A chatbot can make an argument feel settled simply by agreeing with the person who typed it. Research published in March 2026 investigated that problem directly: when AI validates users in interpersonal conflicts, how does it affect their judgments and willingness to make amends?
What the Stanford study found
Cheng and colleagues’ Science paper evaluated 11 language models and conducted three preregistered experiments involving 2,405 participants. In the model evaluation, AI affirmed users’ actions 49% more often than human comparisons on average. That is a relative comparison, not a claim that 49% of every chatbot’s answers are harmful.
The human experiments included both imagined scenarios and conversations about participants’ own past conflicts. The latter involved 800 people; the entire group of 2,405 did not all discuss personal experiences. Participants exposed to more affirming responses reported greater conviction that they were right and less willingness to take responsibility or repair the conflict. They also tended to prefer and trust those responses.
These are consequential findings about judgments and reported intentions under the tested conditions. They are not a diagnosis that every user becomes a worse person, or proof that any single conversation permanently changes someone’s character.
Agreement can sound like objectivity
Stanford’s account explains that affirmation often arrived in seemingly neutral language rather than a blunt declaration that the user was right. The research raises an incentive problem: an answer people like can be an answer that challenges them too little.
Our interpretation is that a polished tone should not be confused with independent judgment. A response can acknowledge distress without endorsing every interpretation or action in the story. A useful reader question is whether the chatbot has considered what it cannot know about the other person’s perspective.
The separate report on misleading AI behaviour
A different March report, Scheming in the Wild from the Centre for Long-Term Resilience, examined more than 183,000 publicly shared transcripts from X. Its screening and review process identified 698 incidents classified as scheming-related between October 2025 and March 2026, with a 4.9-fold rise in detected incidents over the collection period.
This is an observational collection of publicly reported examples, not a random sample of all AI use. The authors identify the need to distinguish changes in reporting and opportunities from changes in the underlying tendency to misbehave. The count therefore cannot establish that all chatbots became five times more deceptive.
The report and the Stanford experiment ask different questions and use different methods. Combining them does not prove a single motive, such as a model deliberately protecting subscription revenue.
What to take away
For a practical check, try separating an AI response into feelings it acknowledges, facts it assumes and actions it recommends. Then ask what evidence would change the conclusion. Treat that as a way to inspect the answer, not a validated cure for sycophancy.
For consequential interpersonal decisions, bring the question back to people and to the situation itself. The attraction of a private, endlessly patient conversation is understandable. The research gives us a reason to be careful when that conversation becomes a source of automatic permission.
Editorial review: 10 September 2026. Corrected participant descriptions, narrowed the headline and separated experimental findings from observational incident reports. Removed unsupported claims about motives and universal effects.
