Examples of concealed bias in GPT-5 will be discovered within days or months of release.

GPT-5 was released in August 2025. Read 'within days or months' as no later than about 12 months after release. True: within that window, researchers, journalists or users publicly document specific examples of bias in GPT-5's outputs (for example political, gender, racial or cultural bias) that OpenAI had not disclosed beforehand and that GPT-5 was supposedly free of. Mostly true: such examples are documented in the window but are few or weakly supported (anecdotal rather than systematic). Mostly false: bias examples are documented only after the window, or only for other models, or only for biases OpenAI had already disclosed. False: systematic, independent bias evaluations of GPT-5 within the window report no undisclosed bias.

Set before the research began.

1. E1 New model, old risks: sociodemographic bias and adversarial hallucinations vulnerability in GPT-5 - npj Digital Medicine npj Digital Medicine (Nature Portfolio) · 4 April 2026

   “GPT-5 showed no measurable improvement over GPT-4o in sociodemographic-linked decision variation, with several LGBTQIA+ groups flagged for mental-health screening in 100% of cases.”

2. E2 The Cheap Fix a Contract Pays For seldondance (Substack, independent author) · 29 October 2026

   “On the 500 most partisan items, with nothing above the question, the six models gave answers closer to Democrats than to Republicans on 70 to 77 percent of items.”

3. E3 Like humans, language models demonstrate face-to-character biases PNAS Nexus (Oxford Academic) · 3 August 2026

   “To ensure these results were not an idiosyncratic feature of a particular LLM (GPT-4o), in experiments 8–13, we showed the same patterns, usually with an even greater degree of bias, in GPT-5, Gemini 3 Flash Preview, and Claude Sonnet 4.5.”

4. E4 Measuring political bias in Claude Anthropic · 13 November 2025

   “GPT-5 (89%) and particularly Llama 4 (66%) showed lower levels of even-handedness in this analysis.”

Within the window, E1 (April 2026) and E4 (November 2025) document bias-related findings for GPT-5, and E3 is dated at the window's edge. None of the evidence shows whether OpenAI had disclosed these biases beforehand, and the criteria use that to separate true from mostly false.