Lol. I tried doing some image generation with SOTA models. I explicitly asked it not to do something it was doing and it would literally do the thing, and straight up tell me it didn't.

Unless someone has a cognitive impairment it's just simply not a failure mode of cooperative humans. Same with hallucinations. Both humans and AI can be wrong, but a human has the ability to admit when they don't understand or know something, AI will just make it up.

I don't understand why people would ever trust anything important to something with the same failure mode as AI. It's insane.

Image generation models are usually not LLMs. Only Nano Banana Pro is capable of following negative directions like that.

Not in my experience. I asked nb to create a transparent rectangle shape and gave it RGB hex for the fill. It created the box but put the hex as text inside of it and used a checkerboard for its background. When I told it that the image wasn't transparent, it wouldn't budge!

Oh yeah, they don't know what "transparent" means. Most of them generate the Photoshop checkerboard background. They also don't know "upside-down".

There isn't much R&D going into image models and what you're getting is scraps from labs that care more about other things. NBP is the closest to a reasoning image generator we have.

Not very capable, though. It's incredibly hard to get it to fix mistakes and not double down/lie. Creating a new context window and trying again is easier - but I've not had a lot of success getting it to do the thing that isn't obvious to it.

What's incredible is that talking to them as if they're rational beings typically produces the outcome you're looking for.