After an argument, an answer that takes your side can feel like clarity. Before acting on it, open the messages it cites. Check what happened before and after, what the answer leaves out, and whether the same reading fits both people. Apply that check to Grey Mirror too.
Can AI relationship advice be biased toward agreement?
Yes. Research has found that models can affirm a user’s account too readily. A structured report is not automatically free of that problem: counts depend on the input, and model labels and explanations can be wrong. Look for dated source messages, a clear denominator, missing context, and a plausible alternative reading.
What the research found
A March 2026 Science study tested 11 models on interpersonal advice. On the general-advice and Reddit-based prompts, the models affirmed users’ actions 49% more often than human responses, on average. This is a relative comparison for those test conditions, not the probability that any particular answer agrees with you.
Read the study and Stanford’s account in the references below. Model versions, prompts, and the comparison group matter when interpreting the result.
In experiments with more than 2,400 participants, people talking to sycophantic AI reported less willingness to apologize or make amends.
Participants could find an agreeable answer trustworthy even when it reinforced their own position.
The study did not evaluate Grey Mirror. It provides no evidence that our reports prevent agreement bias or improve relationship outcomes.
Separate a count from a judgment
A timestamp calculation and an interpretation of an apology answer different questions. A precise number does not make the surrounding explanation precise.
Check an interpretation in five steps
Use this process for a chatbot reply, a Grey Mirror finding, or your own first reading of a difficult exchange.
Open the cited messages. If there is no source, treat the statement as an unverified interpretation.
Read the surrounding turns in order. An answer may refer to an earlier question, and a later message may change the meaning.
Check both people’s contributions. Would the explanation still sound fair if the names were swapped?
Find a counterexample. A pattern should survive looking at ordinary days as well as difficult ones.
Write down what remains unknown: a call, an in-person conversation, deleted messages, an attachment, or a change in circumstances.
What to expect from a Grey Mirror report
Grey Mirror organizes an exported conversation into measurements, timelines, and interpretations with supporting evidence where available. That structure gives you material to inspect. It does not make every interpretation correct.
Grey Mirror uses machine learning and generated explanations. Those components can share biases with other AI systems. We do not claim the product is bias-free.
Check the imported date range and participant mapping before using a comparison.
Distinguish a calculated value from a model-assigned label or generated explanation.
Open the available evidence and compare the wording with the scope of the claim.
Treat unavailable evidence or low coverage as a reason to withhold a conclusion.
Read a model probability as confidence in a particular judgment, not a measured probability that the relationship is healthy or unhealthy.
Turn a finding into a question you can use
Choose one dated exchange and one observable concern. Ask about the missing context before deciding what the exchange means. A report is more useful as a starting point for reflection than as a verdict to win an argument.
Name the behavior you can point to, such as an unanswered question or a conversation that stopped.
Keep the time period explicit. One week of messages should not become a claim about someone’s entire character.
Leave room for disagreement about interpretation while keeping the original wording available.
If the source is incomplete or the finding does not fit what you know, leave it unresolved.
Separate a count from a judgment
Finding
What to check
What it cannot establish alone
One person initiated more exchanges
What to checkHow an exchange starts, date range, missing messages, and each person’s count
What it cannot establish aloneWho cares more or why someone was quiet
Replies became slower
What to checkMedian or average, overnight gaps, timezone, and changes in work or routine
What it cannot establish aloneLoss of interest or deliberate punishment
An apology was detected
What to checkThe exact wording, attribution, acknowledgment, and what followed
What it cannot establish aloneSincerity, acceptance, or lasting change
A recurring conflict was suggested
What to checkSeveral dated exchanges and counterexamples, rather than one selected excerpt
What it cannot establish aloneA diagnosis, motive, or single cause
Frequently Asked Questions
What is AI sycophancy?
It is a tendency to agree with or flatter a user too readily. In relationship discussions, an answer can reinforce the user’s account without adequately considering the other person’s perspective or missing context.
Does the 49% figure describe every chatbot answer?
No. It is an average relative difference in affirmation compared with human responses on particular prompts in the 2026 study. It is not a universal error rate or a probability for your conversation.
Is Grey Mirror free of agreement bias?
No such claim has been established. Grey Mirror provides structured measurements and evidence to inspect, but its model labels and generated explanations can still be wrong or biased.
Does a full conversation export remove uncertainty?
It provides more text context than a selected screenshot, but still leaves out calls, offline events, deleted messages, and intent. More messages do not justify a diagnosis or a claim about motive.
What should I do with a finding I cannot verify?
Leave it unresolved. Look for the original messages and surrounding context, check coverage and participant attribution, and avoid treating unsupported wording as a fact.
A fluent reply may mirror the framing you supplied without checking the whole conversation.
02
A measured observation
A count or dated pattern should identify its input, denominator and supporting exchanges.
03
A useful challenge
An explanation should leave room for counterexamples and missing context, even when that is less comforting.
Conceptual reading guide. This diagram explains what to inspect; it is not a customer report, diagnostic test or measured relationship score.Use this guide on the question you came with
Ask what evidence would change the conclusion. Confidence and warmth do not establish that an interpretation is correct.