The Automation Ledger

The problem

Ask a simple question: how many American jobs were eliminated by automation last year?

Nobody knows. Not the Bureau of Labor Statistics, which measures employment and separations but never asks why a position ceased to exist. Not the Department of Labor, whose WARN notices record mass layoffs and plant closings but not their cause. Not the Census Bureau, whose surveys of business technology adoption are voluntary, lagging, and not linked to headcount. Not the companies, most of which have no obligation to say and every incentive not to.

So the most consequential economic transition of our lifetimes is being argued entirely on anecdote. One side cites a call center that cut staff after deploying a model. The other cites the aggregate unemployment rate and declares the whole worry a moral panic. Both are reasoning from data that cannot answer the question, and neither can be proven wrong, which is why the argument never ends and never improves.

This is not a new failure. America spent three decades arguing about whether trade or technology destroyed manufacturing employment, and the argument was only settled — partially, late, and after the political damage was done — because researchers eventually assembled the data by hand from customs records and county employment files. We are now walking into a larger transition with worse instrumentation.

The absence is not neutral. It systematically favors whoever prefers no policy, because every proposal can be met with "you have not shown the effect is real." That is not skepticism. It is a structural advantage handed to inaction by a measurement gap.

The proposal

Require large employers to report automation-driven changes in headcount, on the same machinery that already reports layoffs.

The Worker Adjustment and Retraining Notification Act of 1988 already requires employers with 100 or more employees to give 60 days' notice of a plant closing or mass layoff. The reporting relationship exists. The threshold exists. The enforcement mechanism exists. What is missing is a single additional field: why.

How it would work

1. Extend WARN with a cause code. Employers filing a WARN notice select from a short, defined list: demand decline, relocation, merger, closure, automation or software substitution, or other. Where automation is selected, a brief structured description of the function replaced is required. This is a checkbox and a sentence, not an impact assessment.

1. A standing annual filing for the largest employers. Firms above a higher threshold — say 5,000 employees — file annually on roles eliminated, roles created, and roles materially changed by automation, by occupation code and location. Aggregated and published; never individual-level.

1. Publish it like an economic statistic. The Bureau of Labor Statistics releases the series quarterly, on a fixed calendar, with documented methodology, prepared by career statisticians. It becomes a number like the unemployment rate — contestable, revised, and shared, so that the debate is at least about the same reality.

1. Reporting only. No penalty attaches to the number. This is the design decision that determines whether the proposal works or is gamed to death. An employer who reports automating 400 roles faces no tax, no mandate, and no liability for the disclosure itself. The moment the number carries a penalty, the number becomes fiction.

1. Protect the firm, publish the aggregate. Individual filings are confidential to the statistical agency, as with existing BLS establishment data. What is published is industry, occupation, and regional aggregates. Trade-secret objections are answered the way they are answered for every other establishment survey.

The numbers

This is the cheapest proposal in the catalog. WARN filings already run through state dislocated-worker units; adding a coded field and a structured annual return for the largest firms is a marginal administrative cost, plausibly in the low tens of millions of dollars a year — a rounding error against the $70 billion of the Caregiver's Wage (№ 021) or the billions moving in the Grid Dividend (№ 024).

What it buys is the ability to answer questions that currently have no answer. Which occupations are actually contracting, and how fast. Whether displacement is concentrated in a few metros or spread thin. Whether the firms automating most are also hiring most, which is the central empirical dispute and is currently unresolvable. Whether the effect is arriving as layoffs or, more likely, as quiet attrition — positions that are simply never refilled, which no existing instrument detects at all.

That last mechanism is why WARN alone is insufficient and the annual filing matters. A firm that automates by not replacing 800 people who leave over three years triggers no WARN notice, generates no layoff statistic, and appears in the data as a healthy company with low turnover costs. That is the likeliest shape of the transition and it is currently invisible.

The honest objections

"Employers will call everything 'other' and the series will be garbage." The most serious objection, and partly correct — self-reported cause data is noisy everywhere it exists. Three mitigations, none complete. The categories must be few and concrete, since long taxonomies collapse into "other." The occupation-and-function detail makes miscoding checkable against the firm's own job postings. And a false statement on a federal filing already carries consequences, which is a different thing from penalizing the underlying fact. We expect the first two years of data to be poor. Every economic series began that way; the unemployment rate took decades to become trustworthy.

"'Automation' cannot be cleanly defined. Every efficiency is automation." True, and the line between a new software tool and a replaced worker is genuinely blurry. We would rather collect a contested definition consistently than collect nothing precisely. The definition should be set by the statistical agency, published, and revised openly, exactly as the definitions of unemployment and underemployment have been.

"This is a compliance burden that will land hardest on mid-sized firms." Which is why the annual filing sits at 5,000 employees rather than 100, and why the WARN change is a single field on a form already being filed. If the burden estimate comes back higher than a few hours per filing, the design is wrong and should be simplified rather than defended.

"You are building the evidence base for a robot tax." We are building an evidence base. Where it leads is a separate argument, and this catalog's own answer to displacement is ownership (№ 001) and a landing (№ 026), not a levy on machines. But the objection deserves a direct answer rather than a denial: yes, better data makes some interventions more likely, including ones we would oppose. We think a policy argument conducted on evidence beats one conducted on anecdote even when the evidence cuts against us. That is the whole premise of this organization.

"By the time the series is reliable, the transition will be over." Possibly. This is the argument for starting now rather than after the next election, and for accepting a noisy series immediately over a clean one in 2032.

Sources

• Worker Adjustment and Retraining Notification Act of 1988, 29 U.S.C. § 2101 et seq. (dol.gov)
• Bureau of Labor Statistics, Job Openings and Labor Turnover Survey and establishment data confidentiality practice (bls.gov)
• Census Bureau, Annual Business Survey technology-use modules (census.gov)
• Acemoglu & Restrepo, "Robots and Jobs: Evidence from US Labor Markets," Journal of Political Economy (2020); Autor, Dorn & Hanson on the China shock, on the difficulty of attributing employment loss after the fact
• Proposals № 001 (The American Dividend), № 021 (The Caregiver's Wage), № 024 (The Grid Dividend), № 026 (The Landing)