GPT-2 detection model flags autistic writing at elevated rates

The study's most underreported implication is not that AI detectors make mistakes, but that their error rates are unevenly distributed across human populations, meaning the systems function as intended on average while quietly failing specific groups at rates that aggregate metrics will never surface.

This connects directly to our coverage of 'Innocuous-Seeming Data, Latent Ideology' from the same day, which showed that models absorb and amplify value systems embedded in training data without explicit instruction. The same mechanism is at work here: GPT-2 detection models trained on neurotypical writing distributions encode an implicit baseline of what 'human' text looks like, and autistic communication patterns fall outside that baseline statistically rather than semantically. The ideological generalisation paper framed this as a deployment risk for fine-tuned LLMs; this study shows the same latent-encoding problem applies to safety tooling itself, which is arguably the higher-stakes surface. Both findings converge on a single uncomfortable point: bias audits that measure aggregate false-positive rates are insufficient when the affected populations are small enough to disappear inside the average.

Watch whether platforms using GPT-2-era detection in content moderation (Reddit is named in the dataset) publish disaggregated false-positive audits broken down by community or writing style within the next 12 months. If they do not, that absence is itself informative about how seriously the finding is being operationalized.