Raising an eyebrow at physiognomy
The famous philosopher Aristotle claimed it was possible to infer character from features, a concept which became known as physiognomy and came to fame in the 19th century as a belief that potential criminals could be identified by facial features.
Physiognomy has been widely discredited, yet still it keeps recurring, and these days it is dressed up with Artificial Intelligence and the power of neural networks. A couple of famous studies conducted in 2017 claimed that AI could be used to correctly predict if someone was gay or straight, just by studying a photo of the person's face, with accuracies of up to 77%. Whilst that isn't stunningly accurate, it was significantly better than random guesswork, and also better than humans were achieving with the same data.
The study couldn't offer any reason why it would be appropriate to use such a tool to identify someone's sexuality, but at least it wasn't predicting someone's criminality, (unless you are unfortunate enough to live in a nation where homosexuality is considered a crime). However, other researchers are more ambitious, and are re-inventing the pseudoscience of the 19th century physiognomists in the process. For example, researchers at Shanghai University used AI to analyse some 2,000 photos of chinese citizens, of whom around one third were convicted criminals, and concluded AI was capable of identifying potential criminals from facial features 89.51% of the time, (a ridiculously precise number that sounds more significant than it is). Features which they claim are good for predicting if someone is predisposed to criminality are "the curve of the lip, the distance between the corners of the inner part of the eye, and the nose-mouth angle".
The researchers argue that although they don't know the scientific basis for the AI's predictions, they can nevertheless be trusted because they are compiled by an unbiased self-learning algorithm which doesn't come loaded with the inevitable human prejudices and arbitrary judgements such as "shifty eyes" or a distrust of men with beards. This illustrates a danger of the world's infatuation with technology in general, and with the current love affair with AI in particular. We tend to be too trusting of AI, especially when it reinforces our own prejudices, even when any scientific basis or controlled testing of the system is lacking. This unwarranted trust is, in no small part, because the name includes the word "intelligence".
The potential to misuse such technology is huge. Imagine, for instance, that states used it as part of the process to decide whether or not to issue visas, or that employers used it to weed out people who were deemed to have criminal indicators as part of their recruitment processes, or even people who's facial features indicated, according to AI, that they were not cut out to be a plumber, or a programmer, or a postman. Imagine the injustices which would be caused if well-meaning people with no biology or psychology expertise developed this technology to identify the faces of potential paedophiles, or potential terrorists,... and yet researchers around the world are scraping photos off websites and feeding them into AI, believing it can somehow do all of these things.
Back to the sexuality predictor, a Google engineer, Blaise Arcas, has taken issue with the results of those AI studies which used thousands of photos scraped from dating profiles. Arcas argues that the AI software was picking up on superficial features, trends, and fashion styles, rather than anatomical details. For example, he argued that in the sample dataset, it just so happened that straight men were more likely to be wearing glasses than gay men, but you obviously cannot extrapolate that to be a global truth. Other possibilities was that the software was picking up on fleeting fashion features or use of make-up. Arcas demonstrated this potential failing by repeating the experiment with all of the anatomical details blurred out and achieving comparable results.
Spurious correlations are well known to scientists and mathematicians. If you take enough datasets and look for matching patterns, you will always find things which, just by chance, appear to go hand in hand. Plot a graph of the two factors and it will look like they are intimately related. Military Intelligence analyst Tyler Vigen famously scoured public demographic datasets for the period 1999 to 2009 and found hundreds of hilarious correlations which include such morbid gems as:
* Apple iPhone sales correlates with the number of people who died by falling down the stairs
* The per capita consumption of beef in the USA correlates with the number of deaths caused by lightning
* The divorce rate in Maine correlates with per capita consumption of margarine
* The number of people who died by falling out of their bed correlates with the number of lawyers in Puerto Rico
* The age of Miss America correlates with the number of murders by steam and hot objects.
* The number of people who drowned by falling into a pool correlates with the number of films Nicholas Cage appeared in.
These are so ridiculous that we know its an illusion, a chance occurence, but remember, AI researchers say that AI systems examine huge amounts of data without any preconceptions or bias, and can spot relationships which would be invisible to the human eye.
Arcas has highlighted the need for the application of ethics amongst AI researchers. A similar message came from Steve Grobman, of the security company McAfee, when speaking at this year's RSA conference in San Francisco. He warned that AI has no moral compass. "We can't allow fear to impede our progress, but it is how we manage the innovation that is the real story. We must embrace AI, but never forget the limitations. It is just maths."
Trivia footnote: You may have heard or used the term "fizzog", used in Britain as slang for "face". It it thought that fizzog is a corruption of the word physiognomy.
28th March 2019
This article comes from the SKILLZONE email newsletter, published monthly since January 2008, and covering topics related to technology and the internet. All articles and artwork in the SKILLZONE newsletter are orignal content. If you would like to receive the newsletter direct to your inbox each month, please SUBSCRIBE here. It is free, and you don't get added to any other mailing lists. It uses best-practice confirmed opt-in only, and you may unsubscribe at any time.