Bear in mind when you ask an LLM to explain why it did what it did (or didn't), you're asking faces on toast to explain faces on toast.
It's not explaining. It's plausibly completing.
And you fell for it twice.
Aug 7, 2026, 12:15 PM:1000:
It's generating plausible fanfic about an imaginary AI that actually had reasons for generating the previous plausible fanfic about whatever you prompted it with.
Aug 9, 2026, 10:14 AM@jasongorman I am going to remember “faces on toast” for a long while 💅
Absolutely this.
Although in fairness, human beings are also notoriously bad at retroactively constructing false narratives to explain their past choices and behaviors.
@jasongorman In my case it kept gaslighting me...
I don't quite understand this post, but if it means that LLMs are always wrong, then the post is wrong
In reality, LLMs are incredibly powerful tools that when used right can do awesome things
@jasongorman Pareidolia? I learned that word on mastodon!
Where I currently work they think asking the chatbot to provide references means that we can trust the output because it has references.
@jasongorman remarkably similar to humans then.
@jasongorman One of the major eye-openers for me when it comes to human cognition was when reading Oliver Sack's fantastic "The man who mistook his wife for a hat" and the research done on split-brain patients.
It turns out that not only are you correct about LLMs, but humans do exactly the same. We will happily make up whatever plausable story explains our behavior after the fact - fully believing it to be true ourselves.
https://en.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife_for_a_Hat
@troed Hey, we've all done it
@jasongorman I originally read that as “faeces on toast”. Your version makes perfect sense, but I think it still works either way…
I once (in earnest) got the great suggestion of getting the LLM to grade it's confidence in the answers it was giving. Just append a "include a confidence score of 0-100% for each answer you give"
It did indeed include them (80% confidence)
Buffoons.
Aug 11, 2026, 5:23 AM@jasongorman if we burn enough toast eventually a slice will look kind of like jesus, and that’s very scary and also the future of all toast and you need to pay me to see it or be left behind
@zzt toastmaxxing
Aug 11, 2026, 3:25 PM@jasongorman Thanks. Completing was correct.
@kktk Completing was a pattern learned from the training data.
@jasongorman I read this while eating my Vegemite. Thanks