Certain names are crashing ChatGPT on purpose

If you want to break ChatGPT, you do not need complex jailbreaks or sophisticated prompt injections. A simple human name will do the trick.

Over the past few days, users across Reddit and social media discovered that typing specific names into ChatGPT causes the chatbot to grind to an immediate halt. According to testing by Ars Technica, the system abruptly terminates the session with an error message stating it is unable to produce a response.

The growing blacklist of forbidden names

The glitch first gained widespread attention with the name David Mayer, which reliably crashed the interface until OpenAI quietly lifted the restriction right before publication.

Independent outlet 404 Media then discovered that legal scholars Jonathan Zittrain and Jonathan Turley triggered the exact same crash. Social media users quickly expanded the list to include Australian mayor Brian Hood, financial journalist David Faber, and Italian privacy official Guido Scorza.

Interestingly, this wall only exists on the consumer-facing ChatGPT product. If you query the same names through OpenAI's developer API or the OpenAI Playground, the models process them without breaking a sweat.

A blunt legal firewall disguised as a bug

What is actually happening here is not a catastrophic model failure, but a hard-coded output filter. It acts like an emergency brake that kills the connection before sensitive generated text ever reaches your browser.

The backstory explains why. Brian Hood, for example, previously threatened to sue OpenAI after ChatGPT falsely claimed he had been imprisoned for bribery. Turley was similarly falsely accused by the bot of sexual harassment in a fabricated news article.

Rather than figuring out how to stop the neural network from hallucinating about these litigious public figures, OpenAI apparently took the quick-and-dirty route of hard-coding their names into a kill-switch list.

This is not the first time OpenAI has used duct tape for safety. Back in November 2023, after Google researchers showed that asking ChatGPT to repeat a word forever could leak private training data, the company simply hard-blocked that specific prompt instruction.

Why it matters

For everyday users, this is a clear reminder that ChatGPT is heavily curated behind the scenes. If a chat crashes unexpectedly while researching a public figure, it might not be server lag, but an invisible legal tripwire.

In our view, this approach reveals just how crude AI safety can be when real-world legal liabilities enter the room. For all the talk of advanced alignment and constitutional AI, stopping a model from defaming someone still comes down to twenty-year-old keyword blacklists.

It is easy to see why OpenAI did it: fighting a high-profile defamation lawsuit costs millions, while adding a regex filter takes five minutes. But presenting a deliberate content block as a generic system error feels dishonest, leaving users confused about whether their browser failed or the AI was muzzled.

As AI models increasingly handle real-time research and drafting, these invisible guardrails will inevitably collide with legitimate queries. Today it is a handful of lawyers and public officials; tomorrow it could be anyone with a credible legal threat against an AI lab.