Select Page

Almost all developments in the world of artificial intelligence meet some degree of backlash, some immediate, and others just contributing to the “millions of blue and white collar jobs will be displaced by AI”-predicament looming in the not-so-distant future. However, a recent breakthrough ushers in a new era of ethical and political concern.

 

Link to study: https://osf.io/fk3xr/

I guess “Breakthrough” wouldn’t exactly describe the study co-authored by Yilun Wang and Michal Kosinski from Stanford University, as they used deep neural networks, a type of widespread machine learning technique. The subject matter itself stands out from the study, which has the title: ‘Deep neural networks are more accurate than humans at detecting sexual orientation from facial images’. Essentially, this team trained a neural network using 35,326 images of faces (each tagged with the self-identified sexual orientation) from a dating website to see if it can correctly guess sexual orientation on additional pictures of faces. Surprisingly, this resulted in an algorithm that could distinguish between heterosexual and homosexual individuals just from five images of a person’s face, with roughly 91% accuracy for men and 83% for women. Essentially the study points to the existence of a correlation between sexual characteristics and certain facial features that, though our brains miss, are caught by an algorithm. As a control, human judges had much lower accuracy, sitting around 61% for correctly identifying a homosexual man, and 54% for correctly identifying a homosexual woman, which more or less amounts to a coin toss.

I have no intentions of commenting on the implications of this correlation and instead plan to focus on the implications of the algorithm’s existence. There still exist countries where public knowledge of one’s homosexuality essentially equates to a death sentence. What would happen if policymakers in these countries began to use this software? Not only does the academic nature of this report mean that the code is likely also published, its existence means that anyone with a working knowledge in neural networks can recreate it. This all results in a grave threat to privacy if implemented maliciously, and could create unprecedented scenarios in the judicial systems of these countries. The group announces that they intend to continue this research, expanding it to find potential links between physical characteristics and behavioral tendencies, psychological conditions, and personality traits. In the footnotes of the study: “We were really disturbed by these results and spent much time considering whether they should be made public at all. We did not want to enable the very risks that we are warning against.”

I have faith that regardless of our beliefs/preconceived notions, we can all agree algorithmic tools with super-human distinguishing abilities must not be used with hateful or selfish motives. The last thing technological innovations need to do is increase the sociopolitical friction that exists in our society.