We learned that “being nice” is a technical decision, not a personality trait

While building our AI matchmaker at LoveActually.ai, we ran into an unexpected engineering lesson.

“Niceness” in AI isn’t free.

It’s a tuning choice.

Most models default to minimizing user discomfort.

In dating, that often means preserving feelings while repeating the same failure patterns.

We experimented with shifting the system slightly:

less emotional smoothing, more pattern reflection.

What changed wasn’t accuracy — it was friction.

Users either felt:

• “finally, something honest”

• or “this feels too real”

From a systems point of view, that raised a bigger question:

Should AI coaching systems minimize friction — or manage it?

Because friction isn’t always a bug.

Sometimes it’s the signal.

Curious how other builders here think about this tradeoff.