Affective computing
The term comes from Rosalind Picard at the MIT Media Lab, who in 1995 defined affective computing as “computing that relates to, arises from, or deliberately influences emotions.” The definition is broad on purpose. It covers machines that recognize emotion, machines that express it, and machines designed to change how people feel.
Expression and influence often meet. In a study in IEEE Transactions on Affective Computing, the field’s own journal, our colleague Fotis Efthymiou and his coauthor Christian Hildebrand gave a synthetic voice a slight tremble. Listeners saw the agent as more vulnerable, felt more empathic concern, and were more willing to donate to charity. In a large field test, the trembling voice also raised click-through on charity ads.
Three decades on, affective computing is an established discipline. When people use the term today, they usually mean the science: the methods, models, and evidence that the other two terms depend on.
Emotional AI
Emotional AI, often shortened to emotion AI, is the applied side. Andrew McStay, who directs the Emotional AI Lab at Bangor University, defines it as technologies that “use affective computing and artificial intelligence techniques to sense, learn about and interact with human emotional life.” In practice the term mostly describes systems that read emotion, such as software that scores a caller’s frustration for a call center or tracks viewers’ faces while they watch an ad. Sentiment analysis, which sorts text into positive, negative, and neutral, is its simplest form.
How well it works depends on the signal. Faces are a weaker guide than is commonly assumed. A major 2019 review found that people do smile when happy and scowl when angry more often than chance would predict, but so inconsistently across cultures, situations, and individuals that a facial configuration cannot be read as a reliable sign of one emotion.
Newer, data-driven research maps emotion far more finely. In a study across five countries and three languages led by our colleague Jeff Brooks, a model trained on thousands of recorded sighs, laughs, and other vocal bursts identified 24 distinct kinds of vocal expression, and the meanings people attached to them were 79% preserved across cultures. Semantic space theory, developed by Dacher Keltner, Jeff Brooks, and Alan Cowen, generalizes the point: emotion comes in upward of 20 distinct kinds that blend into one another, not in a handful of sharply bounded categories.
Our own measurement work builds on this. We track up to 48 distinct emotional expressions in a speaker’s voice, sentence by sentence, and in a working paper we show that this granularity more than doubles the predictive power of traditional emotion measures.
Empathic AI
Empathic AI is defined by what a system does with emotion, not by how it detects it. It acknowledges frustration before solving the problem, slows down for someone who is upset, or stops pressing when someone hesitates. Our explainer on what AI empathy is covers the definition in detail.
The two capabilities come apart. A text-only language model can respond with empathy without any emotion recognition technology, by inferring feelings from what people write. A dashboard that flags angry customers for a supervisor is emotion AI that never responds to anyone. Systems that combine both are newer. Empathic voice interfaces, such as the Hume AI technology used in many of our experiments, measure the emotion in a speaker’s voice and use it to shape the reply as the conversation happens.
Standards bodies use a third phrase for the same family of systems. The IEEE’s 7014-2024 standard addresses “emulated empathy” and brings affective computing, emotion AI, and related fields under one ethical framework. Its 2026 extension, IEEE 7014.1, applies it to AI products marketed as companions, personal AI, and assistants.
Side by side
| What it is | The question it answers | An example | |
|---|---|---|---|
| Affective computing | A research field, named in 1995 | How can computers sense, express, and influence emotion? | Testing whether a trembling synthetic voice changes what listeners do |
| Emotional AI | Technology that reads emotion | What is this person feeling? | Software that scores a caller’s frustration for a call center |
| Empathic AI | A way of responding to people | What response would this person experience as understanding? | A voice assistant that acknowledges a failed request and changes its approach |
Where the law draws the line
European law already treats these as different things. The EU AI Act defines an emotion recognition system as one that identifies or infers people’s emotions or intentions from biometric data. Since 2 February 2025 such systems have been banned in workplaces and schools, except for medical or safety reasons. Since 2 August 2026, anyone using one elsewhere must tell the people it is used on.
The definition turns on biometric data. Under the European Commission’s guidelines, a system that infers emotion from written text falls outside it, while the same inference from a face, a voice, or even the way someone types falls inside. The Act’s preamble is unusually direct about why it restricts these systems, citing “the limited reliability, the lack of specificity and the limited generalisability” of emotion recognition.
Empathic AI is regulated through a different door. The same Act prohibits AI systems that use purposefully manipulative or deceptive techniques to distort people’s decisions in ways that cause significant harm, and requires that people be told when they are talking to an AI. We cover where that line falls in when AI persuasion becomes manipulation.
Common questions
Is sentiment analysis the same as emotion AI?
It is the simplest form of it. Sentiment analysis sorts text into positive, negative, or neutral. Emotion AI usually goes further, estimating specific emotions and how strong they are, often from the voice or face as well as text.
Are “emotional AI” and “emotion AI” different things?
No. The two names are used interchangeably for technology that detects or infers human emotion.