era factors alone.

SPONSORED

A professional home for employer branding. The Institute of Employer Branding Professionals is being built and you can help shape it.

Find out more

One more result deserves a mention because of what it does not show. CAIT found no large disproportionate increases in high-exposure claims by race, ethnicity, gender, or age. On age in particular, the authors note that younger workers are less likely to claim UI benefits, through lack of awareness or eligibility, which limits what this dataset can see about the group other studies have flagged most loudly.

How CAIT fits the wider evidence

CAIT does not sit alone, and its real value for an employer brand audience is as a tiebreaker in a noisy debate. The research splits into two broad camps, and CAIT lands a foot in each.

On one side sits the no-aggregate-disruption camp. The Budget Lab at Yale, tracking occupational mix and unemployment duration through 2025 and into 2026, found that AI exposure showed no clear relationship with employment or unemployment, and described a labour market defined by stability, with no sign of major disruption. Administrative data from Denmark, analysed by Humlum and Vestergaard, found essentially zero effect on earnings or hours across exposed occupations through 2024. CAIT's statewide finding is consistent with both.

On the other side sits the concentrated-disruption camp. The Stanford Digital Economy Lab's "Canaries in the Coal Mine" study, using ADP payroll records for millions of workers, found that early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a relative employment decline of around 13%, while older and less-exposed workers held steady or grew. In their February and March 2026 updates, the Stanford authors reported the effect reaching roughly 16% by October 2025 with no reversal, and argued that interest rate changes do not explain it well. Anthropic's own Economic Index work, in the Massenkoff and McCrory measure released in March 2026, observed a slight decline for the most exposed young workers and stressed the wide gap between what AI can do and what it is observed doing.

CAIT reconciles these camps more than it contradicts them. Its statewide null result matches Yale and Denmark. Its elevated subgroup claims among educated, Bay Area, and tech-sector workers match the direction of Stanford's findings, while drawing on a different data source: layoffs that reach unemployment insurance, where Stanford counts payroll headcount. The Stanford work captures slower hiring and attrition combined. CAIT captures people who lost jobs and filed for benefits. That the two line up in the same places, among the most exposed workers, strengthens both.

SPONSORED

Helping HR, talent acquisition, employer branding, and company culture professionals find careers worth smiling about.

Find People-First Jobs Now

There is also an employer-sentiment counterweight that talent leaders should hold alongside the claims data. The Strada Institute for the Future of Work surveyed nearly 1,500 executives and senior talent leaders in spring 2026. Among employers that had at least explored AI, 46% reported that AI increased their entry-level hiring in 2025, against 13% reporting a decrease, and 2.7 times as many expected AI to increase entry-level hiring in 2026 as expected it to decrease. The same survey found the work itself changing: 42% said analytical and judgment-based responsibilities are growing for entry-level staff, and 41% said routine and administrative tasks are being stripped away. Strada's research director Andrew Hanson summarised the shift by describing entry-level roles becoming more like mid-level roles.

That is the synthesis worth carrying into a hiring strategy. The displacement signal is concentrated, not universal. The job-design signal is universal. The content of junior work is changing almost everywhere, even where headcount is not falling.

The limitations, stated plainly

The report is unusually candid about what it cannot do, and citing it accurately means carrying those caveats.

Exposure is not adoption. The measures capture what AI could do, or how often Claude is used on similar tasks, not whether a specific employer adopted AI or used it to justify a specific layoff.

UI claims are a partial lens. They capture only workers who lose a job and file for benefits. They miss people who do not file, who find new work quickly, who leave the labour force, who are self-employed or gig workers, or who are ineligible. Younger workers are underrepresented for exactly this reason.

Causation is unproven. The authors describe their findings as an early, descriptive signal of potential AI-related disruption, not a full accounting of AI-related job loss and not causal evidence. They cannot separate post-pandemic tech-sector adjustment from emerging AI pressure with the data in hand.

Employer branding works better when people share what they know. EBClub brings practitioners together to exchange ideas, compare experiences, and learn from the people doing the work.

Explore EBClub

A validation check adds confidence without resolving the causation question. The team identified six large employers that publicly announced AI-driven mass layoffs between March 2024 and April 2025, located the layoff dates through legally mandated WARN filings, and confirmed clear spikes in claims from high-exposure workers around those dates. Roughly 58% of affected employees filed for UI within 10 weeks. That shows the instrument can detect AI-linked layoffs when they happen. It does not show that such layoffs are yet large enough to move statewide totals.

What this means for employer brand and talent teams

The tracker is a labour-market instrument, and it carries direct uses for the people who shape how an organisation is seen by candidates and staff.

Calibrate the AI message to the evidence. Candidates are anxious, and that anxiety is rational given the public statements coming from AI companies. CAIT gives talent teams a defensible position to hold: no broad collapse, real pressure in specific corners, and honest uncertainty about cause. An employer value proposition that acknowledges the uncertainty will read as more trustworthy than one that promises AI will only ever augment and never displace.

Treat the entry-level pipeline as the live issue. The strongest convergent signal across CAIT, Stanford, and Strada is that the first rung of the career ladder is changing fastest. The Washington Monthly framed it through Strada's data as "experience creep," with employers wanting more experience for jobs that once developed it. An employer brand that still markets graduate roles as simple on-ramps will misrepresent the work. Teams that redesign junior roles around judgment and AI-direction, and say so openly in their hiring content, will recruit better and set fairer expectations.

Audit which roles sit in the high-exposure band. The exposure rankings are public. Computer programmers, customer service representatives, data entry keyers, and statistical assistants score high. An organisation whose workforce concentrates in those occupations will face sharper questions from candidates and current staff, and its internal communication needs to get ahead of them.

Use the regional and sector lens for benchmarking. A technology employer in the Bay Area is operating in the hottest part of this dataset and should expect more candidate scrutiny than a manufacturer in an inland region. Finance and Insurance carries the largest share of high-exposure claims of any sector. Knowing where an organisation sits helps set a realistic tone.

The Exit Interview - EBN

Anonymous practitioner takes on employer branding, talent, and recruitment. The things that rarely survive the official version. Published by Employer Branding News.

Employer Branding NewsUnnamed CHRO

Borrow the report's discipline for internal communication. During restructures, the instinct is to smooth the story. CAIT shows the alternative: state what is known, name what is uncertain, and avoid claiming more certainty than the evidence supports. For any communication touching layoffs or AI adoption, that posture protects long-term trust with both staff and the external market.

The bigger picture

California has done something other states and countries will likely copy. Bloomberg Law has already reported on New York weighing how to track AI-related layoffs, and the report frames California as a model for others. For employer brand and talent functions, the arrival of public, monthly, government-grade tracking changes the terms of the conversation. The question is moving from speculation toward measurement.

That shift cuts both ways. It gives talent leaders harder evidence to counter the loudest predictions. It also means that when disruption does show up in a region or a sector, it will show up in a public dashboard that candidates, journalists, and competitors can read. The organisations that come out of this period with their reputations intact will be the ones already speaking about AI and work in the careful, evidence-led register that CAIT itself models.

Two questions are worth sitting with. If the displacement signal is currently concentrated among educated workers in exposed jobs, what does an employer brand owe the early-career candidates who can already see the ladder changing beneath them. And when a tracker like this one eventually does register a clear AI-driven shift in a given sector, will the organisations in it have built the trust to be believed when they explain themselves.

Takeaways

SOURCES