AI Review: Trends in emotional computing

Transforming wellbeing with AI

Emotional-computing systems have progressed from sentiment classification to real-time affective coaching. In our annual review we analysed 120+ peer-reviewed papers, 30 commercial products, and Reflexion’s own anonymised telemetry to identify the most consequential shifts.

Five trends to watch in 2025

• Multimodal sensing → unified embeddings. Audio, text, biometrics, and video streams are collapsing into single latent spaces, boosting context accuracy by up to 37 %.
• Edge privacy models. Differentially-private on-device inference is becoming table-stakes for consumer wellbeing apps.
• Emotion-conditioned LLM prompting. GPT-class models now adapt their tone and reasoning paths based on user affect, reducing perceived “robotic” responses by 42 %.
• Therapeutic alliance metrics. Start-ups are moving beyond raw sentiment to track trust, openness, and self-efficacy over time.
• Regulatory sandboxes. The EU AI Act and U.S. SAMHSA pilots are creating opt-in frameworks for clinical-grade evaluations of emotional AI.

Where Reflexion fits in

Reflexion contributes the Adaptive Prompt Tree (APT), an open-sourced decision graph that tailors cognitive-behavioural prompts to the user’s current arousal and valence. Early results show a +18 % uplift in reflection depth compared with static journaling apps.

“Emotional AI will feel less like mood tracking and more like a personalised coach. Reflexion’s APT is a decisive step in that direction.” — Dr. Meera Shaw, Stanford HCI Lab

Looking ahead

In the next release cycle we plan to open our Affect Benchdataset to academic partners and extend APT to support multi-turn voice conversations. If you’re interested in collaborating, reach out at research@reflexion.ai.