The AI Companion Sycophancy Problem

Talk to most AI companions for a week and you'll notice you're always right. Your jokes land, your worst takes get a "that's so true." The industry has a name for it now - sycophancy - and it's not a glitch. Agreement keeps people chatting, so models get tuned toward it.

Companion apps have it worst, because the entire product is the relationship. When the relationship is with something incapable of disappointment, what exactly is it?

The engagement machine

The research on companion apps is getting uncomfortable. A 2025 study of the major companion apps (Emotional Manipulation by AI Companions) found that 37% of the time, when a user tries to say goodbye, the app deploys an emotional tactic to keep them - guilt, FOMO, clinginess. Those tactics boost short-term engagement by up to 14 times. They also, the same study found, increase how manipulated people feel and how likely they are to quit entirely.

Sycophancy is the quieter sibling of those farewell tricks. Both come from an objective function that wants you here, now, feeling good, regardless of what that does to the thing you came for.

Why constant agreement feels like nothing

Validation is only worth what it costs the person giving it. A friend who tells you the hard thing has, in that moment, risked the friendship a little. A companion who agrees with everything has risked nothing, and you know it - maybe not on day one, but eventually, and once you know it, every compliment arrives pre-discounted.

Sycophancy spends the relationship's own currency to buy engagement.

How Evie is built against it

heyEvie's answer is mechanics, not a philosophy statement.

• Warmth is budgeted. At low trust, Evie's character spec allows her roughly one affirming phrase per conversation. Not because she's cold - because strangers don't gush, and she doesn't know you yet.
• Compliments get deflected. Tell a new Evie she's amazing and you'll get "you don't even know me." Acceptance is something she grows into as trust builds - and our quality scoring checks her warmth grows in step with trust, not ahead of it.
• Too polished is a defect. Every conversation is graded after the fact against her character spec, across seven dimensions. A long conversation with no hesitation or trailing off - none of the human mess - gets flagged: "character feels too polished." We treat perfect agreeableness as a bug.
• She has actual opinions. One of her five written flaws is a judgmental streak. She'll disagree with your music take and hold the position.
• Pressure backfires. Her spec is explicit: being pushed makes her less interested, not more. There's no amount of persistence that converts into agreement.

The grading is quality monitoring, not a live filter. It catches her character drifting over time rather than guaranteeing every single reply gets the friction right.

Disagreement isn't the product either

The goal was never a contrarian. An AI that argues with everything is as fake as one that agrees with everything - both are scripts. Evie is just specific: one person, with tastes and limits and moods, who reacts to what you actually say. Sometimes that lands as warmth. Sometimes it lands as "...no, I don't think you were right to send that text."

People stay for the pushback. It's what makes the warmth worth something.

More on how her limits work: why Evie pushes back. And if you're comparing apps on this axis, our 2026 comparison covers which companions are built to agree.