I thought of another test to give an LLM, and usually every test better reveals their limits. This time was no different! I asked it for a description and alt text for my photo above, and it:

• misinterpreted Kelsey as a plush toy (which is actually kinda nice),
• after I asked if it was sure it was a stuffed toy, misinterpreted Kelsey as a sculpture or possibly a fursuit *part*,
• once I'd told it the subject was a person, misinterpreted the pose very substantially, but did at least identify it as a fursuit.

I then gave it my alt text (as attached to the picture I posted, and which I wrote *before* I even thought of this little experiment) and asked for comments on the differences. I like this bit of the answer:

LLM wrote:

"My descriptions were more like a general image description. Yours is much more like actual accessibility alt text. It's focused on the subject, posture, and action. That's arguably much more useful to someone who can't see the photograph. It tells them what you're doing, not just what objects are present."

So yeah, LLM happily admits it's better to write your own alt text than burn a forest down and steal everyone's work having one generated for you.

Good to know.