Ralph Adolphs says AI can diagnose your personality from one page of text

Your face may be lying, but your paragraph is apparently a snitch. Ralph Adolphs says recent work suggests large language models like GPT or Claude can infer a person’s personality from about a page of diary-like text, roughly as well as someone who knows them.

That claim lands with extra force because Adolphs, a Caltech neuroscientist on Huberman Lab, is not a Silicon Valley mystic selling you a chatbot best friend. He spends much of the discussion taking a brickbat to the idea that emotions are neatly readable from faces, then turns around and says text gives AI a surprisingly good personality read. The machine, in this telling, doesn’t need your selfie. It wants your notes app.

people generated I think it’s about a page or so of sort of diary type written text and asked just large language models like GPT or Claude to analyze that text without knowing anything else and just tell me the the personality profile and it’s as good as the ground truth or as another person or a clinician would be

The word “diagnose” is the spicy part, and Adolphs uses a version of it himself. Still, don’t hear this as “Claude is your psychiatrist now.” The example he gives is Big Five personality, traits like extraversion, conscientiousness, and neuroticism, not a DSM label delivered by a toaster with venture funding. The believable version is narrower and, honestly, more useful: your writing contains stable behavioral residue. Cadence, topic choice, emotional temperature, the little ways you arrange reality on the page. Not magic. Pattern.

you can tell that AI can tell that just from a transcript of text quite well. So the information is there clearly in what we say, what we write and the order in which that progresses that tells a lot about a person enough that AI can basically diagnose your personality from just text.

The face loses to the group chat

The funny part is that Adolphs is much harsher on the thing humans think we’re great at: looking at someone and just knowing. He revisits the Paul Ekman tradition of basic emotional faces, the famous anger, fear, disgust, sadness lineup that launched a thousand corporate emotion-detection decks. His complaint is simple and devastating. Those faces are posed. They are extreme. The test is often multiple choice. Congratulations, you matched the emoji to the label.

our conviction vastly outstrips our accuracy.

This is where the AI claim stops sounding like sci-fi and starts sounding like a nasty little truth about daily life. A static face is one frozen frame. Text is behavior over time. Huberman brings up the way people read romance through changed texting frequency, terse morning replies, weirdly formal emails, the terrifying difference between “Thanks!” and “Regards.” Adolphs connects that to prediction: we notice when the pattern breaks. The model notices too, except it doesn’t need to pretend it isn’t analyzing your punctuation.

That doesn’t make artificial intelligence emotionally intelligent in the human sense. It does make it very good at exploiting a thing humans already use all day: deviations from baseline. Your partner doesn’t need a perfect theory of the amygdala to know something is off when your texts suddenly sound like they were written by a regional bank. The model just scales that instinct, without the hurt feelings.

Adolphs is not saying the robot has a heart

Adolphs’ bigger theory of emotion helps explain why he can entertain odd-sounding AI claims without sounding like he’s been trapped in a Stanford dorm room. He defines emotions less by how they feel and more by what they do: they prioritize behavior, scale in intensity, persist over time, and help an organism respond flexibly to the world. That framework lets him talk about animals, humans, and even machines without making consciousness the price of admission.

large language models have functional emotions as well.

This is the most debatable claim in the room. A chatbot can prioritize, persist, and produce emotion-like outputs, sure. It does not have a kidney stone, a marriage, a haunted-house panic response, or the private humiliation of sending a text with the wrong tone. Adolphs’ point is not that AI feels sad while drafting your meeting summary. It’s that if you define emotion as a functional control system, the border gets weirder fast.

The practical stake is less philosophical and more annoying. If Adolphs is right, the most revealing digital version of you may not be the photo you chose, the bio you polished, or the face you make on Zoom. It may be the page you dashed off when you thought nobody important was reading.