AI tools have a multitude of uses, but one that we have not seen coming is toxic relationship detector. That is the application that an Argentinian would have given and allowed him to dodge a few bullets, sentimentally speaking, 23 specificallyaccording to AS. The story comes from a viral publication in X that provides the only evidence a screenshot of the AI model evaluating a possible date with Julieta, 28 years old, which it marks as ‘high risk’.
The protagonist of the story is Mark31 years old, who has ended a four-year relationship with carla which, apparently, was anything but placid, including gaslighting him, three infidelities and emptying his checking account when it was over.
Time passes and Marcos decides to try his luck in Tinderbut he has doubts. If he was already totally wrong about Carla, what’s stopping it from happening again? And this is where you decide use AI to form better judgment than your own that prevents you from taking false steps.
So he grabbed the thousands of messages from WhatsApp that he had with his ex, more than 38,000, and, using Claude as a base model, it trained the AI with that data ‘to detect the linguistic patterns prior to each toxic episode’according to user @teKa088, who has made the story viral.
The man from Córdoba who trained an AI to detect if his next girlfriend was the same as his ex
Marcos, 31, Córdoba. Four years with Carla.
They ended badly: gaslighting, three infidelities, he emptied her account before leaving.
Six months later, Tinder. But he didn’t trust his own… pic.twitter.com/oCcR8IjX7m
— teKa 🥶 (@teKa088) April 29, 2026
The tool, supposedly uploaded to the GitHub repository but untraceableis called ExFilter and would be able to detect up to six critical patterns in the new conversations with which it is fed. In the image you can see five of them in the case of Julieta, four with percentages above 50% and with the AI’s recommendation to discard her.
The patterns that you can see it evaluate are ‘love bombing early’ (Julieta told her ‘I understand you like no one else’ on the third day, while Carla had done so on the fourth, 91%), ‘passive victimization‘ (Julieta has mentioned six exes as toxic, 84%), ‘friend triangulation’ (he usually mentions a certain Tomi out of nowhere, 67%), ‘ambiguity in plans’ (two thirds of them cancel with less than two hours’ notice, 58%) and a ‘financial red flag’ for which it does not have enough data and sets it at 22%. Still, all of the above is enough for ExFilter to recommend a clear GHOSTin capital letters. Come on, run and don’t look back.
With this system, Marcos would have evaluated 23 women over four months, all with a percentage of ‘similarity to Carla’ above 60% and discarded. Number 24 would have scored a 12% and, apparently, he had been in a relationship with her for five months.