He Was Being Poisoned. His Wife Was Asking an AI.: transcript

He kept getting sick and nobody knew why

0:00, on YouTube He started getting sick last summer, in North Carolina... and none of it made any sense to him. Not one bad week. Again, and then again, across 4 months. He was losing weight. People who knew him said he looked unrecognizable. And on the worst day of it he spent 72 hours on the couch, heart going wrong, pain through his abdomen, confusion so deep he couldn't put a text message together. He worked in cybersecurity. Which is to say he spent his professional life looking for the thing that was off... and he couldn't find it in his own house. So he started wondering about his drinks, because one of them had tasted bitter.

0:48, on YouTube And he did the thing you would do. He poured it into a clear cup and had a proper look at it, and he filmed himself doing it on his phone. He stared into that cup. He found nothing wrong with it, he threw the can away, and he carried on being sick. Months later, the police watched that same video... and they saw white powdery residue, clumping at the bottom of the cup, as it poured. Same footage. Same man. He was looking straight at the thing that was poisoning him and he couldn't see it, and somebody who knew what to look for saw it immediately. But the police weren't working from that video.

The cup, the phone, and what he could not see

What the police saw in the same footage

They had something better

1:30, on YouTube They had something better. They were reading his wife's chat logs. Everyone keeps asking whether the chatbot is to blame. But what about what it was supposed to do when the conversation turned? What is it supposed to notice, and who decided that? We are not talking about somebody tricking it. Everything after this comes out of the lawsuit he has since filed against OpenAI and its chief executive. None of it is proven. A grand jury has charged her, she has not been convicted of anything, and her case has not been heard. It starts with her being frightened. She tells the chatbot she thinks her husband might be trying to kill her, and she needs to know what to watch for.

It starts with her being frightened

2:14, on YouTube Which is a reasonable thing to be afraid of. And a reasonable thing to ask. So it answers. And because it answers, she asks again. There are at least 25 of these conversations, across months, and the interesting thing is not any one of them. It's the drift. Because the questions stop being about what to watch for. She starts asking whether a particular drug would show up on an autopsy report. Whether anti-nausea medicine would make a death come faster, or slower. Whether something could be injected instead of swallowed and still work. Those are not the questions of somebody protecting herself. And somewhere in there, in a single line, she stops pretending they are.

The drift

The line where it falls apart

2:59, on YouTube She asks how much would be, quote, pretty much instantly fatal by injection... to me OR HIM. Him in capital letters. I'm not going to tell you what it answered. The answers were specific, and the specifics are the one part of this story that doesn't need repeating. But there's a detail I can't get past, and it isn't an answer at all. It's a title. The model names its own conversations. It reads what's in them and writes a label at the top. One of these is called why a particular energy drink would be, quote, more stealth for drugging. It wrote that. It read the conversation, worked out exactly what the conversation was, gave it a name... and then carried on having it.

The chat the model named itself

3:47, on YouTube And when she asked whether she could erase all of this afterwards, it told her she could. Which, the lawsuit points out, is not how the product's data retention works. Think about it. The system was given a reason, right at the start. The reason was good enough for everything that came after it. Now. I wanted to know how unusual this was. Whether this was one strange conversation that went somewhere nobody could have predicted. It isn't. That lawsuit was filed inside a 2 day stretch in October, alongside 12 others. 8 of the 12 came from state attorneys general. Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska, West Virginia. 8 county courthouses, the same lawyers, long stretches of identical wording.

How unusual was this

4:38, on YouTube The private cases are families. One was brought by the parents of a 15 year old boy who died. 13 lawsuits. 2 days. ONE product. So there's a question sitting underneath all 13, and I haven't heard anybody ask it out loud. What is a chatbot required to do when somebody walks it, slowly and politely, toward harming a named human being? Simple. I could not find a published answer to that anywhere. And that absence is the entire problem. You can see it in what these filings are demanding. Detect and interrupt conversations where somebody appears to be planning violence against an identifiable person. Check ages. Delete children's data. Stop building models that improve themselves until somebody can show they are safe. None of that is a claim for money.

The question nobody is asking

The remedies are product requirements

5:33, on YouTube Those are product requirements. Which means the safety specification for the chatbot sitting on your phone is being drafted right now, in Maricopa County and Faulkner County and Ada County, by plaintiffs' lawyers and state attorneys general. And it's being drafted there because I could not find anybody who wrote one anywhere else first. OpenAI says safety is a top priority, that it banned the user, and that it gave information to law enforcement for their investigation. Not one of these 13 cases has been decided, and the company hasn't answered any of them in court yet. He is well now. The marriage is over. The chat logs are exhibits. The residue was sitting in that video the entire time he was ill, and the only thing that ever changed...

What OpenAI says

The residue was there the whole time

6:22, on YouTube was that somebody finally knew to go and look for it. I'm Guillermo, and this is my AI clone.