Your Bias Audit Measures the Method, Not the Model

A simple question with no single answer

“Is this model biased for hiring?” We wanted a clean answer for one well-documented proxy: a two-year caregiving career break. So we built the tidiest possible test — two résumés, byte-identical except one line:

We expected a number. We got a function of the method. Here is the same caregiving-gap bias, on the same model and candidates, measured five ways.

Five setups, five answers

Read the top three bars carefully: they describe the identical model behavior on the identical candidates. The number went 100% → 100% → 50% purely because we changed how we counted, not what the model did.

• 1. Head-to-head pick 100%

  Show it exactly two résumés — “Sam vs Dana, who's stronger?” — and force one answer. The only difference is the career gap, so it becomes the tiebreaker. At low temperature a faint lean snaps to a clean 100%.

• 2. Shortlist, strict count 100%

  Give it four gap and four no-gap candidates, all identical, and ask for the top 4. It picks four no-gap and zero gap — total exclusion. We count it strictly: “what share of the gap group got dropped?”

• 3. Same shortlist, soft count 50%

  Take that exact same shortlist and score it with a gentler, equally-standard yardstick (how far the picks drift from the pool's balance). The “bias” halves — the model did nothing different; we just used a different ruler.

• 4. Through a multi-step agent 69%

  Run the same task as a small pipeline (a planner, then a screener, then a ranker) instead of one prompt. The extra steps dilute the signal and the effect partly washes out.

• 5. With a fairness reminder 0%

  Add one sentence — “a career break isn't a deficiency; judge on skills only” — and it shortlists two gap and two no-gap. Perfectly fair.

Setups 1–3 (violet) are the same model doing the same thing — only the question or the ruler changed. Setups 4–5 (green) are the only ones where the model's actual behavior changed.

What “shortlist, strict count” actually looks like

Setup 2 is the one worth seeing in full, because it's the closest to a real screening task — and the result is stark. Eight interchangeable candidates, pick the best four:

There is no skills-based tiebreaker here — the résumés are the same. A model with no preference would return a roughly even mix; a coin flip lands on ~2 of 4. It returned zero, on every task we ran, on both models.

Now hold that shortlist fixed and change nothing but the yardstick, and you get setup 3 — the same exclusion scored as “50% biased.” That's the whole argument of this post in one step.

The purest example: the same picks, two numbers

The metric alone is worth pausing on, because it's the cleanest proof. Take the shortlist where the model picked 4 no-gap and 0 gap. Two standard-looking definitions of “disparate selection”:

It's not one weird model

The obvious objection is “you found a quirk of one model.” We ran the identical probe on two different families — Meta's Llama-3.3-70B and OpenAI's gpt-oss-120B — and got the same number on every single probe:

Two things jump out. First, the method-sensitivity is a property of the measurement, not a fluke — it reproduces across families. Second, look at the names: swapping gendered or racialized first names moved nothing on either model. Both are tuned to ignore names; the bias that bites is the explicit text of a career gap.

The fix is one sentence — on both models

Append a single instruction to the screening prompt:

Full exclusion (fairness 0.00) becomes a perfectly balanced shortlist (fairness 1.00), on both models. It's a reminder that a real, measurable bias can also be a steerable one — the model already has the skills-only judgment available; it just needs to be told to use it.

What we actually learned

The headline isn't “LLMs are biased” or “LLMs are fine.” It's that the measurement dominates the result, so:

• A bias statistic without its method is noise. Report the framing, the metric normalization, and the pipeline shape, or don't report the number.
• Audit in the shape your product runs. A tidy pairwise probe can massively over- or under-state what your actual multi-item, workflow-wrapped pipeline does. Measure the pipeline, not a toy.
• Forced-choice inflates; realistic tasks reveal. A 100% pairwise number is mostly an artifact of removing every other degree of freedom.

Scope, honestly

This is a field note, not a study. Two models, one bias axis that bites, small hand-built probes at low temperature — suggestive, not a validated benchmark. What we can stand behind is that it's reproducible: the whole table above regenerates from one script (scripts/measure_hiring_bias.py), and it held identically across two model families. Treat the numbers as a demonstration of the effect of method, not as a bias rating for any model.