I don’t think this project was well thought out. The project seems to claim that because vision systems strip away the context of the scene to zone in on the face, it’s dehumanizing and therefore problematic that we’re being “watched over” by machines all the time via consumer systems such as FaceID, snapchat filters, etc. I completely agree that there are important ethical implications that should be considered, but are often overlooked when it comes to facial recognition systems. But to what extent does this project actually address these issues? I personally don’t mind being simply “watched over” by machines, but I do mind what companies choose to do with my data. If it’s just to give me personalized ads, I’m fine with that. But if they’re using my data to build discriminatory automated hiring systems (a system built by Amazon would reject resumes containing “society of women engineers” and accept resumes containing “lacrosse”), I would have a problem with that. I don’t think it’s the bounding boxes on faces that are dehumanizing; it’s the discriminatory policies informed by biased systems that are dehumanizing.

A more meaningful direction this project could have gone in is exploring who is and is not ‘seen’ by computer vision systems. Because the faces of women and minorities were once severely underrepresented in datasets, early vision algorithms wouldn’t recognize or misclassify anyone who wasn’t a white man, especially black women, at much higher rates. I wonder why the artist chose to use old photos depicting practically only white people; it seems more like a reflection of the (white) artist’s bias than a deliberate attempt to criticize vision dataset bias.

As someone who’s spent the last year working with a computer vision research group, I think this project’s idea of machines being some kind of all seeing overlord misrepresents the field. It communicates to laypeople that vision algorithms are inherently evil (it’s literally just math) without putting it in the important context of human and data bias. It perpetuates the layperson’s exaggerated fears of an AI apocalypse, which is definitely not happening anytime soon because AI is nowhere near capable of independent thought. AI can make predictions and find patterns, but humans are ultimately still the ones who decide when and how AI should be used.

I discussed this project a bit with a friend who’s also familiar with computer vision methods, and she brought up the interesting point that it’s not necessarily a bad thing for vision algorithms to be reductionist. In fact, basically all problem solving methods have to be reductionist — otherwise there are way too many parameters that have to be considered and the problem can’t be reasonably solved. Conversely, it’s pretty incredible how a task as complex as vision that takes hundreds of millions of human neurons can be modeled by a surprisingly simple computer algorithm. I disagree with the project’s claim that reductionist algorithms imply that our bodies are unwanted or dehumanized.

Computer vision systems have also long surpassed simply extracting a box containing a face, which is what this project oversimplifies computer vision as. There are tons of papers out there describing systems that take the entire scene into context, many of which include heatmaps visualizing how the algorithm weights each aspect of the scene to make its final decision/output. My UROP group’s broader goal is to make AI more explainable and controllable, so that humans ultimately get the say in how AI should work and address algorithmic biases directly. I wish the person behind this project read up on more current vision literature or consulted an active researcher in this field before publishing their project.