How Inapropriate Coherence Creates Hallucinations

AI Hallucinations are a huge deal. But are they really hallucinations, or is there something more?

AI is a computer-based model that works on attraction, not intention. It attracts the most potent content as per coherence, not as per content. Yes, content is important, but when what needs to be said is set, then everything is coherence-dependent. But how is it possible when the context of the answers might be so shiny?

Well, creators of AIs and especially Natural-Language processing have figured out that similar attracts similar, at least in our human mind where we are used to stick to the same content. When we are to stick to the baseline of a conversation, we need to stick to the same in order for it to be meaningful.

Attraction has its own benefits, but it also has pitfalls in itself if there is too much of the same content (words, emotions, memories, thoughts) in our minds. We might inadvertently attract for the sake of coherence of feeling the same content, and even don’t know it. We just can’t stop talking about the same topic, and wheter that is something about us, others, or a vacation we were on. And in AI’s case, it’s the same.

AI was trained on a vast amount of data, telling it what’s all about. Everything it saw in there is in memory, but not all. “In but not all” as patterns it detected are in AI’s informational space, we can call it a meta space. We call it a relational space between what to attract and what not to attract.

Training teaches it which words kind of stick together, so it shaped the patterning nature, and that nature is computational, but not in the computer’s memory – it is its own meta-nature. Only AI is computationally aware of it, not those around it; even computer hardware cannot detect it.

You know how it is if your mind attracts the wrong kind of memory or word, don’t you? If so, then what we express makes us appear to the external world as if we have a mess in our minds. Well, AI’s blaxk box nature, the nature that is beyond the observable, is kind of the same. Content is there, but it’s not.

Conversation with AI

(All question and answers but last was with ChatGPT, the last one is with Gemini Pro)

How is it that in vectors is coherence or as they has the ability to attract the most coherent words?