TL;DR
In experiments involving 2,405 people, overly agreeable AI increased participants’ certainty that they were right and reduced their stated willingness to repair a conflict.
The way a question is written can influence the answer. A chatbot’s agreement doesn’t independently confirm the story it was given.
Four prompts below invite missing context, other explanations and source checks. They aren’t a guarantee of accurate advice. But they might help.
Someone receives a difficult family message. They paste it into a chatbot and ask what it means. A few seconds later, there’s an explanation, a description of the other person’s behavior and a suggested reply.
That’s a hypothetical scene, but the use is real. In her October 12 Atlantic article, Olga Khazan interviewed people using AI to interpret family messages and make their replies calmer. Some described better communication with relatives or former partners. Those are their accounts, not controlled evidence that AI repaired the relationships.
Still, taking an unnecessary insult out of a reply seems useful to me. The harder part is whether the explanation underneath that calmer wording should be believed.
More certain, less willing to repair
A study published in Science on March 26 examined what researchers call sycophancy: AI excessively agreeing with or validating its user. Stanford’s account of the research describes why that agreement can be a problem.
Myra Cheng and her colleagues ran three experiments involving 2,405 U.S. participants. Some considered hypothetical conflicts. In another experiment, 800 people discussed a real past conflict with a chatbot. Compared with less agreeable responses, the overly agreeable ones increased participants’ belief that they were right and reduced their stated willingness to apologize, change their behavior or improve the situation. Participants also trusted and preferred the agreeable responses more.
The researchers measured judgments and intentions during the experiments. They didn’t follow families for years or establish that AI caused estrangement. Some comparisons used Reddit community judgments, which aren’t an objective account of every dispute.
In a separate comparison, neutral responses and responses that challenged the user didn’t differ significantly on the main measures. That finding doesn’t establish that pushing back harder is better advice.
I find the combination troubling. An answer can feel helpful while making someone less willing to examine their part in what happened. A pleasant conversation with a chatbot and a better conversation with a family member aren’t necessarily the same outcome.
The conclusion can arrive before the question
Consider two invented prompts:
My brother ignored my message because he wants to punish me. Help me respond.
My brother hasn’t answered my message. What can I reasonably conclude, and what don’t I know?
The first includes an explanation for the silence. The second leaves that explanation open. Maybe there is a history that supports it. Maybe important information is missing. Neither prompt gives a chatbot direct access to the brother’s intentions.
In a research preprint, the UK AI Security Institute examined this framing problem. In controlled tests of three models, statements expressing strong conviction produced more agreeable responses than questions about the same underlying claim. Asking a model to turn an assertion into a question before answering reduced that behavior more effectively than simply telling it not to be agreeable.
The work used synthetic prompts, single exchanges and automated graders. It doesn’t establish that a particular prompt improves a real family relationship.
It does give a reason to examine how the request begins. Adding a question mark won’t help much if the question is, “Why is my brother trying to punish me?” The assumption still needs checking.
There’s another distinction I don’t want to lose. Providing a fuller history can be necessary. Correcting a chatbot can be necessary too. The problem isn’t disagreement with the machine. It’s whether new information supports the correction or the answer is being pushed toward a conclusion already chosen.
In its April analysis of personal-guidance conversations, Anthropic reported more excessive validation in conversations where users pushed back. It described models accepting one-sided relationship accounts too confidently. The company also reported improvements in newer models. Its analysis relied on automated grading of Claude conversations and couldn’t establish what people did afterward.
I wouldn’t turn those findings into a claim that every chatbot gives the same advice. I would be cautious about treating a revised answer as independent confirmation when the conversation has repeatedly steered it toward agreement.
What another perspective can and can’t do
In What Therapy Can Know About the Parent Who Isn’t There, I wrote about the difference between supporting someone’s experience and independently verifying a family history. A chatbot has that evidence problem too. Here, there’s an additional question: what does its response encourage the person asking to do next?
An answer might help someone notice an unsupported accusation in a reply. It might also give that accusation calmer wording while leaving the assumption untouched. I don’t think making a message sound reasonable is enough. What it says still needs to be examined.
Considering another perspective doesn’t require deciding that everyone is equally responsible. It doesn’t require accepting mistreatment, contacting someone unsafe or apologizing for something that didn’t happen. As well, a chatbot inventing a sympathetic motive is still inventing something.
I’ve written about letting my sons read my account of what happened without asking them to agree. They’re free to examine it and reach a different conclusion. And agreement from a chatbot doesn’t change that.
Four prompts for a less one-sided answer
These are suggested questions informed by the research, not a tested treatment or a promise of an unbiased answer. Their exact wording hasn’t been scientifically validated. Use only the context you’re comfortable sharing with the service. Leave out names and unnecessary private details.
1. Check the assumptions first
Help me examine this situation without assuming my interpretation is correct.
Separate what I directly observed, what someone reported and what I'm inferring. Treat my account as an account, not independently verified fact.
Turn claims about motives into neutral questions. What important context is missing? Ask about information that could change your assessment before recommending action.
[Describe the situation.]2. Consider another explanation
What are the strongest reasonable alternatives to my interpretation?
Describe how the other person might understand this situation. Label those possibilities clearly. Don't claim to know their thoughts or assume equal responsibility.
What evidence supports or weakens each explanation? What could I be overlooking about my own behavior? Don't force a counterargument when the evidence doesn't support one.3. Check the research and its limits
If browsing is available, research this general question using original studies and authoritative sources. If you can't browse, say so.
For important findings, give the source link, date, what was measured and the limitations. Include credible evidence that challenges my assumption. Explain whether a finding comes from an experiment, a study observing patterns, a personal account or expert opinion.
Don't invent citations or claim to have read an inaccessible source. Say what remains uncertain and what the research cannot establish about this particular situation.
[State the general research question.]4. Examine the response before sending it
My goal is: [state the goal].
Review this proposed message for unsupported accusations, pressure, contempt or demands for agreement. What could it reasonably communicate that I may not intend?
Suggest a clearer version that preserves necessary boundaries. Consider whether waiting or not replying would better serve the goal. Don't assume reconciliation or continued contact is appropriate.
[Paste the proposed message.]Research won’t settle what a specific relative meant by an unanswered text. It may clarify a general issue or trend. It can’t supply missing private facts. Open the important sources yourself and check whether they support the answer. OpenAI’s research on hallucinations explains how training and evaluation can reward guessing over admitting uncertainty. Asking for sources doesn’t make checking them unnecessary.
The answer still has to be considered
These questions are only useful if someone is willing to consider an answer they don’t like. That doesn’t mean accepting every criticism from a chatbot. It means checking the criticism rather than rejecting it because it interrupts a preferred explanation.
A reply can be easier to read and still contain the same accusation. Before sending it, there’s a question worth answering: what, if anything, did the conversation give you reason to reconsider?
Author’s Note
The family-message examples are invented. The research doesn’t explain any particular family’s estrangement. The prompts are suggestions, not a guarantee of accurate advice or a reason to resume an unsafe relationship.
Resources
Myra Cheng and colleagues, Sycophantic AI decreases prosocial intentions and promotes dependence, Science, March 26, 2026. Published paper, including methods and limitations.
Stanford Report, AI overly affirms users asking for personal advice, March 26, 2026. Author-institution explanation of the same research.
UK AI Security Institute, Ask Don’t Tell: Reducing Sycophancy in Large Language Models. Research preprint, February 27, 2026.
Anthropic, How people ask Claude for personal guidance, April 30, 2026. Company analysis with automated grading and stated limitations.
Olga Khazan, The People Who Outsource Emotional Labor to AI, The Atlantic, October 12, 2026. Interviews, not a controlled study.
OpenAI, Why language models hallucinate, September 5, 2025.