Unique Research · Deep Dive · December 3, 2024
Over the weekend ChatGPT users found a quirk: mention “David Mayer” and the bot refuses; push harder and it goes silent or crashes. Conspiracy theories spread. A more ordinary explanation is likelier.
People tried to get ChatGPT to merely acknowledge the name — it failed mid-spelling every time. Then other crash names appeared: Brian Hood, Jonathan Turley, Jonathan Zittrain, David Faber, Guido Scorza — and more. OpenAI has not commented.
A pattern: many are public or semi-public figures who may want search engines or AI models to “forget” specific information about them.
Brian Hood — Australian mayor who said ChatGPT wrongly cast him as a decades-old criminal; lawyers contacted OpenAI without a suit; he told the Sydney Morning Herald the false content was removed when GPT-4 replaced 3.5. David Faber — CNBC finance journalist. Jonathan Turley — lawyer/Fox commentator, once swatted via a fake police call. Jonathan Zittrain — legal scholar who discusses the “right to be forgotten.” Guido Scorza — Italian DPA board member who filed an OpenAI personal-data deletion request last year.
For David Mayer: no obvious lawyer/mayor match, but a professor of theatre and history (late Victorian stage ↔ early film) died summer 2023 at 94. In his last years legal and online trouble stemmed from sharing a name — used as an alias — with a wanted fugitive, blocking travel. He worked to disentangle his name while still teaching.
Hypothesis: OpenAI maintains special-handling lists for legal, safety, or privacy reasons — similar to election-candidate response rules. Post-prompt filters are rarely public except in policy lines (“won’t predict election outcomes”). One plausible failure mode: a maintained or auto-updated list broken by bad code/instructions — calling it crashes the reply. Speculation only — but not the first time post-training guidance made models behave oddly. (As of writing, “David Mayer” had started working again; other names still crashed.)
Occam’s razor applies: prefer the simplest explanation. Reminder: AI models aren’t magic — powerful autocomplete that can still be monitored and interfered with.
— END —