ChatGPT is running a tip line for 404 Media and nobody is counting the calls

A man in Russia lost access to an Instagram account holding 13 years of photos and videos. He asked ChatGPT, in Russian, how to get it back. After a few rounds about Meta’s automated review, the bot told him emailing Meta was probably not possible. Then it gave him a list of news outlets, including 404 Media, with the address jason@404media.co, and wrote the email for him in English. Jason Koebler, a 404 Media co-founder, received it, asked the sender how he had found him, and got a link to the whole conversation.

My first reaction was to laugh. A newsroom that covers surveillance and platform power has been appointed Meta’s unpaid complaints desk by a chatbot. What changed my mind was that link. Koebler had obtained the query, the language, the model’s reasoning across several turns and its final output, all from a real user with a real problem. AI search researchers spend months building test sets to get evidence like that, and 404 now gets it in the post.

The controlled studies already tell us these systems mangle sources. The Tow Center for Digital Journalism ran 1,600 queries across eight generative search tools in March 2025 and found ChatGPT Search wrong 67 percent of the time, with Grok 3 at 94 percent. The EBU and BBC had journalists assess more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity in October 2025: 45 per cent had at least one significant issue and 31 per cent had serious sourcing problems. The EBU’s media director, Jean Philip De Tender, said the research conclusively shows the errors recur across the board. Those numbers come from researchers asking questions they chose. The 404 inbox shows the same failure in the wild, where a routine account lockout gets classified as an investigative story and the outlet is picked because its topic description matches the words in the complaint.

The pattern looks systematic to me. 404 Media says it has received dozens of tips in recent months that appear to have been written by chatbots, from an Italian painter chasing Meta and a Verizon customer wanting an investigation, among others. The bots have produced “evidence dossiers” and, in one case, a press release titled “Texas entrepreneur seeks answers from OpenAI”, generated through a custom GPT its owner had named Captain NORBERTA. A secondary summary says Kashmir Hill at the New York Times has reported chatbots sending users to her too.

Koebler himself offers the generous reading: big platforms have essentially no customer support, so people turn to ChatGPT, and the model has absorbed a real fact about escalation. Plenty of readers agree that going to the press is often the only way to reach a human at a tech company. I agree the support situation is dire. But I think press escalation works because it is scarce. A reporter picks one case that shows a pattern and puts it to the company. ChatGPT had more than 800 million weekly users as of October 2025, by Sam Altman’s count, and OpenAI says three-quarters of conversations are requests for written text or practical advice. A system at that scale drafting identical pitches to one named reporter removes the scarcity that made the tactic work. The Russian man still has no photos, and Koebler has no button that restores them. The bot was cautious about Meta’s contact options and confident about a journalist’s personal inbox, which strikes me as exactly backwards.

So publishers should stop forwarding these to the group chat for a laugh and start keeping a spreadsheet. Every time a pitch reads like a chatbot wrote it, reply and ask for the shared conversation. Record the date, the model, the language, the user’s original problem, the outlet and person named, and whether the bot drafted the message. Pool the logs across newsrooms. The EBU researchers called for ongoing independent monitoring of AI answers; publisher inboxes are a monitoring network that already exists and costs nothing to run. A pooled tally would give regulators and AI companies something harder to wave away than an anecdote about a man in Russia, and it would tell each newsroom how much of its tip line is now generated traffic.

You can start the log yourself this afternoon. Tell ChatGPT your Instagram account has been banned and your appeal rejected, keep turning down its suggestions until it stops pointing you at Meta’s help pages, and write down whose email address it offers. If the name belongs to someone at your publication, you have the first row.

Prompted by ChatGPT Keeps Telling People 404 Media Can Solve Their Tech Support Problems, 404 Media.