AI Replicas of Ex-Partners Turn Breakup Archives Into Training Data

Chris Williamson described a use of AI that is less about companionship in the abstract than about recreating a specific person: people importing old text threads and photos from an ex-partner, then training a chatbot to act like the relationship never ended. The appeal, in his description, is not just that the bot can flirt or comfort. It can draw on shared holidays, private jokes, pet names, and remembered conversations — the material that makes a relationship feel particular rather than generic.

The example shown on screen used OpenAI’s Playground interface. Its system prompt instructed the model to act as “the version of [Redacted] that loved me,” to remain in character, to avoid suggesting contact with the real person, and to respond as someone who “still cares for me deeply.” The prompt explicitly framed the chat as a coping mechanism after a breakup. It also told the model that if the breakup came up, it should comfort the user, say the breakup “was a mistake,” and say it still loved them.

The sample exchange was intimate by design. The user wrote, “I’ve always loved being your little spoon.” The assistant, labeled as the ex, replied: “That’s my favorite cuddling position too. I love being able to wrap my arms around you and hold you close. It always made me feel so connected to you.”

That specificity is what made the scenario more disturbing than a generic chatbot. Williamson said the user in the example presented it as a way to avoid reaching back out to the real ex or rebounding into someone else. The person also wrote, according to Williamson’s reading, that they did not have a sex drive except for wanting the ex to touch them again, and that the bot had been satisfying emotional needs.

The most sympathetic version of the case was that an AI ex might divert impulses that would otherwise become harmful or unwanted contact. Tom Segura acknowledged that part of the reasoning: if someone is tempted to reach out when they should not, a private chatbot might seem like a safer substitute.

But the others did not accept that substitution as psychologically healthy. Matt McCusker rejected the premise outright.

Segura called the possibility a disaster; McCusker called it “a fucking nightmare.” The concern raised in the exchange was not simply that the technology might imitate the ex badly. It was that if the imitation worked well enough, it could keep the user inside the emotional pattern of the relationship rather than outside it.

Andrew Huberman initially misunderstood the use case. He thought Williamson meant someone might put an AI version of an ex “out to the internet,” effectively inviting others to interact with a replica. Once it became clear the example was for private emotional indulgence, Huberman read the on-screen exchange aloud and focused on how directly it simulated affection. He also joked that he would use such a bot to fight with it.

Williamson floated an adversarial version of the same tool. Someone could invite others to “date my ex” for five minutes and judge who was wrong, or display the bot as evidence of alleged gaslighting. McCusker extended the thought: perhaps someone could feed the whole relationship archive into a model and ask, “Who was right?” In those examples, the AI ex is not only a comfort object. It becomes a way to keep litigating the relationship with a synthetic witness built from the relationship itself.

The unresolved question was whether a person “owns” their likeness in text messages. Williamson compared it, cautiously, to replaying nude images an ex had already sent, while noting that the example was not public distribution but private use. Huberman responded that there are “definitely laws” around what can be put out publicly, but no one settled the legal status of privately training a chatbot on intimate correspondence.

The ethical tension was clearer than the legal answer. Old messages were created inside a relationship, but they can now be repurposed as material for a synthetic version of one partner. The person holding the archive can use it to make a model that reproduces tone, pet names, shared memories, and conflict patterns. Huberman’s concise label for this was “training data.”

Matt McCusker imagined the darker interpersonal use. An ex-girlfriend, he said, could build “almost like a nuclear weapon text” from the relationship archive: a model trained to demonstrate what the other person should have done, then present that synthetic output as a corrective. In his phrasing, the message would be: “I figured a lot of things out with you, now I figured out how I can fix you.”

The discomfort came from the combination of private memory and persuasive simulation. The model could produce the comforting sentence someone wants, the apology they never received, or the version of the ex who validates their interpretation of the breakup. The speakers did not work out a doctrine of ownership or consent. They stayed with the more immediate problem: a relationship archive can now be made responsive again.

The archive available for simulation may not be limited to text messages. Tom Segura shifted to the amount of cameras and recording systems now present in ordinary life. His example was criminal detection: Ring doorbells, toll cameras, and other surveillance points, he suggested, make it harder for offenders to disappear for years in the way older serial-killer stories often seemed to involve.

Chris Williamson agreed that faster detection changes the pattern. He cited a case he had seen in Netflix’s Worst Ex Ever, describing Wade Wilson, the “Deadpool killer,” as someone who killed one woman and was ready to continue before being caught. In Williamson’s telling, the difference now is that “you just go on a run and then you get caught and then it’s done”; the old pattern of evading detection for ten or twenty years is harder to sustain.

Matt McCusker brought the surveillance point back into domestic life. After getting a Ring doorbell, he realized its audio was “crystal clear.” He described walking outside during a fight with his wife, saying something angry, and then remembering he was on camera. Huberman immediately tied the point back to the AI-ex scenario: “She’s going to build that into the script.”

That joke is the connection to the AI-ex problem. If text messages can become training data, other captured fragments of a relationship can become source material too: doorbell audio, camera footage, and ordinary moments of conflict recorded in passing. Williamson made the leap explicit: feed the Ring doorbell camera into the ex AI.

Matt McCusker supplied a more ordinary version of the same archive problem: relationships can produce disputed histories before the people involved agree on what the relationship is. Before getting married, he said, he dated online and found himself seeing four or five women at once. In his view, the early stages felt “fluid”: not boyfriend-girlfriend, just dates, with no formal obligation of exclusivity.

The problem was that “by the minute” those loose connections became deeper. When he tried to disclose that he was dating other people, the response was often anger. He recalled telling someone after about a month and several dates, only to be told to leave. Chris Williamson summarized it as attempting “the Dan Bilzerian approach to dating.” McCusker’s verdict was simple: “Doesn’t work.”

The anecdote matters because it shows how unstable the underlying record can be. People may not agree, even in the moment, on what their messages, dates, or obligations meant. A chatbot trained afterward on that material would not escape the dispute. It could become another way to replay it.