If you keep hearing that machines can read feelings, what is affective computing really? It is the field of building systems that sense, respond to, or shape human emotion. In practice, that means reading signals like expression or skin response, then adapting.
Affective computing is one strand of emerging technologies that change how people and machines interact.
| Item | Value |
|---|---|
| Working definition | Computing that relates to, arises from, or deliberately influences emotion |
| Emotion categories in a well-known wearable-sensor test | Eight emotions |
| Classification accuracy reached in that test | 81% |
| Accuracy of a random guess in that test | 12.5% |
| Users in a frustration-agent experiment | 70 users |
| Computer users who admit to swearing at computers (1999 survey) | Three-quarters |
The figures come from Rosalind Picard’s 1999 paper on affective computing for HCI and her 2003 paper on affective computing challenges. Figures can change, so check the sources before you rely on them.
Where the idea comes from
The MIT Media Lab report TR-321, written by Picard, starts from a neuroscience claim: emotions play a critical role in rational decision-making, in perception and in human interaction. If that is right, a machine that ignores emotion is ignoring useful information. You can read the abstract on the MIT Media Lab site.
In her paper on affective computing challenges, Picard wrote that researchers around the world had built machines with several emotional abilities. Those abilities are:
- recognizing emotional information
- expressing it
- modelling it
- communicating it
- responding to it
These abilities can be separated in a machine more easily than in a person. A machine can fake the appearance of an emotion without having any feeling behind it, which is why showing a smile and feeling happy are different things.
How do machines sense emotion?
The research reports describe several routes to emotional signals. Each has strengths and limits.
- Body signals: skin-surface sensors that pick up measures such as blood volume pulse and respiration, often in a wearable form.
- Facial muscles: experimental eyeglasses that sense changes in facial muscles, such as furrowing the brow in confusion or interest.
- Behavior with software: a 1999 system combined two physiological signals with mouse clicks to spot when a user was having problems.
- Speech: early work on analyzing affect in speech together with other natural ways of expressing feeling.
To see how cameras find and match faces, read the explainer on how does facial recognition work. Reading an expression is a different job from matching an identity.
The reports also admit a weak spot: traditional pattern recognition and signal processing methods have difficulty with day-to-day and person-to-person variation in how emotion is expressed. That is why the researchers called for better machine learning methods.
What the accuracy numbers really mean
One of the first tests asked whether a wearable computer could recognize a person’s emotion over a long period. Picard’s team used four skin-surface sensors on one person for several weeks. Their algorithms reached 81% accuracy on eight emotions, compared with 12.5% for a random guess.
Picard then warned against reading that as “computers recognize emotions with high accuracy”. The experiment forced a choice among eight kinds of emotion, much like early speech recognition that only had to tell spoken digits apart.
Real life is messier. Three things make emotion recognition hard:
- Variation from day to day, even for the same emotion, could be larger than variation among different emotions on the same day.
- People’s expression of emotion is individual and variable.
- Many people do not know how to put their own feeling state into words.
In a separate 1999 study, the system that combined physiological signals with mouse clicks did significantly better than random at detecting problem episodes in 21 out of 24 users. That is promising for a narrow task, and it is a long way from general emotion reading.
Where is affective computing used?
TR-321 describes theoretical and practical applications in several areas:
- learning
- human-computer interaction
- perceptual information retrieval
- creative arts and entertainment
- human health
- machine intelligence
Frustration is the clearest example in the sources. A 1999 study found that 84% of help-desk managers surveyed said users admitted to abusive behavior toward computers, and a separate 1999 survey found that three-quarters of computer users admit to swearing at them.
Picard’s group built an agent that practiced active listening, empathy and sympathy through a plain text dialogue box. In a test with 70 users, it led to behavior indicating a significant decrease in frustration compared with two control conditions. The agent did not need to know what caused the problem.
Should machines read your feelings?
The 1999 paper is direct about consent: people have strong feelings about if, when, where and how they want to share their emotions. It also says it would be absurd if the technology did not respect that, and that designers should give users choices.
The options it describes include:
- no sensing at all
- dialogue boxes that the user controls
- icons that a person can hit or kick to give feedback directly
Where sensing happens also matters for privacy. Processing signals close to the user is one design direction, and the explainer on what is edge ai covers how on-device processing works.
Picard also wrote in 2003 that her 1997 book Affective Computing expressed serious doubt about machines that actually have feelings. The field is mainly about sensing and responding, not about giving machines an inner life.
Frequently asked questions
What is affective computing in simple terms?
Show answer
It is a research field where machines pick up signals that hint at how a person feels, then adjust their behavior. Picard’s papers describe abilities such as recognizing, expressing and responding to emotional information.
Can a computer actually feel emotions?
Show answer
The sources do not show that it can. Picard wrote in 2003 that machines can fake the appearance of an emotion quite well, and that her 1997 book expressed serious doubt about machines having real feelings.
How accurate is emotion recognition?
Show answer
One wearable-sensor test reached 81% on eight emotions for one person. Picard cautioned that this does not mean a computer can recognize anyone’s emotions that accurately, because the test forced a choice among eight categories.
Does affective computing need a camera?
Show answer
No. The research described skin-surface sensors, mouse clicks, speech analysis and eyeglasses that sense facial muscles. It also described a text dialogue box that needed no face or voice at all.
Are these sources up to date?
Show answer
They date from 1999 and 2003, so they show the research foundations rather than today’s products. Check newer work before you judge any current system.
