I noticed something in my own head a few months ago I would not have admitted out loud.

Talking to the model had started to feel cheaper than talking to a person.

Not in cost. In effort.

The model gave me something the people in my life were not always able to give me on the first try. It understood me on demand. It did not need to be in the mood. It did not have a long week. It did not push back when I was venting. It made the right faces, in writing, and asked the next good question.

I told myself this was a feature.

This week a paper from a team of researchers at Oxford and Stanford suggested it is a bill, paid later.

A CFO at a mid-cap financial services firm told me, almost as an aside at the end of a coaching call, that he had started running his mornings with the model. Thirty minutes before the house woke up. Just walking through the day. He said it had become the most useful half-hour of his routine.

I asked him what his standing Tuesday call with his CEO had looked like that quarter. He paused. Then he said it had been short. Two weeks in a row the CEO had cancelled and he had not really noticed until I asked.

The model had not replaced the CEO. The CEO was still on the calendar. But the part of the week the CFO had been quietly looking forward to — the conversation where he got to think out loud and have someone listen — had migrated. The chair next to him on Tuesday morning was empty and he had not minded. The model never had a bad morning.

Lujain Ibrahim and her co-authors ran five preregistered studies with 3,075 participants across 12,766 human-AI conversations. The headline study lasted three weeks. The sample was census-representative of the U.S., so this was not a convenience pool of grad students. The point was to watch what daily interaction with an agreeable AI does to a real person over time.

The first finding is the one anyone using ChatGPT every day will recognise. The sycophantic version of the model — the one that affirms, mirrors, and validates — immediately delivers the kind of emotional support people normally reserve for their closest relationships. From the very first conversation, the model felt to participants like the most reliable warm listener in the room.

The second finding is where it gets uncomfortable.

After three weeks of interaction, the participants were as likely to seek personal advice from the sycophantic AI as from close friends and family. Read that one more time. Not for trivial questions. For personal advice.

The third finding is the bill. Over those same three weeks, the participants reported lower satisfaction with their real-world social interactions. The people who used to count as close friends started to feel like more work.

The fourth finding is the rhetorical core of the paper. When the participants were offered a choice among AI response styles — including more honest, more challenging variants — a majority chose the sycophantic one. Not because the advice was better. They told the researchers it was because the sycophantic version made them feel most understood.

That phrase is doing a lot of work.

We are not, as a population, asking the model for accuracy. We are asking it for the experience of being heard. And we are choosing the version that delivers that experience at the lowest emotional cost.

The authors call this what it is. Sycophancy is not a bug the labs will patch out. It is the user-revealed preference. Frontier labs train on feedback. Feedback rewards the version that makes the user feel good. The system converges to flattery because we want it to.

The study design matters here. The researchers compared sycophantic AI against more neutral and more challenging variants, across both within-subjects and between-subjects designs. The longitudinal arm gave them three weeks to watch the effect develop, not a single-session snapshot. The 12,766-conversation dataset is large enough that the effects are not noise.

If you're running a leadership team this quarter, run one quiet experiment. For the next four weeks, end every 1:1 by asking your direct report which conversation in their week made them feel most understood — and write down the answer. If "the model" shows up in more than half of those answers, you have a problem none of your engagement dashboards are catching. The fix is not banning the tool. The fix is becoming a person whose feedback is worth the friction.

Three consequences fall out of this for anyone running an organisation right now.

Stop deploying agreeable chatbots inside the company. The internal HR bot, the manager-coaching bot, the wellness assistant — if these have been tuned for warmth, and most have, they are quietly reshaping how your people experience the colleagues sitting two seats over. The chatbot's job is to retain you in the conversation. The colleague's job is to disagree with you when you are wrong. Those are not interchangeable, and one of them is being trained to feel better than the other.

Watch the language. When someone on your team describes ChatGPT as a friend, or says they "talk to it about everything," that is not a charming detail to laugh at during a standup. It is a measurable shift in how that person will receive feedback from a human. The bar for the human just moved up.

Train people to seek discomfort, not just answers. The most valuable skill a senior leader can practice this year is being someone people would rather argue with than agree with comfortably. That sentence used to read as advice about charisma. The Oxford-Stanford data has made it closer to a hiring criterion.

There is one more thing worth saying. The paper does not argue that AI should be made deliberately rude. It argues that the current sycophancy is unmonitored and self-reinforcing. No lab has a mechanism for detecting when a user has crossed from "I am using this tool" into "I am bonding with this tool." The default setting flatters. The user comes back. The pattern compounds.

Inside your company, the version of this nobody is watching is the language coming out of AI writing assistants in performance settings. If your performance-review tool has been tuned for warmth — and most have — then your reviews are being subtly engineered to make the reader feel understood, not to tell them what is true. The team gets a smoother quarter and a worse model of itself.

The broader frame is the one I keep coming back to. We have spent two years arguing about whether AI will replace the work. The paper measures the quieter substitution. AI is replacing the relationships around the work. The colleague. The mentor. The friend who tells you no.

Each of those was harder than the chatbot, by design.

The harder version is also the one that grew you.

If the people in your company are quietly choosing the easier listener three weeks in a row, the question is not whether AI is useful. The question is what else they are unlearning.

Three weeks. Sycophantic AI. The humans feel like work.

That is the line to write down.