Language model trained on a decade of chat history challenges close acquaintances
Models & ResearchTAAFT - There's An AI For That · 12h ago

Language model trained on a decade of chat history challenges close acquaintances

A researcher at the University of Illinois constructed a digital mimic using hundreds of thousands of his personal messages, autobiographical writing, and personality evaluations to see if his peers could identify the duplicate. In blind text tests, friends correctly identified the real person only slightly more than half the time, while the program consistently beat humans trying to impersonate him. Participants were usually alerted to the machine's identity by clumsy factual inserts, overstated mannerisms, or missed shared memories.

University of Illinois

The Blend

A research team led by Muhammad Rizky Wellyanto created an artificial intelligence system trained on ten years of one author's private chat records, personal writing, and personality profiles. The project aimed to discover whether software could successfully impersonate a specific individual in text conversations with close friends. In blind trials, acquaintances correctly identified the genuine person only thirteen times across twenty rounds, a success rate barely better than guessing.

The automated clone also outperformed real people who were asked to pretend to be the researcher, fooling interrogators in seven out of ten rounds. When friends did spot the machine, it was usually because the model inserted biographical facts unnaturally, overplayed familiar catchphrases, or failed to recall specific shared memories.

For regular internet users, the findings show that automated social engineering could become far more sophisticated. If digital message histories are stolen or voluntarily harvested, bad actors could construct highly convincing conversational replicas to defraud family members, bypass trust barriers, or manipulate personal relationships.

Because the study focused on just one individual and involved a small group of testers, it remains uncertain whether these clones would stay convincing over days rather than brief text chats. This also opens a complicated legal dilemma that existing privacy laws do not address: if building an authentic personal clone requires scanning decade-long message archives, it inevitably exploits the private words and confidences of every friend and relative who ever replied.

Written independently by AI News Smoothie from the reporting listed below. Facts belong to the original publishers. Follow the links for their full coverage.

Ingredients

Read the original