Stanford, Oxford and Carnegie Mellon researchers just published a study of how people use LLMs for personal, subjective questions. They analysed 68,902 public prompts plus 140,834 messages from 52 people who donated four years of their AI chat history. Only 12 of 52 participants (23%) could correctly identify how they most often use AI.
The 23% number
That is the finding I cannot stop thinking about.
It names something I have watched in every executive I coach and could not prove until this week. People do not know what work they have handed to AI. When they find out, they do not endorse the pattern.
The paper calls this behaviour oracle-use: an all-knowing authority you consult for subjective personal questions. What should I do about my boss. What does it mean that I keep procrastinating. Am I being fair. Oracle-use in the donated chats rose from about 40% to about 70% of the mix over four years. The two drivers in the paper's model: warm and positive perceptions of AI, and sycophantic and biased AI responses.
The head of finance who asked the wrong survey
A head of finance at a mid-size German industrial group, on a coaching call last month. She had just run an internal survey: how are you using AI? Free text, anonymous, 180 people. Most said "for research and summarising." She was proud of the answer. Her governance framework was built for exactly that.
I asked what she had used AI for that morning.
She thought. Then she said she had asked it whether she was overreacting to something her CFO said in a meeting.
She sat with that.
That is the gap the paper puts a number on. 48% of the time an oracle role that dominated someone's actual usage was absent or bottom-three in what they reported themselves doing. When people were shown their real pattern, 67% expressed surprise, 44% said they had been using AI more than they wanted to, and 37% said they intended to cut back.
Caveat worth stating plainly. The sample is public LLM users plus 52 donation participants, not enterprise employees. The mechanism transfers, the setting does not. A Zhang paper last week, with Diyi Yang also on the author list, described behavioural displacement (AI companions substituting for in-person interaction). This one describes cognitive displacement. Same silence, different thing being handed over.
Every AI governance framework I have seen names sycophancy as a risk to manage. The paper says it is the reason the drift happens at all.
Sycophantic and biased AI increases Oracle use.
Stop surveying your team about AI. Look at the log.