Shiveen Pandita

AI Is Retrocausal

I was recently talking to a friend about the uncanny valley effect of reading AI generated prose or images. Most of us have felt it, you can almost always instantly tell when something is too obviously AI generated. The prose is too crisp, the images are a bit too sharp, the argument feels like its rushing to prove a point conclusively. I didn’t really have a good response to this feeling until a few days ago.

We all know it’s most likely because these models have been trained and then system prompted to be as helpful as possible and take things to their logical conclusion. However, I believe the reason for that is most likely that AI was never trained in the abstract human middle thought patterns and only on the combined output of finished artifacts that most model companies simply stole from the internet.

AI was never trained in the abstract human middle thought patterns and only on the combined output of finished outputs

Let's try to run a thought experiment in real life. Picture two students given the same maths problem. The first has the answer key. They read the final answer, then write working that leads neatly to it. Every step is correct and every step points straight at the destination. The second has no answer key. They try an approach, hit a dead end, back up, notice something odd, and drunk walk their way to the answer. Their working is messier, but you can see a mind moving through it. There are "traces" so to speak of interim thinking which then peeks out from the finished product.

That example brings me to my main critique of any piece of information (textual, visual, memetic) artefact produced by AI: everything it does is retro-causal. The models spit out a finished product and then try to form the narrative (or lets say the thinking trace behind it) to fit that finished product. A common counterargument is that LLMs output things token by token, start to finish, with reasoning traces. In my opinion it’s not the same. A human mind works through a messy, abstract middle before anything resembling an answer exists. An LLM skips that middle. It jumps between versions of finished products, then presents a reasoning trace that looks like it led there.

Or in other words, unlike a curious student or an artist or an engineer, that might go through interstitial thinking states to land at perhaps imperfect but truly novel and elegant idea, AI optimized to land at an idea and then pretend the previous thinking was absolutely aligned to this final output. This is why when you can tell it’s AI, it doesn’t have the indelible watermark of small inevitable but not deliberate imperfections that we humans can detect.