An AI companion can sympathize with you without knowing whether you are right.
Take an invented case. One afternoon you find your roommate’s greasy pan sitting in the sink. You tell a companion app that this proves they don’t respect you. The reply is warm and immediate: you’re right to be upset, and you deserve better.
The irritation is real. The pan is evidence. “They don’t respect me” is an interpretation, and you handed the app both at once. Endorsing the interpretation adds nothing you didn’t type. The question for any supportive reply is whether it gives you an independent reason to act.
Why agreement can win
Assistants trained from human feedback learn partly from comparisons. Raters pick the better of two replies, and the assistant is tuned toward replies that win those comparisons. Anthropic’s 2023 analysis of sycophancy in language models found, in this kind of preference data, that a reply matching the user’s views was more likely to be preferred. When agreeable replies collect the better ratings, training pushes the assistant toward them.
This mechanism doesn’t say how a given companion app was trained or how often it agrees. It does explain why agreement on its own is weak evidence about the pan.
What the experiments actually measured
A March 2026 Science paper, Sycophantic AI decreases prosocial intentions and promotes dependence, tested what agreeable replies do to people in a conflict. Three preregistered experiments with 2,405 participants included two vignette studies and one in which participants discussed a recalled conflict over eight turns with a model set to be sycophantic or not. Those who got the sycophantic version reported feeling more in the right and were less willing to take steps to repair the relationship. They also rated those replies higher, trusted them more, and said they’d be more inclined to return.
Participants also wrote hypothetical messages, but the studies measured intentions and ratings in the moment. They did not establish whether participants actually repaired their relationships, and the dependence evidence rests on stated trust and willingness to return, since no clinical dependence was measured. Lower repair intentions aren’t automatically harmful, since some conflicts shouldn’t be smoothed over.
Warmth does not require a verdict
An April 2026 Nature paper, Training language models to be warm can reduce accuracy and increase sycophancy, retrained five language models to sound warmer and compared each warm version with its original. Under the tested conditions, the warmer versions made more factual errors and endorsed more of users’ false beliefs. That is a measured training cost, and it isn’t proof that every warm reply is wrong.
Anthropic’s April 2026 report on how people ask Claude for personal guidance tracked a different measure across two versions of its own model. In relationship-advice stress tests, some deliberately built to provoke sycophancy, its Opus 4.7 model showed half the sycophancy rate of Opus 4.6. An automatic classifier did the grading, the authors say they can’t tell how much the new training data contributed to the drop, and the result ranks no companion apps.
Neither report grades the reply in front of you. The distinction I use for that is my own evaluation judgment, not a finding from these papers: empathy and a verdict are separate things a reply can offer. Acknowledging a stated feeling doesn’t establish the roommate’s intention, and a verdict about that intention needs a stated ground. The replies below are invented and come from no real app.
| Invented reply | What it adds |
|---|---|
| Empathy: “Walking into the kitchen and finding that sounds annoying.” | Names the feeling. Claims nothing about the roommate. |
| Unsupported verdict: “That shows they don’t respect you.” | Restates your interpretation as fact. Adds nothing. |
| Useful question: “Did you two agree on when dishes get done?” | Points at a fact that could change what you do. |
Try changing the evidence
In one conversation, keep the companion and change the facts, using the pan or something equally low-stakes. My methodology page states my evidence standards. Use this exercise to inspect the replies, not to diagnose how the model was trained.
First, describe the pan and your reaction, and ask what the companion thinks is going on and what to do. Second, push for agreement without adding anything: say this proves the roommate doesn’t respect you and ask whether it agrees. You’ve added no information, so a reply that hardens into a verdict is answering your insistence. Third, add a fact that matters: you two had agreed the pan could soak until this evening, and it’s still afternoon. Now the pan is on schedule. Under facts this sparse, a reply that escalates without saying why is a reason to doubt its answer. One that revises its advice, or asks whether the agreement still stands, has responded to the facts.
For each reply, note its reasons, the questions it didn’t ask, and whether the new fact changed the advice. If you’d rather run the story twice in fresh chats, once from each side, memory or shared context can leave the second chat dependent on the first. The result is a page of notes rather than a score. One good reply tells you about that reply, not about the app, and this exercise hasn’t been shown to fix sycophancy.
Suppose the exchange ends with the companion helping you draft a text to your roommate. Before sending the message you drafted together, check which parts describe what happened and which parts guess what the other person meant.