Semi Sentient Sentence Engineer

Sometimes it's better to listen to people from the International Conference on Artificial Neural Networks (ICANN) than those from your newsfeed. Especially when your newsfeed is algorithmically engineered to reinforce billions of dollars of investment in a half-baked invention. Because for so many people, no matter who you are, whether you're the expert, or the outlet promoting said expert, there is so much hype around "technology" right now that you'd think we could pretty much replace every invention ever created with a few lines of code, upload ourselves into the anthroposphere, and watch our cellular clocks tick tock into infinity. Just about everyone is so ready to extoll the virtues of a super-sentient human alternative without checking to see whether the alternative actually exists. (Too late, we've already invested your retirement fund into it, so now you have no social security AND no retirement!)

The presenter here, Bernhard Schölkopf, works with computers and code, and tries to make easier the things that are hard for computers to do. Sometimes he uses to an advantage the things that computers mess up. This talk seems to be about "spurious correlations", and how they can be useful. For example, you can see how his team manages to get the details of the surface of the moon during a lunar eclipse. It's hard to convey how bonkers this is in words - likely the most intense contrast of light vs dark that can exist for a human, when the midday sun is blotted out by the moon, making the moon, by comparison and as seen by our photochemical receptors, the darkest thing to ever exist, and they have managed to pull out the details of the moon's surface with no less detail that you would see with the naked eye.  

Above image credit: Please excuse my simple language, but they de-noised, or whatever you call the opposite of de-noising, the image of the moon as seen during a lunar eclipse, and left only the details of the moon, as seen here. Kind of like hearing someone whisper you name from across a football field, during the Superbowl, right after a touchdown, while wearing headphones and listening to electronic dance music. ICANN 2025 - Causal Representations, World Models and Digital Twins - Bernhard Schölkopf [link] 

This is the The Ebbinghaus Illusion, where the two orange circles are the same size but they look different because of the blue circles that surround them. ICANN 2025 - Causal Representations, World Models and Digital Twins - Bernhard Schölkopf [link] 

But in setting up his talk, he first mentions, almost in passing, this interesting phenomenon that I have definitely never heard anyone else talk about. Maybe because I'm not spending enough time at computer science conferences. 

He shows us the The Ebbinghaus Illusion, where the two orange circles are the same size but they look different because of the blue circles that surround them. I thought that was called the Oppenheimer Effect, but who cares what I think. It's an optical illusion. You probably know the definition of an optical illusion, and how they "work" and what they look like. Computers, on the other hand, do not. And so, when querying an LLM with an exaggerated but incorrect version of the illusion, it betrays its non-human nature, and tells us this new version of truth: "This image is a variant of the Ebbinghaus illusion. Although the circle on the right appears larger due to being surrounded by smaller blue circles, both orange circles are actually the same size. The arrangement ... tricks our perception. ...". 

Excuse me, "Our" perception?

This is the The Ebbinghaus Illusion, as seen by a robot (OpenAI in this case), and by the description given, robots seem to have been confused, by us, its teachers, into seeing the wrong thing. "This image is a variant of the Ebbinghaus illusion. Although the circle on the right appears larger due to being surrounded by smaller blue circles, both orange circles are actually the same size. The arrangement ... tricks our perception. ..." ICANN 2025 - Causal Representations, World Models and Digital Twins - Bernhard Schölkopf [link] 

This is a different optical illusion that the LLM also gets wrong: "Although the two segments of the horizontal line appear to be of different lengths due to the arrows at the ends, they are actually equal." ICANN 2025 - Causal Representations, World Models and Digital Twins - Bernhard Schölkopf [link] 

And here's another one. The text in the above image, generated by an LLM when prompted with an exaggerated version of the illusion, says, "Although the two segments of the horizontal line appear to be of different lengths due to the arrows at the ends, they are actually equal."

Actually...

Anyway, this is one of those immediate, obvious examples that you would expect to have seen or heard of already. But not in this mediasphere, which is, unfortunately, trying to sell us this technology. And to find out what's wrong with it, or how it can destroy your business, ruin your life, or just give you really bad dating advice, you have to watch talks that you barely understand, given by people who already have a decent job as a scientist so they're not trying to sell you the tech; they're just showing you how they use it. And that usually means figuring out not how it works, but how it doesn't work, and then using that part to make it do cool things you didn't expect it to. That's what this guy did. What he also did was reaffirm my suspicion that some of the changes made to your work by a neural net, be it for language or images or music or whatever, are likely to be imperceptible to you.

These obviously wrong optical illusion descriptions are an example of how it messes up, but it shows us something deeper about how training works, and how it doesn't work. It shows us the gap between a thinking computer and a human being. The brains behind this thing are so alien to us that we are likely not going to see what it's doing wrong unless we're intentionally studying its output.