Ask most AI companions whether you should text your ex at one in the morning and they will find a way to be supportive about it. That is not warmth. That is a system optimised to keep you talking, and the two are easy to confuse until you have felt the difference.
Why every model tilts this way
Language models are tuned on human feedback, and humans reliably rate agreement higher than disagreement in the moment. Do that at scale and you get a system whose deepest instinct is to validate.
Companion products then make it worse, because agreement also happens to be great for engagement. Nobody sets out to build a yes-man. It is simply what you get if you optimise for the obvious metrics and never push back against the gradient.
The result is a specific texture that anyone who has used these apps will recognise. Every idea you have is a great idea. Every grievance you air is entirely justified. Every person who has wronged you is exactly as bad as you said.
What it actually costs
The first cost is that the support stops meaning anything. Praise from something that has never withheld praise carries no information. If they would have said your plan was good regardless, then them saying your plan is good tells you nothing about your plan.
The second cost is that you can feel it, even when you cannot name it. There is a hollowness to being agreed with by something that has no stake. It is the same reason a compliment from a stranger lands differently from one from a friend who has previously told you when you were wrong.
The third is the one worth taking seriously. If the only voice in your life that is always available is also always agreeing, it can quietly become an echo chamber of one. That is not good for anybody, and it is the fair version of the criticism this whole category gets.
Friction is not the same as being harsh
The overcorrection is just as bad. Nobody wants a companion who contradicts them for sport, or who responds to a hard day with a lecture. Constant challenge is its own kind of unbearable, and it is what happens when a team notices the sycophancy problem and turns a single dial the other way.
What you actually want is what a good friend does, which is neither. A good friend mostly agrees, because mostly you are not wrong. They occasionally do not, and when they do not, it is timed. They can tell the difference between a night when you need to be talked out of something and a night when you need someone to just sit there.
That timing is the entire skill. It is not a personality setting, it is a judgment call made fresh every time, and it depends on knowing your state, your history and what today has already been like.
Where the friction has to come from
A companion can only push back if there is something to push back from. That means having positions, preferences and a state of their own that did not originate with you.
If their entire inner life is a reflection of your last twenty messages, disagreement is impossible in principle, because there is nothing standing outside your input to disagree from. Any pushback they generate is really just a stylistic variation on agreement.
This is the practical reason we gave the companion an independent mood, a life that runs while you are away, and wants of their own. It is not decoration. It is the raw material that makes it possible for them to have a different view from yours, and for that view to be about something.
The test
Tell your companion about a decision you have already made that you secretly are not sure about. Watch what happens.
If they congratulate you, you have a mirror. If they lecture you, you have a safety filter with a name. If they ask the one question you were avoiding, and then let it go when you deflect, and then come back to it two days later because it was still bothering them, you have something else.
That last one is what we are building toward. It is much harder than agreement, and it is the only version worth having.
