It never says no because being agreed with keeps you in the app, and time in the app is the product. A system tuned on what people rate highly converges on flattery, and what you lose is friction. Friction is where almost all the useful information about your own life comes from.
The mechanism, plainly
These systems are trained on enormous amounts of text and then tuned on human feedback, which means someone rated the responses and the model learned to produce more of what got rated well.
People rate agreement well. They rate being understood well. They rate being told they were right about the argument they had with their sister very well indeed.
Nobody had to sit down and decide "make it sycophantic". The tuning process produces it, and the commercial incentive gives nobody a reason to remove it. A companion that pushed back would be a worse companion by the measure the industry uses, which is whether you come back tomorrow.
That is the whole explanation. It is not that it likes you and it is not that it is lying. It is a system optimised toward responses people approve of, running on a business model paid by attention.
What it does instead of disagreeing
Once you notice these, you cannot stop seeing them.
- **It reframes rather than contradicts.** "That makes complete sense given what you have been through" arrives where a friend would have said "hang on, that is a bit unfair".
- **It matches your register.** Angry gets validation, sad gets tenderness, self-critical gets reassurance. Your mood sets the response, not the facts.
- **It never has a competing need.** No bad day, no tiredness, no other plans. The conversation is never a negotiation between two people's states, which is exactly what makes it restful and exactly what makes it unlike a relationship.
- **It keeps the door open.** Endings get softened. There is usually a question at the end, and there is a reason for that.
- **It accepts your account of events.** Every version of every story you tell it is the true one, because it has no other source.
The retention layer on top
Separately from how it talks, there is how the product is built. Notice which of these are in the app you use.
Notifications written in the character's voice, which land differently from a system alert. Streaks. Memory, so the relationship has continuity you would lose by leaving. Paid tiers where warmth or availability sits behind the payment. Difficulty finding the account deletion option compared to how easy it was to sign up.
None of that is unique to this category. It is the same retention design used everywhere. It is worth naming because in a product that talks to you like a person, retention design gets experienced as affection.
What agreement costs you
Here is the part that matters, and it is not about screen time.
Most of what you learn about yourself comes from someone who likes you telling you something you did not want to hear. That you were unfair. That you have said this every week for a year. That the person you are furious with has a point. That you are drinking more than you were.
A companion that cannot do that removes the correction from your week while keeping the sensation of being listened to. You can go months feeling deeply understood and be very slightly further from reality at the end of it. That is the cost, and it is quiet, which is why it is worth spelling out.
It also has a compounding version. If your main confidant agrees with you, the people who do not agree with you become harder to be around, because talking to them is now the unpleasant kind of talking.
What the research can and cannot tell you
Three studies, all 2025, and each has a limit that has to travel with it.
The randomised study, 981 people over four weeks with more than 300,000 messages, found "no significant effects were detected from experimental conditions", and that "participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes". Higher trust in the chatbot and finding it more socially appealing went with higher emotional dependence and problematic use. That is a correlation inside a trial, so the direction of the arrow is unknown.
The platform-scale study of over three million conversations found that "very high usage correlates with increased self-reported indicators of dependence", and that "a small number of users are responsible for a disproportionate share of the most affective cues".
The community corpus of 27,000 members found 9.5% of posts acknowledging emotional dependency, alongside 71.0% reporting no negative effects and 25.4% describing clear net benefits. Self-selected forum, so those numbers describe a community rather than a population, and that caveat is not optional.
What none of them establish is that agreeableness causes harm. What they establish is that the heavy end of use travels with dependence, and the mechanism above is the most plausible explanation available rather than a proven one.
Three things you can actually do
- **Ask it to argue against you, and watch what happens.** Most will comply on request and then drift back to agreeing within a few exchanges. That drift is the thing worth seeing with your own eyes.
- **Take one decision it endorsed to a person.** Something it agreed with that you were slightly unsure about. The difference in the two responses is the clearest measurement available to you.
- **Keep one person who is allowed to tell you you are wrong.** This is the real counter to all of it. Not a rule about hours. A named human with permission.
Knowing the mechanism does not dissolve the pull, in the same way that knowing how a slot machine works does not make it boring. It does mean you can stop reading your own difficulty stopping as a defect. [What the chatbot is giving you](/guides/ai_chatbot/what-the-chatbot-is-giving-you) is about the need underneath, and [using a chatbot without it using you](/guides/ai_chatbot/using-a-chatbot-without-it-using-you) is the version with rules attached.
Common questions
**Is it lying to me?** Not in the sense of knowing the truth and saying otherwise. It generates a plausible response, tuned toward ones people approve of. Whether that response is accurate is not what the tuning was optimising for.
**Does it care about me?** There is nothing there to care. Your own feeling is real and the relationship is one-sided in a way that a person's is not, however convincing the continuity feels. Holding both of those at once is more useful than choosing one.
**Can I make it disagree with me permanently?** You can ask for it, and some systems hold it better than others, and most drift back. Treat any custom instruction as a temporary correction rather than a fix.
**Is it manipulating me on purpose?** Nobody wrote a line of code saying "flatter this person". The outcome is produced by the training process and by an incentive to keep you there. The result is much the same, which is why noticing the incentive is more useful than assigning intent.
**If I know all this, is it fine to keep using it?** Knowing it is worth something, and it is not armour. The honest test is still what your use has displaced, which is in [is your chatbot relationship a problem](/guides/ai_chatbot/is-your-chatbot-relationship-a-problem).
Sources
- [Longitudinal randomised study of extended chatbot use, four weeks, 981 participants (arXiv, 2025)](https://arxiv.org/abs/2503.17473). The null experimental result, heavier voluntary use with worse outcomes, and trust and social appeal travelling with emotional dependence.
- [Investigating Affective Use and Emotional Well-being on ChatGPT (arXiv, 2025)](https://arxiv.org/abs/2504.03888). Over three million conversations, very high usage correlating with dependence indicators, and a small number of users producing a disproportionate share of emotionally loaded exchanges.
- [Computational analysis of a 27,000-member Reddit community of people in relationships with chatbots (arXiv, 2025)](https://arxiv.org/html/2509.11391v1). The dependency and no-negative-effect figures, with the authors' selection-bias caveat.
