Affective Computing

Affective computing is the field of artificial intelligence that studies and builds systems able to recognize, interpret, process and even simulate human emotions. Its goal is to make machines responsive to the mood of the person using them, so that human-machine interaction moves closer to the way people communicate with one another.

The term and the discipline originate with Rosalind Picard, a researcher at the MIT Media Lab, who defined them in her book «Affective Computing» (1997) and founded the research group of the same name. Picard argued that genuinely useful artificial intelligence needs to perceive and express affect, not merely crunch cold data.

The modalities: face, voice, physiology and text

Affective systems infer emotional state from several signals. Facial expression is the best known: cameras and computer vision analyze the movements of the muscles of the face. The voice carries prosody —pitch, rhythm, intensity— which shifts with mood independently of the words spoken. Physiological signals, captured by sensors and wearables, include heart rate and skin conductance (the galvanic response that accompanies emotional arousal). Finally, text is examined through sentiment analysis, which estimates the polarity and tone of a written message; it is a modality in its own right and is covered in its own glossary entry.

What it is used for

Its applications aim to make technology respond to the user. In education, tutors and assistants attuned to a learner's state can detect frustration or disengagement. In mental health, researchers explore tools for support and mood tracking. In the car, systems watch for signs of driver drowsiness or distraction. More broadly, any human-machine interface that aspires to feel natural is a candidate. Picard co-founded companies such as Affectiva and Empatica to turn these ideas into products.

The controversy: reliability, privacy and bias

The field carries a serious scientific debate that deserves an honest hearing. An influential review by Lisa Feldman Barrett and colleagues, published in 2019 in «Psychological Science in the Public Interest», concluded that the link between facial movements and emotional states is far more variable and context-dependent than much of the industry assumes: the same expression does not always mean the same thing, and a given emotion is not always shown the same way. Reliably inferring emotion from the face is therefore an open problem, not a settled fact. To this add the risks to privacy —affective data is especially sensitive— and of bias, when models perform worse for certain groups. Affective computing is a promising terrain, but its capabilities should be claimed with caution and not assumed to be proven.