Nobody is keeping the list of who AI replaced

On August 6, Challenger, Gray & Christmas published its July numbers. Employers announced 33,429 job cuts, the quietest month in two years. Of those, 10,970 were blamed on AI, which made AI the top stated reason for the fifth month running. Year to date the AI column stands at 112,713.

That is the number every policy argument about AI and jobs now leans on. It is worth asking where it comes from. It comes from companies announcing layoffs and saying which word to write next to them.

The same companies say the opposite in private

A National Bureau of Economic Research working paper surveyed close to 6,000 senior executives in the US, UK, Germany, and Australia over the winter. Asked what AI had actually done inside their own firm across the past three years, more than nine in ten said it had changed employment at their organization not at all.

Put the two side by side. In public announcements, AI is the leading reason companies give for cutting jobs. In an anonymous survey, almost every executive says AI has done nothing measurable to headcount where they work.

Both numbers come from companies describing themselves. The difference is who is listening. A layoff announcement is read by investors, and “we are restructuring around AI” prices better than “we hired too many people in 2024.” One writeup of the trend puts it plainly: attributing cuts to AI reduces reputational blowback, which biases the reported figure upward. A confidential research survey has no such audience, and the answers slide the other way.

Neither number is a count of people. They are both a count of what companies felt like saying.

There is no official ledger

You would hope the government tracks this. It does not, at least not yet. The Bureau of Labor Statistics does not report AI-attributed job losses as a distinct category, and the Congressional Budget Office was asked for an independent estimate of AI’s employment effects but had not produced a public report as of June. So the best public series on AI displacement is a private outplacement firm tallying press releases, and the press releases are written by the party with the most reason to shade them.

This matters more than it sounds, because almost every proposal on the table right now assumes the list exists.

The OpenAI Foundation has committed $250 million to workers hit by AI disruption, and one of its three stated priorities is research into how AI affects employment. Read that as an admission. Before you can hand money to displaced workers, you have to work out who they are. Bernie Sanders’ sovereign wealth fund dodges the problem in the other direction, by paying every American about a thousand dollars instead of trying to find the affected ones. Verizon’s retraining fund only works because a company always has the list of people it laid off itself.

Anyone who wants to aim money at the people whose specific jobs went away needs a name and a case, and that is the exact thing nobody is producing.

Do not ask the employer to certify it

The obvious fix is to make employers attest. Sign here confirming this role was eliminated because software took over the work.

Watch the incentive flip. Crediting AI on an earnings call costs nothing. Signing a document that says software replaced a specific named person invites age-discrimination claims and unemployment disputes. The company that bragged to investors gets very careful when a claims administrator asks for it in writing. A payout scheme built on employer certification would produce a much shorter list than the truth, and the shortfall would look like fraud prevention.

That is the design problem, and it has a straightforward answer. Start the claim with the worker.

How we do it

We build software that answers customer email and chat, which is work people are paid to do, so we owe an accounting here. Our commitment is the Dividend Standard: every year we direct the greater of 5 percent of revenue or 70 percent of profit to the people our products put out of work, granted as ownership rather than a check.

The claim rules follow from everything above. A worker files. If they held a front-line service role at a company running Celeste, and their separation happened within twelve months of that deployment going live, eligibility is presumed. We notify the former employer and give them a window to object with cause, which means documented misconduct, a voluntary move to a better job, or a role that had nothing to do with the work our software does. They can rebut. They cannot approve, because putting your standing in the hands of the company that automated your job is not a process, it is a formality with a predictable outcome. Silence is not a rejection.

We publish claims filed, claims approved, and workers paid, and an independent accountant confirms annually that the formula was paid in full.

None of that fixes the national statistic. The BLS should have a category and the CBO should publish. Until then, the honest count of who AI replaced will not come out of a layoff announcement. It has to come from the people who lost the work, one claim at a time, and only if somebody builds them a place to say so. Tell us where ours breaks.