UPDATE 2026: This article was originally written in 2023. In just 3 years the models are now good enough to get text right a lot of the time. In that time, the training and usage of these AI models, whether the corporate offering or the self-hosted ones, an immense amount of resources have been expended. The result? AI still hallucinates. It still confidently fabricates information. Ai is another crypto or NFT; it relies on hype to exist. Hundreds of billions of dollars were spent on them, and 99% of NFTs are now worthless. Most crypto currencies are scams and vehicles for laundering money. AI is just another such technology, except it’s have even more dire real-world impacts.
You know the weird almost-letters that generative images (AI) have in them instead of actual text? Yeah, that’s hext. Don’t bother commenting that some AI, like Adobe’s Firefly does text really well, it only does when you’re trying to apply a style to existing typefaces; not the same thing.
Hext because hallucination text, but also because hex text. Hallucinating is a term used in the AI community referring to the point when AI loses adherence to the prompt given, usually producing horrific or weird images. The hext does often look like the mad writings of something designed to curse or hex. It looks like what I imagine people who use Theban script think their “spells” and shit look like to other people.
ChatGPT et. al.
I’m also not talking about the responses of generative text AIs after they appear to lose the plot of your thread of “conversation” with them. That’s a form of hallucination or short-term memory loss that AIs like ChatGPT experience when you continue an interaction beyond their normal buffer.
Okay, some examples:
Why though?
Hext comes about because AI doesn’t actually know what it’s generating. That’s one of the main arguments anti-AI folks are making in copyright-related issues, and they’re right. In fact, when a word appears, like the word “not” in one of the examples above, it could be because such a prevalent word is well-trained into the model being used, but more often than not in current generations of models, it’s a fluke.
