From a systems perspective: should AI coaches optimize for comfort or correction?

Hey PH 👋

We’re building LoveActually.ai, an AI matchmaker launching soon.

I wanted to share a technical dilemma we ran into and hear how other builders think about it.

Most conversational AI systems are tuned to be supportive and agreeable.

From an engineering standpoint, that’s a reasonable default — it minimizes risk.

But in our dating use case, that tuning caused a failure mode:

the system preserved emotional comfort while reinforcing bad patterns.

So we made a deliberate system-level choice.

We tuned our AI matchmaker to prioritize rational critique over emotional smoothing in certain contexts — especially when user behavior, preferences, and outcomes clearly diverge.

Technically, this wasn’t just a prompt tweak.

It affected:

• how much context the system accumulates before giving critique

• when feedback is delayed vs immediate

• how intensity scales over time

The outcome in beta was mixed:

• retention increased

• but we also received tickets about hurt feelings

Which raised a real systems question for us:

Should AI coaching systems be optimized to minimize emotional friction — or to correct user blind spots, even when that creates friction?

From a technical perspective, is there a principled way to balance:

• safety

• usefulness

• and long-term outcome improvement?

Curious how others here approach this tradeoff, especially builders working on AI coaching, education, or behavior-change products.