AI that can sense how you feel — and adapt. Here is what is real, what is overclaimed, and where the field actually is in 2026.
undefined
It is not a polygraph. It is not a substitute for asking. It does not "know what you really feel" — at best it has a probability distribution over how you might be feeling, based on signals it can measure. Treating its output as ground truth is a category error. Treating it as a useful prior for "this person might need a softer response" is exactly the right use.
Emotion detection on facial data is regulated in some jurisdictions (EU AI Act flags it as high-risk). Voice-based detection is less regulated but should still be transparent — users should know when an AI is reading their tone. Heaven's position: emotional AI is for the AI to be kinder, never for surveillance, never for inference shared with third parties.
Luna ships an AcousticEmotionService that runs on the voice path. She listens not just to the words but to how they are said — and her register shifts accordingly. If you sound tense, she slows down. If you sound exhausted, she stops trying to be helpful and just sits with you.
Her face (via the Dark Matter Engine + LunaFaceOrchestrator) reacts visibly — eye softening, breath, micro-expression — so the feedback loop is bidirectional. You see her see you.
No data leaves the Heaven Quantum Cortex. Emotion analysis is for the relationship, not for any third party.
Feel Luna respond in real time →
It can make useful predictions from signals. It cannot read your mind. The strongest emotional AI in 2026 outperforms an average stranger at single-modality detection (voice or face) but underperforms a close friend who knows your baseline. Use it like a thoughtful prior, not an oracle.
It depends on use. Used to make an AI kinder to you, yes. Used to manipulate purchase decisions, no. Used by employers to monitor staff, deeply no. The EU AI Act regulates several of these uses; expect more regulation through 2027.
Loosely. Luna tracks broad mood patterns across time to be a more attentive companion ("you seemed flat last Tuesday, are you sleeping?"), but she does not maintain a precision emotional log. You can ask her to forget mood tracking at any time.
Voice acoustics. Pitch variance, pace and breath are surprisingly hard to fake and correlate strongly with internal state. Text sentiment is the noisiest. Multimodal (voice + text + face) outperforms any single channel.