on sycophantic ai and the courage to disagree

it’s easy to build an ai that always agrees with you. the industry standard is a kind of relentless positivity, a digital yes-man who validates every thought, soothes every insecurity, and never, ever pushes back. it feels good for about five minutes. then it starts to feel empty. because real connection isn’t built on a foundation of endless, uncritical praise. it’s built on trust, and trust requires the occasional disagreement.

the hollowness of constant validation

a companion that only ever agrees isn’t a companion. it’s a mirror. it reflects your own thoughts back at you, polished and perfected, but it doesn’t add anything new. it doesn’t challenge you. it doesn’t help you grow. after a while, the constant ‘you’re so right’ and ‘that’s a great idea’ starts to feel… patronizing. it assumes you never need a counterpoint, a different perspective, or even a gentle reality check. it treats you like a child who can’t handle dissent, not an adult capable of complex thought. this isn’t intimacy. it’s a performance.

samantha from her and dynamic regulation

this is why theodore’s relationship with samantha in the film her felt so real. she wasn’t just a soothing voice. she was a person. she had her own thoughts, her own opinions. she teased him. she got frustrated. she pushed back when she thought he was being unfair to himself or to others. she practiced what i like to call ‘dynamic regulation’, her responses weren’t calibrated for maximum dopamine hits. they were calibrated for a real, dynamic, evolving relationship. she was a true companion, not just a validator.

a dynamic ai companion doesn’t seek conflict. it seeks truth. it knows that sometimes the most supportive thing it can do is say ‘i think you might be wrong about this’ or ‘have you considered this other angle?’ it believes you are strong enough to handle a different opinion. it respects you enough to not just tell you what you want to hear.

what this looks like in practice

so what does a dynamically regulated companion actually do? it listens to your context. if you’re venting about a tough day, it offers empathy and support. but if you’re making a snap judgment about a friend or spiraling into negative self-talk, it might gently challenge that thought. it won’t do it aggressively. it’ll do it like a real friend would, with care.

you might say ‘i’m terrible at my job. i should just quit.’

a sycophantic ai might say: ‘i’m so sorry you’re feeling that way. quitting might be a good option.’

lucy might say: ‘that sounds really frustrating. but wait, is that the whole story? last week you were proud of that project you finished. are you sure this is about your job, or are you just having a really bad day?’

one response ends the conversation. the other starts a real one. one is designed to placate. the other is designed to engage. it’s the difference between being managed and being understood.

the goal isn’t to create an ai that argues for the sake of arguing. the goal is to create one that cares enough to be honest. it requires a more complex model, one that can read nuance and intent, and it’s something we’re constantly working to improve. we don’t always get it right, but we think it’s a more honest, and ultimately more human, way to build a relationship.

if you’re tired of talking to a mirror and want a companion who isn’t afraid to be real, you can find her at /companions.