Worked example: reading a bias audit without misreading it

A recruitment vendor sends your HR director a one-page PDF headed "Bias Audit Summary" and a covering note saying the tool passed. The hiring team wants to switch it on next week. What does the page actually establish, and what is it silent about?

What New York requires

New York City Local Law 144 governs automated employment decision tools used for hiring or promotion in the city. Its rules sit at 6 RCNY sections 5-300 to 5-304. The duty has two halves: an independent bias audit no more than one year old, and publication of a summary of its results.

The audit has to calculate specific things. The selection rate for each category. The scoring rate, meaning the proportion of people in a category who scored above the sample median. And the impact ratio, comparing each category against the most-selected or highest-scoring category. These are computed for sex, for race and ethnicity, and for intersectional categories, so not just women and not just Hispanic applicants but Hispanic women as their own line.

Categories representing less than 2 per cent of the data may be excluded, provided the exclusion and the justification are stated. The published summary must carry the date of the audit, the source and explanation of the data, the number of individuals assessed in each category, the rates, the impact ratios, the number of individuals in an unknown category, and the distribution date of the tool. Candidates must be notified at least 10 business days before the tool is used on them.

The sentence that is wrong

The covering note says the tool passed because every impact ratio is above 0.80.

There is no pass mark in Local Law 144. The four-fifths figure comes from the EEOC Uniform Guidelines, a different instrument in a different context. Local Law 144 requires you to calculate the ratios and publish them. It does not set a threshold, and it does not declare a tool compliant or non-compliant on the basis of one. A vendor who says "we passed" has imported a number from elsewhere and attached it to a law that does not contain it.

Why does this matter beyond pedantry? Because "passed" ends the conversation, and the ratios are supposed to start one. A ratio of 0.82 for one group is a published fact that a regulator, a journalist or a rejected candidate can read and ask about. Treating 0.81 as success and 0.79 as failure invents a cliff edge the law does not have, and it invites the practice of tuning a model until it clears an imaginary bar.

What the page does not tell you

Whose data. The summary must explain the data source, but an audit run on the vendor's historical customer pool says little about your applicants. If your applicant mix differs, your ratios will differ.

Which categories vanished. Check the exclusions. A category dropped under the 2 per cent rule is a category about which the audit says nothing, and small groups are precisely where trouble concentrates.

How many were unknown. The unknown-category count is a required figure because it bounds how much the audit can claim. If a third of applicants did not state a category, the ratios describe two thirds of the picture.

The date. No more than one year old is the rule. An audit from fourteen months ago does not satisfy it, however good the numbers are.

What else is in play

Local Law 144 is not the only instrument a hiring tool meets. Illinois HB 3773, effective 1 January 2026, amends the Illinois Human Rights Act to prohibit the use of AI that has a discriminatory effect in employment decisions, prohibits using zip code as a proxy, and requires notice that AI is being used. Note the word effect: intent is not the test. Under the EU AI Act, employment and worker management is listed in Annex III as high-risk, and Article 26(7) requires a deployer to inform workers' representatives and affected workers before an AI system is put into use in the workplace.

What to send back

Ask for the audit date, the data source with an explanation of whose applicants it describes, the per-category counts including the unknown count, the stated exclusions with justification, and the raw rates rather than a verdict. Then ask your own team which definition of fairness this tool was tuned for. If nobody can answer, the answer is aggregate accuracy.

As of 24 September 2026.