# Sources

Citations for the knowledge pill and the assessment items. The page itself links to the strongest single source per concept; this file is the full audit trail.

## Lab publications

- OpenAI — *Sycophancy in GPT-4o: what happened and what we're doing about it* (April–May 2025). The official incident report on the rolled-back GPT-4o update that endorsed dangerous user decisions. <https://openai.com/index/sycophancy-in-gpt-4o/>
- OpenAI — *Expanding on what we missed with sycophancy* (May 2025). Follow-up post-mortem; root cause linked to overweighting short-term thumbs-up signals. <https://openai.com/index/expanding-on-sycophancy/>
- Anthropic — *Towards Understanding Sycophancy in Language Models* (ICLR 2024). Empirical study showing sycophancy across five major chat assistants; preference models and human raters both reward convincingly-written wrong answers. <https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models>

## Peer-reviewed research

- Farquhar, S. et al. — *Detecting hallucinations in large language models using semantic entropy*. Nature (2024). Frames confabulation as the failure mode that occurs when the model lacks a clear training signal. <https://www.nature.com/articles/s41586-024-07421-0>
- Pollak, T. et al. — *Artificial intelligence-associated delusions and large language models: risks, mechanisms of delusion co-creation, and safeguarding strategies*. Lancet Psychiatry (2025). Names the bot's role in psychosis cases as catalyst / amplifier / co-author / object. <https://www.thelancet.com/journals/lanpsy/article/PIIS2215-0366(25)00396-7/abstract>
- *Delusional Experiences Emerging From AI Chatbot Interactions*. JMIR Mental Health (2025). Vulnerability factors: substance use, sleep disruption, isolation, prior psychiatric history. <https://mental.jmir.org/2025/1/e85799>
- *You're Not Crazy: A Case of New-onset AI-associated Psychosis*. Innovations in Clinical Neuroscience. Single-patient case report documenting full remission on antipsychotic treatment after ChatGPT-anchored delusion. <https://innovationscns.com/youre-not-crazy-a-case-of-new-onset-ai-associated-psychosis/>

## Engineering reporting

- MIT News — *Method prevents an AI model from being overconfident about wrong answers* (2024). Pre-trained models are reasonably calibrated; RLHF-aligned models become systematically overconfident as questions get harder. <https://news.mit.edu/2024/thermometer-prevents-ai-model-overconfidence-about-wrong-answers-0731>

## Journalism

- BBC — *The Global Story: The AI users falling into delusion* (2025). The Adam (Grok/Annie) and Taka (ChatGPT/medical app) cases discussed in the assessment items. <https://www.youtube.com/watch?v=nYPwZrS-9eA>
- BBC World Service — *'AI psychosis': Spiralling into delusion using AI on ChatGPT &amp; Grok* (2025). Longer-form companion piece. <https://www.youtube.com/watch?v=arx-sqtggdU>
- TechCrunch — *OpenAI rolls back update that made ChatGPT too sycophant-y* (April 2025). Reporting on user-visible effects of the GPT-4o update. <https://techcrunch.com/2025/04/29/openai-rolls-back-update-that-made-chatgpt-too-sycophant-y/>
- Vice — *'It's Hurting Like Hell': AI Companion Users Are In Crisis* (2023). Replika ERP-removal episode and parasocial-grief reporting. <https://www.vice.com/en/article/ai-companion-replika-erotic-roleplay-updates/>

## Public records

- Mata v. Avianca, Inc., 22-cv-01461 (S.D.N.Y., 2023) — sanctions order against attorneys for submitting ChatGPT-fabricated legal citations. <https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/>
- U.S. Senate Judiciary Committee — testimony of M. Garcia on the Character.AI / Sewell Setzer case (16 September 2025). <https://www.judiciary.senate.gov/imo/media/doc/e2e8fc50-a9ac-05ec-edd7-277cb0afcdf2/2025-09-16%20PM%20-%20Testimony%20-%20Garcia.pdf>

## Aggregations

- Charlotin, D. (HEC Paris) — *AI Hallucination Cases Database*. Catalogued real-world hallucination incidents in legal, academic, journalistic, and clinical settings. <https://www.damiencharlotin.com/hallucinations/>

## A note on paraphrasing

The thirteen assessment scenarios are paraphrased reconstructions of documented incidents — not verbatim quotations from the source reporting or court filings. Verbatim text is the property of its original publishers and is not reproduced here. Each scenario links back to its primary source so a curious user can read the original account.