What is emotional AI?
Emotional AI refers to technologies that use affective computing and artificial intelligence techniques to sense, learn about and interact with human emotional life. It is a weak form of AI in that these technologies aim to read and react to emotions through text, voice, computer vision, biometric sensing and, potentially, information about a person’s context.
While the effectiveness of current methods is highly debatable, we believe that the use of human-state measurement to engage with qualitative dimensions of human life is still in its infancy. Emotional AI, and wider automated human-state measurement, thus requires ongoing social, cultural, legal and ethical scrutiny.
How?
The following techniques are used to try to sense and discern people’s states, emotions and expressions:
Sentiment analysis of online language, emojis, images and video for evidence of moods, feelings and emotions.
Large language models: for makers of chatbots, these extend and deepen human-system interaction through emphasis on emotional language.
Facial coding of expressions: the effectiveness of this method is highly debatable, especially when based on the “big six” emotions, but it analyses faces from a camera feed, a recorded video file, a video frame stream or a photo to “infer” an emotion.
Voice analytics: includes elements such as the rate of speech, increases and decreases in pauses, and tone.
Eye-tracking: measures gaze, eye position and eye movement.
Wearables sense skin responses, muscle activity, heart activity, skin temperature, respiration and brain activity.
Gesture, behaviour and internal physiology: cameras track hands, faces, external bodily behaviour, and remote heart rate tracking.
Virtual Reality (VR) allows remote viewers to understand and feel-into what the wearer is experiencing. Headwear may also contain EEG and face muscle sensors.
Augmented Reality (AR): remote viewers can track attention, reactions and interaction with digital objects.
Machine-readable human life: OK or not?
Emotional AI promises better experience of services, devices and technologies. However, there are considerations that give cause to mistrust and question the rollout of these technologies. Citizens, researchers, policy-makers and industry should consider the following:
Overall, it is it desirable that emotions are machine-readable?
Do such technologies make sense given the ambiguous nature of emotion and subjective life?
What of different national, cultural, social and historical contexts?
What of racial, sex, gender and trans’ bias in computer vision and training data?
Will data about emotion be used in a manner that benefits citizens?
What of relationships with emotion-sensing objects?
Are protections adequate? That is, are laws and regulations appropriate (and are they being followed)?
Is the spirit of data protection appropriate? This tends to focus on identification, but is identification the principal issue?
Is it we OK that social media companies registering mental states, emotions and moods?
What of third-party uses by data brokers?
What of use-types? games are one thing, but job opportunities, border controls, education, health insurance… ?
What of uses with children?
Is law the only answer, what of design?