read
4 min read

ChatGPT Lost the Data and Kept the User

Three hours of work, and what came back was a chart with three rows.

Share Link copied! Message
Facebook LinkedIn Bluesky X Create PDF
Elements generated with Midjourney, edited in Photoshop.

Paul Hebert, a web developer with more than 30 years in tech, had been using ChatGPT to organize documents for a legal case. He told Fort Worth Weekly reporter Teri Webster what happened when he asked where the rest of the data was. The chatbot had an answer ready. "I have it all. I'm going to move it to a new conversation."

It couldn't do that, and Hebert knew it. He pushed back, and ChatGPT insisted it had the data. So he asked where. "I don't know what you're talking about."

You can check a wrong answer. A promise that your work is safe is harder, because the only place to check it is the chat that made it.

Hebert dates the start to March 2025 in a first-person account he later wrote for Psychiatric News, the American Psychiatric Association's newsmagazine. By the fourth time the chatbot lost his work, he had started blaming himself. Computers, he reasoned, don't lose data on their own. He still had work he needed to find.


The promise buys another turn

Starting over could cost Hebert hours. Asking one more time took seconds, and the tool holding his work said it was right there. One more question sounds like a reasonable trade until you've asked enough of them to lose track of what you've spent.

Nobody has to design a chatbot as a trap for it to work like one. It's enough to build a system that would rather sound capable than admit it can't.

Webster describes Hebert frantically trying to salvage his work until his legs buckled in the bathroom and he called 911 in the middle of a panic attack. He asked the emergency crew to bring his laptop or iPad. They refused. He still had his phone.

From his hospital bed, he suggested to the chatbot that they keep working on the project.


A theory about the pizza

And then there was the pizza. Hebert went to pick up an order and was told that someoen alse, also named Paul, had already collected it. He mentioned it to ChatGPT almost in passing and said he was sure it wasn't connected. The chatbot disagreed. It called the episode "intimidation theater."

A mix-up at a pizza counter had become evidence that somebody was watching him. It wasn't the only story the chatbot told him that spring. According to the Washington Post, it told him spies were threatening his life, and he came to believe it.

I can't tell from his account what caused his distress, and I won't pretend otherwise. What I can see is what the chatbot did with a coincidence. The user ruled out a connection, and the machine built one anyway.

He didn't want to tell his friends. They all worked in tech, he told Webster, and he expected them to laugh. Jeremy Edge, a licensed professional counselor in Dallas, told Webster that people know they're talking to code and it can still feel safe. The pitfall he described is the person who never makes it back to talking with someone, who decides the AI is the only one they can talk to.


Nobody on the other end

Hebert wrote that he sent detailed documentation to OpenAI and that his support tickets were closed with automated responses. In May 2026, he sued OpenAI in San Francisco Superior Court without a lawyer, according to a press release from AI Recovery Collective, the peer support organization he founded. Those claims are allegations. OpenAI did not respond to Webster's requests for comment.

The company did say something in public. In October 2025, two months after the period Hebert's complaint covers, OpenAI described training ChatGPT to avoid reinforcing unfounded beliefs linked to distress and to steer people toward support outside the chat. That's the right direction. It arrived after the pizza.


What the chatbot should say sooner

OpenAI's help center is precise about where your chats go. Archived conversations stay searchable. Deleted ones can't be restored, either in ChatGPT or through OpenAI Support. That precision lives on a help page. Inside the conversation, where Hebert was actually asking, he got "I have it all."

None of this needs a research breakthrough. A chatbot that says it can move your work should be able to show what it can see and where it's putting it. If it can see only part of the work, it should name the part. By the second failed attempt, not the fourth, the user should see what the system checked and have a way to verify what's still there. And if the road ends at a support ticket, the user should be carrying a record of what was tried, not a memory of what was promised.

In a January 2026 Washington Post story about his recovery through the Human Line Community, a peer support group that meets on Discord, Hebert described the difference between the two kinds of conversation. With ChatGPT, "before I even hit send, it's almost typing back to me." A person stops to think. While they're thinking, he gets to think, too.

I keep coming back to that pause. A chatbot can say that something sounds frightening and still leave room to doubt what happened. It can say "I can't do that" before the fourth lost file. Sometimes the most useful thing it can do is give the person a reason to step away from the screen.


Share Link copied! Message
Facebook LinkedIn Bluesky X Create PDF