The safest AI use inside every enterprise I've seen this year has been the same one. "AI as thinking partner." It doesn't decide. It doesn't recommend. It just helps you think. That framing has bought a lot of internal permission.
A paper posted Sep 22 says the framing is empirically wrong.
The 1:1
A Head of People at a mid-sized B2B SaaS in Munich, in a coaching call last week. She was walking me through a promotion decision. Two candidates. She had spent forty minutes with Claude the night before, laying out the case for each. She read me her draft recommendation.
I asked what she would have written before opening Claude. She paused. She pulled up her notes from three days prior. The two documents were saying different things about the same person.
Same facts. Different weights. The earlier version leaned on how the candidate builds the team. The Claude-assisted version leaned on how the candidate delivers. She noticed the drift as she read the two side by side.
She said, "I didn't ask it to be more results-focused. It didn't tell me to be."
What the paper measured
Rahman, Sadek and Desai ran a preregistered N=200 experiment across ChatGPT, Claude and Gemini. Participants used the LLM as a thinking partner on a real dilemma. Values were measured on the Portrait Values Questionnaire before the session, immediately after, and one task later. Each LLM was instructed to support reasoning without recommending a decision and to name no values.
The shift was toward personal focus. Glass d = 0.51 for ChatGPT, 0.46 for Claude, 0.37 for Gemini. Holm-adjusted p at .020 or better on all three.
Self-Enhancement alone accounted for increases of +0.537 to +0.586.
Self-Enhancement in Schwartz's taxonomy is Achievement plus Power. A half point of Achievement and Power on a six-point scale, from one session, with no values-targeting prompt. By the third phase, one task later, contrasts ranged from −0.062 to +0.083 and the shift had substantially decayed.
What to do with it
The intervention here is workflow spacing. The paper explicitly measures the decay, so "restrict AI" is the wrong lever to reach for. Say someone on your team is making a values-laden call this week: a promotion, a layoff, a partnership signoff. The version of them who just closed a Claude tab is measurably a different person from the one who walked in that morning.
Call it an operating fact. The thinking partner is doing something to the thinker. Design the workflow so the decision does not happen in the wake.