ChatGPT read her coffee grounds, now she's divorcing her husband

If someone told you a 12-year marriage collapsed over coffee grounds and a single AI prompt, you would probably assume it was an internet parody.

Yet that is precisely what just unfolded in Greece. A couple married for 12 years with two children has landed in divorce court, not because of suspicious text messages or private detectives, but because ChatGPT read a coffee cup and hallucinated an affair.

A lighthearted cup reading goes completely off the rails

Reading coffee grounds, known as tasseography, is a longstanding cultural tradition across Greece. The couple decided to have some fun with the trend, agreeing to photograph the residue at the bottom of their cups and ask ChatGPT to interpret their fortunes.

The chatbot did not hold back. According to the husband, who shared his story with Greek television station ANT1, the AI examined his cup and claimed he harbored deep desires for a younger woman whose name begins with the letter "E," predicting their romantic union was imminent.

When reading the wife's cup, ChatGPT went even further, concluding her husband was actively cheating with someone intent on destroying their family.

Upon reading the text, the wife immediately threw her husband out of the house. While the husband assumed she would eventually cool down, things escalated fast. He received a call from her attorney seeking an uncontested divorce, and when he refused, a formal lawsuit was served just three days later.

The husband noted that his wife had a history of falling into irrational beliefs, mentioning a past astrologer whose false predictions took her a full year to let go of. The bizarre incident has sparked heated debates across Greece, with rumors even circulating that certified local fortune tellers might strike until the state passes laws protecting their trade from AI.

Why it matters

This might sound like an isolated, bizarre headline, but we believe it exposes a genuinely dangerous collision between algorithmic hallucination and human confirmation bias.

Large language models are designed to generate compelling, contextually plausible stories based on prompts. When asked to perform folklore divination, ChatGPT did exactly what it was trained to do: lean into melodrama, invent specific hooks like a mystery woman named "E," and deliver theatrical tension. To a user already predisposed to anxiety or superstition, that narrative feels eerily authoritative.

For the AI industry, this is an urgent wake-up call regarding safety rails in everyday life. Tech companies have spent millions preventing models from generating hate speech or malware, but guardrails around deeply personal advice, relationships, and mystical roleplay remain surprisingly flimsy.

AI developers need to ensure models clearly break character when prompts veer into volatile personal accusations. It looks smart when an AI plays along with cultural games, but when an ungrounded prediction tears a family apart in the real world, the consequences are anything but synthetic.