My children are not AI – #3 They both struggle with reading emotions.

One day, as our little family was having dinner, my older son dropped from his chair and somehow hurt himself. Nothing serious, but you can imagine how a 5-years old boy can make a drama about it. He started yelling and crying very loudly, making dreadful faces. In front of him, my then 6-months old baby was looking at him in awe, with his big blue eyes wide open. After a few seconds, he started laughing and wriggled his little arms and legs. This of course angered my older son who thought he was mocking him. My baby then stopped laughing, seemed to think for a few seconds, and then started to cry with real tears flowing from his eyes.

Reading emotions is not innate. We start learning how to recognize basic emotions from an early age and keep learning during our whole life. Reading emotions is the first step to empathy. Like my little baby is experiencing it for the first time, to bond with other human beings, it is not only important to understand what they say, but also what they feel, and, even better, share what they feel. This is what makes us human, this is what makes us social creatures.

So what does AI have to do with it? How can we expect AI to understand what we feel? Why would it be useful? At this stage of research, AI is far from being able to feel anything human. However, in a similar way to how my little baby does it, AI uses something called Emotion Recognition Technology (ERT), to learn to read emotions. This has already been used in several areas: Is this marketing campaign effective? Are pupils attending remote classes during the pandemic really engaged? This multi-billion-dollar industry is growing at an extremely fast pace, but a number of experts and organizations are warning us about the dangers and the need to regulate its use.

First of all, ERT is still very biased. As an example, in a 2019 study using a facial recognition and analysis tool called Face++, on average black men were flagged as having negative emotions more often than white men. Today, ERT is only being trained on 6 basic human emotions (happiness, sadness, fear, surprise, disgust and anger). However, humans are far more complex than that. Joyous laugher is not the same as mocking laughter, we can cry out of sadness but also through fits of laughter. It all depends on the context and culture. We are also pretty good at faking faces. The Emojify project explains it very well in a short video and offers a couple of games to see if you can beat the machine. Knowing all these limits, we can easily understand the controversy around the usage of ERT for very sensitive purposes: using ERT to assess if this job applicant is dependable,  if this air traveler is suspicious, if this prisoner can be trusted on parole…  Science fiction? Not really, these are real use cases that have already been tested and heavily criticised.

I recently discover an application called “cry analyzer”. It claims to be able to interpret baby cries with an impressive 80% accuracy score for less than 6-months old babies. I tested it with my baby and, I must admit, I was quite impressed. But this made me wonder about the role technology is playing in the relationship parents are building with their own child. When we become parent for the first time, we are bombarded with (un)solicited advice, raising doubt and making it sometimes difficult to grow confident in our own role as a parent. If technology is now also telling us they know better, can we still experience and learn by ourselves how to communicate with our children? Could ERT change our human nature, if we start superseding our own judgement by their analysis?

With AI getting more and more ubiquitous in our daily life, we should take a step back and consider, beyond the immediate benefits, what is at stake. What can we be losing in the longer term if we let machines take the lead on fundamental human skills.