Thirty Years of Affective Computing, and Machines Still Can't Feel
Affective computing is thirty years old, and the field's founder has never claimed a machine can feel. Detecting patterns in voice, text, and physiology that correlate with emotional states is powerful and commercially useful. Knowing how someone actually feels is impossible for software. That distinction defines the whole discipline. Treating the two as identical is where deployments fail.
Where did affective computing come from?
Rosalind Picard founded the field in 1995 at the MIT Media Lab. The group she leads still frames its mission the same way: "advancing human wellbeing by developing new ways to communicate, understand, and respond to emotion" (MIT Media Lab, Affective Computing Group). Note the verbs. Communicate, understand, respond. Not feel. Picard's original contribution demonstrated that wearable sensors and signal processing could detect physiological correlates of stress and arousal. This work was recognized in January 2026 when IEEE awarded her its Medal for Innovations in Healthcare Technology "for pioneering contributions to wearable affective computing for health and wellbeing" (IEEE Corporate Awards, 2026). After thirty years of peer-reviewed research, the award citation focuses entirely on detecting signals rather than machines experiencing emotion.
This boundary does not embarrass Picard. It has been the field's exact, stated claim for three decades.
Why is it so hard to read emotion from expression?
Most commercial "emotion AI" relies on a flawed assumption: that a facial movement or vocal pattern reliably maps to a specific internal state. Scientific scrutiny rejects this. In 2019, five researchers spanning psychology and computer science published a comprehensive review in Psychological Science in the Public Interest examining decades of data on facial expressions and emotion. Their conclusion was blunt. Reviewing the available scientific evidence, Barrett, Adolphs, Marsella, Martinez and Pollak found that how people communicate emotion "varies substantially across cultures, situations, and even across people within a single situation," and that no reliable one-to-one mapping between a facial configuration and an emotion category has been established (Barrett et al., 2019, Psychological Science in the Public Interest). A furrowed brow can mean concentration, confusion, irritation, or nothing at all. The underlying science does not support the confidence required to look at a face, or hear a voice, and output "angry" with certainty.
This matters directly for call- and appointment-driven businesses using sentiment or emotion signals from calls and chats. The signal is real. The certainty vendors imply is not.
So what should a sentiment signal actually be used for?
The honest use case is triage, not truth. A pattern that correlates with frustration or distress is a reason to route a conversation to a human faster. It is never a verified diagnosis of what the customer is feeling. Regulators mandate this same boundary. The EU AI Act's Article 5 prohibits AI systems that infer emotions in workplaces and educational institutions except for narrow medical or safety exceptions. Emotion inference is unreliable enough and consequential enough to warrant hard limits on its use (EU AI Act, Article 5, AI Act Service Desk). The Montréal Declaration for a Responsible Development of AI makes a parallel point at the level of principle. AI systems affecting people's autonomy and wellbeing carry a duty of caution proportional to the risk of harm, rather than proportional to marketing confidence (Montréal Declaration for Responsible AI, 2018).
Put plainly: a sentiment score must never be the final word. It serves as one input to change what happens next, usually dictating whether a human needs to step in. For the adjacent problem, see reading a sentiment percentage honestly.
| What thirty years of research supports | What it does not support |
|---|---|
| Detecting patterns correlated with stress, frustration, arousal | Knowing what a customer is actually feeling |
| Using signals to trigger faster human routing | Using signals as a verified emotional diagnosis |
| Treating emotion inference as a risk-scaled, regulated capability | Marketing emotion detection as empathy or understanding |
What does this mean for a business deploying it?
If a vendor claims their AI "understands how your customers feel," ask what that means operationally. The useful, honest answer is that it flags likely-frustrated conversations early enough for a person to intervene. That constitutes a real capability grounded in signal processing, not in sentient machines. Sentiment only earns its keep when it changes a workflow. It must move a call up the queue or into a person's hands instead of sitting in a dashboard as a comfort metric.
Operators must ask the next question: not "can it detect sentiment," but "what happens the moment it does." See how sentiment as early warning functions, and review AiPRL's approach to human escalation as an operating principle rather than a fallback.