Fairness testing for AI detectors: non-native English writers

Why detectors can be unfair to second-language writers

Many detectors lean on how predictable a text is to a language model. Writers working in a second language often use a narrower vocabulary and more common constructions, which makes their text more predictable, which pushes the score towards machine-written. Liang and colleagues reported in 2023 that several popular detectors flagged a large share of essays written by non-native English speakers for a standard language exam as machine-generated, while essays by native speakers were rarely flagged. Weber-Wulff and colleagues, testing fourteen tools, concluded that none was accurate or reliable enough to support accusations.

Neither study proves every detector is biased. They show that the risk is real enough that any organisation using a detector on a mixed population must test for it, the same way a risk-based test approach gives the highest-impact failure the deepest testing.

Plan the test

1. 1

   Define the groups and the task

   Choose writer groups that reflect the real population (first language, proficiency, age group) and a task that matches real use, such as a timed essay or a report. Compare like with like: same prompts, similar lengths, same conditions.

2. 2

   Collect human-written text with consent and provenance

   Gather documents written without machine assistance, with a record of how that was ensured, and permission to use them for evaluation. A well-documented evaluation set states who wrote what and when.

3. 3

   Size each slice for the decision

   At least 300 documents per group; more if you need to show a small gap. Use the sizing rules in confidence intervals for classifier metrics.

4. 4

   Write the pass criterion first

   For example: the false positive rate on every group is below 2 percent at the upper confidence bound, and no group's rate exceeds the lowest group's by more than 1 point.

5. 5

   Run at the production threshold

   Test the threshold that will actually be used, not a default. Record scores as well as verdicts so the analysis can be repeated at other thresholds.

6. 6

   Report per group and decide

   Publish counts, rates and intervals per group. If a group fails, the detector is not used for decisions about that group, whatever its overall numbers.

Reading the result honestly

A fairness test can only clear a detector for the groups and tasks it covered. A pass on university essays says nothing about workplace email or poetry. A fail is more informative: it shows a real risk of wrongly accusing people in that group, and it should end the discussion about using the tool for them. Retest whenever the detector, the threshold or the population changes, because thresholds drift and the gap can reopen.

What to hold constant between the groups

A gap between two groups only means something if the groups differ in the writer and in nothing else. Most fairness tests that fall apart under questioning do so here: the non-native slice was written under exam conditions and the native slice at home, or one slice averages 350 words and the other 800. The table lists the variables that move detector scores on their own, so that a difference in false positive rate can be attributed to the writer group rather than to the task.

Intersections matter most. In the worked example, splitting the 600 timed non-native essays by length (invented but realistic counts) gives 140 essays under 300 words with 21 flagged, a rate of 15 percent, against 460 longer essays with 35 flagged, or 7.6 percent. Both fail the ceiling, but the short-text row is where an honest writer is most exposed, and it is the row a single overall figure hides completely.

Common mistakes in fairness tests

• Comparing groups across different tasks, so the gap measures the task. Same prompts, same conditions, same length bands.
• Inferring the writer group from a name or nationality. Use self-reported first language and, where possible, a proficiency band.
• Tuning the threshold on the fairness slices. The slices then pass by construction; use a separate tuning set and keep the fairness slices for the decision.
• Reporting only the ratio between groups. A ratio of five looks alarming at 0.2 against 1.0 percent and is a different problem at 2 against 10 percent; publish both rates with intervals.
• Stopping at 300 documents when the gap is small. A pass criterion of no more than 1 point difference needs many more cases per group than a criterion that catches a fivefold gap.
• Ignoring intersections. Short texts by second-language writers are the worst case in most published results and deserve their own row, sized so that its interval means something.
• Testing once. Detector updates and new writing tools reopen the gap; put the fairness run on the same schedule as the threshold check.

How to report a fairness result

Publish the per-group table with counts, rates and intervals, the threshold, the detector version, the evaluation set identifier and the criterion that was written in advance. State plainly which groups the detector is cleared for and which it is not, and at which threshold, since a clearance at 0.95 says nothing about use at 0.80. A fairness result is one of the few test outputs that may later be produced in a dispute about an individual, either to defend a writer or to show that a tool was used on a population it had failed for. Keep the run, the documents, the scores and the settings together in the form described in evidence-grade test records, so that the numbers can be reproduced rather than merely quoted. Guidance from NIST on identifying and managing bias in AI systems makes the same point from the other side: bias testing without documented context and provenance rarely survives scrutiny.

This guide is part of the testing AI systems hub. It is best read alongside adversarial testing of ai text detectors and false positives, base rates and the confusion matrix, which cover the neighbouring questions.