Large Language Models and Pareidolia

Have you ever looked up at the sky and seen a face staring back at you from the clouds? Of course you have; you're human. Our delicious meaty brains are hardwired to recognise certain shapes - and faces are a useful shape to recognise. A few false positives are a worthwhile trade-off for such a powerful feature.

Mistakenly seeing faces where there are none is a phenomenon called pareidolia. If you've ever used facial recognition on a computer, you'll know that machines also suffer from it.

I was using an AI tool to scan all my photos. I wanted it to recognise all the human faces so that I could tag my photos with my friends' names. One of the photos it presented for tagging was this:

Are those faces? Undoubtedly yes! Is this a mistake that a human would have made? Absolutely not!

But the above is a mistake generated by Machine Learning, not by Our-Lord-And-Saviour Large Language Models. Surely a language model doesn't suffer from this?

Because Google has no faith in its ability to launch new products, it has forcibly shoved AI into all of its services. There's no way to turn it off. You will use Gemini and you will like it.

At the time of writing, here's what happens if you ask Google "How many i's in teamwork?"

It's easy to see how Google's LLM has gotten this so catastrophically wrong. There are dozens of articles where some business guru ineffectually tries to argue that there is an "I" in team actually. So the statistical model inside the LLM gives weight to that.

Similarly, there are lots of silly articles proclaiming that the I in team is in the A-hole. But LLMs do not understand satire:

Google and Meta search both report that Cape Breton Island has its own time zone 12 minutes ahead of mainland Nova Scotia time because they are both drawing that information from a Beaverton article I wrote in 2024

— Janel Comeau 🍁 (@verybadllama.bsky.social) 2025-06-10T00:50:07.217Z

LLMs are hardwired to regurgitate text which statistically matches what they've seen before. Their makers believe that a few false-positives are an acceptable error rate for such a useful feature. The LLM form of pareidolia is to recognise text as being syntactically and linguistically correct, even though the contents are rubbish. This is an inherent feature of LLMs. No amount of manually tweaking their parameters or prompts can fix this.

At the moment, Artificial Intelligence - whether Machine Learning or Large Language Models - only works well on a narrowly defined set of tasks and with humans checking the output.

Imagine you've just hired an intern. They've graduated top of their class from the best university and, apparently, excel at what they do. Because you're the boss and they're the intern, you ask them to make you a mug of tea. White, no sugar.

They return with the teabag still in the mug. OK, not everyone knows the intricacies of how to serve tea.

The tea tastes funny. You ask them if they sniffed the milk. "Milk? I used Tipp-Ex to make it white!"

At which point, after throwing up, you throw them out.

Most people encountering Gemini's repeated and unacceptable failures will decide, perhaps rightly, that AI isn't even close to good enough yet.