# Can Machines Read Human Emotions?

IELTS Academic Reading - 14 questions - 20 minutes - Multiple choice, Yes / No / Not Given, Summary completion (word list)

Source: Written by Testmigo.

## The passage

You should spend about 20 minutes on Questions 27-40, which are based on the passage below.

In 1995 Rosalind Picard, a member of the faculty at the Media Lab of the Massachusetts Institute of Technology (MIT), argued in a paper titled 'Affective Computing' that computers should take emotion seriously, and two years later she developed the idea into a book of the same name. Her case was that emotion is bound up with intelligence and with the way people communicate, so a computer that worked closely with humans might need to sense emotional signals and respond to them. Picard was proposing a research programme, not a mind-reading device. The industry now marketed as 'emotion AI' owes a great deal to her work, yet its products often claim far more than the original research ever did.

Affectiva, which spun out of the MIT Media Lab in 2009 with Picard and Rana el Kaliouby among its founders, developed software intended to detect emotion from people's faces. Similar tools soon appeared in recruitment, where some systems analysed recorded video of job applicants answering interview questions. In 2021 Affectiva was bought for $73.5 million by Smart Eye, a Swedish company that supplies driver monitoring systems to carmakers. That market has been helped by law: since 7 July 2024, every new motor vehicle sold in the European Union has had to provide a warning when the driver becomes drowsy.

The science behind these products owes much to a paper published in 1971, in which Paul Ekman and Wallace Friesen described work with the Fore, a relatively isolated people in the highlands of Papua New Guinea, who heard short stories and were asked to pick the photograph of a face that suited each one. They mostly chose the faces the researchers had predicted, which suggested that a few 'basic' emotions, usually listed as anger, fear, happiness, sadness, disgust and surprise, are expressed in much the same way everywhere. In 1978 the pair published the Facial Action Coding System, an anatomically based method for describing every visible movement of the face. For engineers, the combination was ideal: a short list of categories to predict and a precise vocabulary for the evidence. There was nothing careless about building on it, since the theory rested on genuine findings that hundreds of later studies broadly repeated.

That consensus was tested in 2019 by a review in the journal Psychological Science in the Public Interest. Its five authors, led by the psychologist Lisa Feldman Barrett, spent two years on the task and, according to press reports, considered more than 1,000 studies. They asked whether a facial movement occurs reliably when a given emotion is felt, and whether it is specific to that emotion. On both counts the evidence was weak. Barrett reported that, on average, people scowl less than 30 per cent of the time when they are angry, so most anger produces no scowl at all. The same movement can also signal different states: a smile may express submission rather than happiness, and among the Trobriand Islanders the face Western researchers linked with fear was read as a sign that someone intended to attack. Observers matched faces to emotions consistently when handed a list of words, but agreed far less when they had to produce a label of their own.

It would be a mistake to read this review as proof that faces say nothing about feelings. Its claim is narrower: the link between a movement and an emotion depends heavily on the situation, the individual and the culture. That narrower claim, however, is exactly what undermines most commercial systems. Detecting that the corners of a mouth have risen is a measurement problem, and software can tackle it. Concluding that the person is pleased is an inference, and that step is where the evidence runs out. A system trained to treat every smile as happiness will learn that assumption and repeat it with mechanical confidence.

Some companies appear to have reached a similar view. In 2019 the Electronic Privacy Information Center, a US non-profit, complained to the Federal Trade Commission about HireVue, whose software assessed video interviews. In January 2021 HireVue said it had already stopped using facial analysis, with its chief executive explaining that non-verbal data had added little predictive value. The following year Microsoft announced that it would retire features of its face service that 'purport to infer emotional states', citing the lack of agreement on how emotions should be defined and the difficulty of generalising from expressions across regions and groups of people. It kept the capability only for limited accessibility tools.

Not every use deserves the same suspicion, in my view. A car that warns a driver who is falling asleep is responding to a physical condition, and its warnings can be judged against what then happens on the road. That is a narrower and far more testable task than deciding whether a job applicant is enthusiastic or a pupil is bored.

Regulators in Europe have drawn a firmer line. The European Union's AI Act, which entered into force on 1 August 2024, has since 2 February 2025 prohibited AI systems that infer people's emotions in workplaces and educational institutions, unless the purpose is medical or concerns safety. That restriction seems to me justified, because employees and students are rarely in a position to refuse to be observed, and a mistaken reading of a face could cost someone a job or a grade. A blanket ban, though, would go too far. The ambition Picard set out in the 1990s, of machines that take human feeling into account, is still worth pursuing. What deserves to be abandoned is the shortcut of treating a face as a dial that displays an inner state to anyone with a camera.

## Questions

### Questions 27-31: Multiple choice

Choose the correct letter, A, B, C or D.

27. According to the writer, what did Picard argue in the 1990s?
   - A. Emotion matters less to intelligence than scientists had once believed.
   - B. Computers should be designed to experience emotions of their own.
   - C. Machines that work closely with people might need to detect and react to feelings.
   - D. Computers would soon be capable of reading their users' minds.

28. Why does the writer mention the European Union's rules on new motor vehicles?
   - A. to explain one reason why the market Smart Eye works in has grown
   - B. to show that cars must now be able to identify their drivers' emotions
   - C. to suggest that Smart Eye paid too high a price for Affectiva
   - D. to contrast the rules for cars with the rules for recruitment

29. According to the writer, why did basic emotions theory suit engineers so well?
   - A. It had been confirmed among people with little contact with the outside world.
   - B. It had already been tested in hundreds of computer-based studies.
   - C. It showed that one facial movement could reveal several emotions.
   - D. It offered a limited set of outcomes and an exact way of describing faces.

30. In the fifth paragraph, the writer is chiefly arguing that
   - A. measuring the movements of the face is still beyond current software.
   - B. the weakness of emotion AI lies in moving from a facial movement to a judgement about feeling.
   - C. facial movements provide no useful information about emotions.
   - D. the 2019 review exaggerated the influence of culture on expressions.

31. What did Microsoft give as reasons for its 2022 decision?
   - A. A complaint about the service had been sent to a US regulator.
   - B. Non-verbal data had added little to its predictions.
   - C. Emotions lacked an agreed definition, and expressions did not generalise well across populations.
   - D. Its accessibility tools had been used for purposes it had not intended.

### Questions 32-36: Yes / No / Not Given

Do the following statements agree with the views of the writer? Choose YES if the statement agrees with the views of the writer, NO if the statement contradicts the views of the writer, or NOT GIVEN if it is impossible to say what the writer thinks about this.

32. The products of the emotion AI industry generally stay within the limits of what the early research proposed.

33. Those who built systems on basic emotions theory had reasonable grounds for doing so.

34. Systems that warn tired drivers can be checked more easily than those that judge a person's feelings.

35. Other governments are likely to follow the European Union's approach to emotion recognition.

36. Emotion recognition should be prohibited in every setting.

### Questions 37-40: Summary completion (word list)

Complete the summary using the list of words, A-J, below.

- **A** reliably
- **B** rarely
- **C** minority
- **D** majority
- **E** deference
- **F** aggression
- **G** generate
- **H** translate
- **I** suddenly
- **J** surprise

A review published in 2019 asked whether a particular expression shows up **(37)** _____ whenever someone feels a particular emotion, and whether it is tied to that emotion alone. It found little support for either idea. People who are angry, for example, scowl in only a **(38)** _____ of cases, and a smile may indicate **(39)** _____ instead of happiness. In addition, observers agreed much less often when they were asked to **(40)** _____ a label themselves rather than select one from a list.

[Practice online with a timer and instant marking](https://testmigo.com/ielts/reading/can-machines-read-human-emotions) - [Answers and explanations](https://testmigo.com/ielts/reading/can-machines-read-human-emotions/answers)