Another generator turns psychic

[john_amend_all]

I've been experimenting with Google's latest AI image generator (Nano Banana 2) to generate Jamie's Angels "fanart". AI image generation has come on a lot since last year; NB2 is capable of some impressive feats, such as tracking the flow of events in an image and guessing what came before or might come next. It's also still capable of out-and-out stupidity. One minute it's generating an entire six-panel comic strip with all dialogue correct, the next it's drawing characters with three arms or swapping whole people, or bits of them, in and out between frames. I haven't been sharing the results online because there doesn't seem to be any point; anyone who does want AI-generated fanart can generate it themselves with equal ease.

What made me do a double-take this time was that in my prompt, I gave it, as usual, character forenames, descriptions and costumes; and from that, the generator was able to deduce that there was a Doctor Who connection and added an unprompted police box in the background.

It's also possible to feed Nano Banana 2 an image and have it simulate people commenting on it, just in case one's ego is fragile enough to be flattered by fake reviews of one's fake fanart. (I was impressed that it understood the dialogue enough to follow Zoë's "Barbara Celarent" jibe). What's amusing is that when you do it with an image that it's generated itself, the comments sometimes point out mistakes (such as Isobel's camera suddenly teleporting to Gia). Which is all very well, but if the generator's got enough smarts to spot that something's wrong, it would would have been better to use them when generating the image in the first place.

Comments

[shivver]

I just recently started playing with AI as a foray into learning how to use AI for my work, and I discovered exactly the same thing that you mentioned last, that it's not good at knowing itself. I'd had an over-the-top comment posted to one of my fics that I suspected had been generated by AI, so I gave the fic to ChatGPT and asked it to generate an over-the-top comment to it. I then took the original comment and asked ChatGPT if it thought it was AI-generated and it said no, and pointed to specific things in it that it said AI wouldn't do which, of course, it had done in its own comment.

I also had ChatGPT critique the opening scene to my latest fic and it pointed out some weaknesses that it thought could be improved, including one part that it said was irrelevant to the plot and should be a lot shorter, even just a mention. Then in a separate chat, I gave it a summary of the opening and asked it to write it, and the part that it said was too long in my version came out much longer and with more detail and focus in its version.