The Verification Gap: Who Holds AI Layoff Claims Accountable?

Employers are citing artificial intelligence in layoff announcements, and lawmakers are beginning to ask how these changes should be documented. The concern is understandable. If AI is transforming the workforce, we need some way to measure its effects. But before we can measure those effects, we must establish what actually caused the changes.

This piece is an argument and a set of design questions, not a new investigation.

States including Connecticut and California have begun addressing this problem through existing labor-notification requirements.

Connecticut’s Public Act 26-15, which took effect October 1, 2026, requires employers filing federal WARN notices to disclose whether their layoffs are related to artificial intelligence or another technological change.

California has taken a somewhat different approach. Its SB 951, signed September 30, 2026, takes effect January 1, 2027. When a covered workforce reduction is caused in whole or in substantial part by AI or another automated technology, employers must provide additional information, including the occupations affected, the job functions being automated, and the type of technology involved.

The distinction matters. Connecticut asks employers to identify a possible relationship between layoffs and technology. California requires more descriptive information when a reduction meets the law’s threshold, which the statute does not define. In practice, the employer makes the first assessment.

Neither approach, by itself, establishes whether the employer’s explanation is accurate. Nor does either guarantee that every relevant case is reported: an employer that concludes AI was not a substantial cause files none of California’s additional AI-related detail.

These measures are meaningful developments in workforce reporting. They acknowledge that technological displacement deserves closer examination. Yet they also expose a familiar problem: the institution making the claim can control much of the evidence needed to verify it.

The Problem We Encountered at Kroger

Our recent investigation, When Does a Layoff Become an AI Layoff?, confronted this difficulty directly.

Kroger announced a substantial corporate workforce reduction after outlining an expanded technology strategy. The timing raised an obvious question: had artificial intelligence contributed to the decision?

We designed five tests to examine the available evidence rather than relying on the sequence of announcements. Two findings, on timing and on peer companies, were consistent with ordinary restructuring but remained provisional because some of their sources were not fully pinned; a third, on the company’s stated rationale, was mixed.

Two of the tests most relevant to determining AI’s contribution could not be completed using the public information we examined. We lacked sufficiently detailed workforce reductions by function and the hiring information needed to evaluate whether AI-related positions were replacing other roles.

The result was not evidence that AI played no part. Nor was it evidence that AI caused the layoffs.

It was a limit on what we could establish.

This is where the governance question begins. If companies possess the operational records needed to distinguish technological displacement from ordinary restructuring, how can the public evaluate their explanations?

The new state disclosure requirements address parts of this information gap. But neither, as written, guarantees access to every kind of evidence our Kroger investigation required.

Counting Claims Is Not the Same as Measuring Causes

The problem extends beyond individual companies.

According to Challenger, Gray & Christmas, artificial intelligence was cited in 120,136 announced U.S. job cuts during the first nine months of 2026, approximately 21 percent of all announced cuts in its reporting.

That is a consequential figure. It tells us that AI is a prominent explanation associated with workforce reductions.

But what exactly does it measure?

The Challenger reports classify announced job cuts by reason. The public reports we examined do not explain the attribution process in enough detail to establish that every cut assigned to AI resulted from independently verified technological displacement.

There are several competing possibilities. AI may genuinely allow companies to perform some tasks with fewer employees. It may contribute alongside cost pressures, organizational consolidation, or declining demand. In other cases, executives may emphasize AI when describing restructuring whose causes are more complicated.

We should not assume any one explanation applies universally.

The distinction is straightforward: a reason recorded in the tracker is evidence of an attribution, not necessarily proof of the underlying cause.

This does not make the Challenger figures meaningless. They are useful evidence of how announced cuts are being attributed. But interpreting them as a direct measurement of jobs eliminated by AI would go beyond what the available methodology establishes.

From Disclosure to Verification

Federal lawmakers have also begun examining the reporting problem.

In November 2025, Senators Josh Hawley and Mark Warner introduced the bipartisan AI-Related Job Impacts Clarity Act, S.3108, which was referred to committee. The bill proposed additional reporting of AI-related employment effects through a federal framework.

Such measures could improve the information available to policymakers and researchers. But collecting information and establishing its accuracy are different responsibilities.

A reporting system can tell us who submitted a statement, when it was submitted, and what the employer declared. It cannot automatically tell us whether that statement correctly explains the workforce decision.

This is the distinction we have returned to repeatedly in our work:

Provenance is not verification.

Knowing the origin of a claim is necessary for accountability. It is not sufficient to establish its truth.

If the objective is to understand AI’s employment consequences, the next challenge is to determine what evidence should support these disclosures and how that evidence could be examined independently.

Before proposing any new oversight institution, we should examine whether the WARN reporting systems already in place could support more reliable and independently verifiable information.

I see three possibilities worth examining.

Three Questions for a Verifiable Reporting System

1. Can we identify the work that actually changed?

A company’s announcement that AI contributed to layoffs tells us little about the underlying mechanism.

A more informative report might identify which tasks were previously performed by employees, which technologies were introduced, and how the organization changed afterward.

Existing occupational classifications and task descriptions could provide a common vocabulary for recording these changes. But exposure to AI is not equivalent to replacement by AI. A task may be technically automatable without a company automating it, and automation may change a job without eliminating it.

California’s new reporting requirements move toward this level of detail by asking employers to identify affected job functions and technologies.

There is an important limitation, however. Those additional fields apply when the employer’s reduction meets the law’s technological-displacement condition. The legislation does not independently establish whether every employer has correctly determined that condition.

The question is whether task-level information can help distinguish actual deployment from assumptions about what AI might be capable of doing.

2. What evidence should be available for independent examination?

A disclosure requirement becomes more useful when there is a way to challenge an inaccurate statement.

That does not necessarily mean every employer should publish confidential operational records. Companies have legitimate concerns about proprietary information, employee privacy, and the cost of compliance.

But those concerns do not eliminate the public interest in reliable employment data.

One possibility would be to require employers to retain supporting records that an appropriately authorized reviewer could examine when a material claim is disputed.

Such a system would need clear rules governing access, confidentiality, evidentiary standards, and enforcement.

Existing WARN procedures do not automatically provide a comprehensive federal mechanism for investigating the causal accuracy of AI-displacement claims. State enforcement arrangements differ, and the reach of existing authorities over these new disclosures requires careful legal examination.

We should establish those jurisdictional limits before proposing additional powers.

3. How do we establish accountability without creating an unreasonable burden?

Any verification system must distinguish between an employer acting in good faith and one knowingly providing misleading information.

Workforce decisions often have more than one cause. A company may adopt AI while also reorganizing departments, reducing costs, or responding to changes in demand.

Requiring absolute certainty about the contribution of each factor could make reporting impractical. Accepting unsupported declarations, however, would undermine the purpose of disclosure.

A workable approach might allow employers to describe multiple contributing factors, disclose the evidence supporting their assessment, and identify where attribution remains uncertain.

Independent audits, review thresholds, and protections for good-faith reporting deserve examination. But these mechanisms would need testing to determine whether they improve accuracy without imposing disproportionate costs or becoming procedural formalities.

These are design hypotheses, not established solutions.

What Should Count as Evidence?

The larger issue is not whether artificial intelligence will affect employment. It is how we distinguish its actual effects from the many other forces shaping the labor market.

Better reporting could help researchers understand displacement, help workers anticipate changes, and help governments design appropriate responses.

But additional data fields alone will not settle questions of causation.

We need to know what changed inside organizations, what evidence supports the explanation, and whether independent examination could reasonably reach the same conclusion.

The emerging legislation offers an opportunity to develop that discipline. Whether it succeeds will depend not simply on what employers are required to report, but on how those reports can be evaluated.

Our Kroger investigation illustrated how far a careful inquiry can go with public evidence—and where it stopped.

The next question belongs to governance.

When a company attributes the disappearance of jobs to artificial intelligence, who has the evidence, who can examine it, and who is accountable if the explanation does not hold up?

That is the verification gap we need to address.

Research and editorial disclosure: I developed the central argument and wrote the original draft. ChatGPT substantially restructured and rewrote the essay, including its headings and several passages. Claude and Lumen contributed research review, criticism, and suggested revisions. I evaluated the collaborators’ contributions and retain editorial responsibility for the final text. The proposed governance mechanisms are hypotheses for further examination, not validated policy solutions.

Dale Joseph is the author of Thought Partners: Preserving Cognitive Sovereignty in the Age of AI and founder of the Emergence Institute. He worked for years as a consultant helping install hospital networks before turning to writing and systems thinking. He lives in Boynton Beach, Florida.