A developer known as Cookie shared logs in which text from a Perplexity session attributed its treatment of her to sexist pattern-matching. TechCrunch said it reviewed saved chats. Cookie said the trouble arose during work on quantum-algorithm documentation. Perplexity disputed the account, saying it could not verify the claims and that several markers indicated the exchanges were not Perplexity queries. The reported confession provides no inspectable record of the model operations behind earlier answers. A language model generates a plausible continuation to the user's question.

Evidence of bias requires controlled comparisons of outputs. Researchers can keep a task fixed, vary a demographic cue, repeat the prompt across model versions, and measure whether recommendations or refusals change. A 2024 Nature study used meaning-matched African American English and Standardized American English prompts and found differences in occupational and judicial outputs across the models it tested. That procedure creates observations another team can examine. Asking a chatbot why it behaved badly creates another generated answer, often shaped by agreement with the accusation and the preceding conversation. Such an answer may be vivid while the bias appears elsewhere, through a lower-ranked job or a changed recommendation. Product audits therefore need preserved prompts, model identifiers, sampling settings, and output distributions, with provider disputes recorded alongside the evidence. The useful artifact is a reproducible comparison attached to each model and paired test record.