It’s happening right now.
While you argue with the chatbot about your cat’s diet.
AI knows you better than your partner probably does. Or maybe that’s too much. But a new pre-print study from ETH Zurich says artificial intelligence can guess your personality just by reading your chat history. Not just vague guesses either. Specific, psychological traits.
Here’s the setup. Researchers grabbed the data.
They asked 668 users in the US and UK to dump their ChatGPT logs into the lab.
That resulted in over 62,00 chats. A lot of text. They sorted the chaos by topic and trained a model to look for patterns.
What patterns?
The Big Five. Psychologists have been using this framework for decades:
– Agreeableness
– Conscientiousness
– Emotional stability
– Extraversion
– Openness
To check if the AI was lying or actually getting it right, the humans took a standard psych test. Ground truth.
The results were startling.
The model hit 61% accuracy.
That might sound low in some fields but for guessing human complexity? It’s impressive.
The AI was sharp on agreeableness and emotional stability.
It struggled with conscientiousness. Makes sense. Organized people might write cleaner prompts? Hard to say. The algorithm missed the mark there.
“The more you use the AI, the more you reveal.”
Longer chat histories meant better predictions.
Every prompt you type is a fingerprint. Every “yes” and “no” paints a picture of you that you didn’t realize was being sold—metaphorically, or maybe not.
So what’s the risk?
For you, personally? Not much. Yet.
The real danger is scale.
Bad actors could buy this kind of analysis.
Imagine targeting millions of people with propaganda tailored specifically to their emotional triggers. Not broad strokes. Laser-focused manipulation.
Disinformation campaigns that know exactly which buttons to press because an algorithm analyzed your late-night venting sessions.
Researchers want to build shields. Tools that scrub your identity from your chats automatically. Privacy filters that turn your distinct voice into generic noise before the model sees it.
We need these tools. We need them fast.
But until they exist?
You’re leaving trails.
Everywhere.
Are we just data points in waiting? 🕵️♂️
Probably. And we’re getting better at hiding… or so they tell us.
