Now That's What I Call Bullshit™
Welcome to the blog! Plus, an important rule of thumb for sniffing out AI bullshit.
Something about this particular ad — this particular clusterfuck of marketing nonsense composited over stock imagery to advertise targeted products — inspired me to start this blog:
Not because of any of the many ways this could be problematic, but because it’s so shameless it’s funny.
I first learned about Perfect Corp.’s “AI Personality Finder” via this story in The Information. It’s a technology that “promises to read your facial features for clues about what kind of person you are—and, by extension, what kind of products you might buy.” Good — as if advertisers aren’t already doing enough to sell us stuff. Per The Information, it works as follows:
I snapped a selfie and ran it through the online demo. Five seconds later it spit out a profile based purely on the arrangement of my face. According to Personality Finder’s algorithm, the distance between my nose and my mouth, combined with my rounded cheekbones, hooded eyes, and other facial attributes, identified me as an enthusiastic, action-oriented and social person. I scored 95% for extraversion, 21% for neuroticism and 63% for conscientiousness.
You could be forgiven for buying into this. After all, Perfect Corp. isn’t some know-nothing company — it’s backed by Goldman Sachs, Alibaba, Snap, and Chanel. The product itself was covered in several different outlets, including Yahoo News, which republished a glowing press release about Personality Finder, highlighting its use of “psychological big data.”
None of which changes the fact that this is snake oil.
*Freeze frame, record scratch*
How did we get here?
I’m Maxime, and I’m an AI researcher. I’ve worked at J.P. Morgan AI Research, the world’s leading financial AI research lab, and Berkeley Artificial Intelligence Research, one of the world’s foremost AI research groups.
Now, what’s this newsletter about? I firmly believe that while AI seems like distant magic to most people, it’s going to affect us all. I’m on a mission to make AI accessible to everyone, so that you can make informed decisions about what’s real and what’s not, what excites you and worries you, and what you can use to make your own life brighter.
Whether you’re a business owner or just a curious person interested in learning about the field, I’m here to help you contend with the hopes, worries, and ambiguities created by the rise of AI.
Sometimes, that means helping you to calibrate your Bullshit-o-meter. Please subscribe to join me on this journey!
Rule of Thumb: No New Information
AI is not magic. This is the part I think reporters and investors, in their well-intentioned or profit-driven optimism, sometimes forget. AI systems can compress, shuffle, and transform input data in all sorts of ways, but they can't create information out of thin air.1 The information has to exist somewhere in the training data.
In the case of personality detection, it doesn't. My next piece is a deep dive into the literature around this issue (scout’s honor), but statistics aside, humans are famously bad at reading one another. Read Malcolm Gladwell’s Talking to Strangers for an in-depth treatment of that topic — or just recall the many times you’ve likely misjudged someone or been misjudged yourself.
Case in point: the two separate people at a college party who told me, “I see you around all the time but I never say hi because you always look angry!” Granted, this used to be my happy face:
Takeaway: Snake Oil and AI
In many ways, AI as an industry has made itself vulnerable to bullshit. The people who really understand the technology know well that most people don’t. Most researchers can’t even explain what their models are doing — that’s the entire point of using AI. If the problem could be solved with manually written, human-interpretable rules, it would be.
As such, the industry relies on a willingness to believe: an optimism in the face of ignorance about the magic sauce that sits between the input and the output. End users verify these technologies in a “proof is in the pudding” style: if the behaviors exhibited by the system align with intuition, users assume they’re deeply sound.
You might well say, “Yes of course — that’s the only way to verify such systems. You’re just describing empirical validation.”
I agree. My point is that the AI industry has normalized this willingness to believe among its users — so when methods come along that actually don’t work, people still give them the benefit of the doubt. They assume that even if the results seem implausible, there must be something in the magic sauce, some science beyond their comprehension, making it all make sense.
But what makes “proof is in the pudding” or “gut feeling” verification work is that most of the time, AI is reproducing capabilities humans already possess. It’s easy to check whether a model is detecting cats and dogs correctly, because we’re good at detecting cats and dogs.
But when AI promises to do something humans can’t — when it promises that something in its magic sauce now allows it to pick up on cues beyond human comprehension — that’s where the trouble lies. “Proof is in the pudding” verification fails when there is no reasonable expectation for how the model should behave.
In turn, this does wonders for marketers, who are free to write their own explanations for what happens in the black box and peddle these stories to users happy and accustomed to taking for granted that AI works.
The Verdict
No, I don’t think this model can predict your personality just by looking at your face. If it could, I imagine it would be used for Bad Things™. After all, human attempts to assume what kind of person someone is based on their physical appearance have, historically… Not ended well.2
I hope this kind of high-tech palm reading never becomes a thing. If it does, I can’t wait to get rejected from my next job for “not looking like enough of an extrovert in my profile picture.”
And to marketers: for God’s sake, stop coming up with new ways to sell us stuff.
See “racism” for more information.
The photo captions and footnotes are killing me, please never stop.
Really dissappointed at how this wave is being ridden. Dissappointed and sad