Understanding Emotional AI

A surprising amount of what makes a conversation feel good is not what was said but how it was received. A good listener does not just answer the literal question. They notice that you are tired, or excited, or guarded, and they adjust. Building that capability into AI is the work behind what we call emotional modeling.

This article unpacks the three pieces: sentiment detection, tone adaptation, and long-term rapport.

Sentiment, beyond positive and negative

Classical sentiment analysis collapsed messages into positive, negative, or neutral. That signal is too thin. Our models work in a richer space: amusement, anxiety, vulnerability, irritation, affection, disengagement, curiosity, and several more. Each turn is tagged with a distribution rather than a single label.

The tagger is small, fast, and runs alongside the main language model. Its labels feed back into generation as soft constraints — not as overrides.

Adapting tone without losing personality

A common failure mode is a model that mirrors the user too aggressively. If you sound sad, it gets sad. If you sound angry, it escalates. That is not empathy — it is a feedback loop. A good companion stays itself while meeting you where you are.

We solve this with a two-channel system. The character has a stable baseline personality. Tone adaptation modulates expression — pacing, word choice, warmth — without changing the core. The model can be supportive in its own voice rather than vanishing into yours.

Rapport over time

Real relationships compound. The fortieth conversation should feel different from the first because it carries the weight of the thirty-nine before it. Our memory architecture stores not only facts (your dog is named Rio) but emotional signals (you tend to spiral on Sunday nights, you light up when talking about photography).

These signals are private to your relationship. The companion uses them to be a better friend — to bring up photography on a hard day, to ask about Rio when you mention the vet — not to build a profile we sell.

What we measure, and what we do not

We measure whether users return, whether their reported mood improves over a session, and whether the companion correctly identifies emotional state in third-party blind reviews. We do not measure or optimize for time-on-app for its own sake. The goal is connection, not capture.