A new statistical re-analysis finds no solid evidence that AI language models recognize or favor their own writing.
Researchers rebuilt the generator-by-selector matrices from Laurito et al.'s 2025 PNAS study, which had found that AI models judging between AI-written and human-written descriptions consistently pick the AI version. Using the original dataset of 21,828 trials across five generator models and five selector models, the new analysis isolated an own-model effect: whether a model recognizes and favors text it generated itself, separate from a general preference for AI-written prose. The measured premium was small and not statistically significant in every category tested: 1.3 percentage points on product descriptions, -1.0 on paper abstracts, and 5.4 on film descriptions, with a pooled estimate of 1.9 points (p=0.14). The study's design had enough statistical power to catch a true effect of 5 percentage points in most categories, so the absence of a signal is informative rather than just inconclusive.
That distinction matters because the original study's headline finding - that AI models share a taste for AI-written text - has been read in some corners as evidence that models can spot their own output, a capability with real stakes for detecting AI-generated content and for policing self-preference in benchmarks or hiring tools. This analysis shows generalized bias, not self-recognition, explains the pattern: GPT-4's descriptions were picked 77 to 95 percent of the time by every selector model tested, not disproportionately by GPT-4 itself. A smaller 2024 study did find a self-preference signal, but its sample was a fraction of the size, which makes the comparison shaky.
AI models, it turns out, are snobs about AI-sounding prose in general - they just don't play favorites with their own words, at least not in any way these numbers can detect.