Is AI hallucinated or is it part of the process?

I used to be fascinated by AI, but after a while it seemed to hit a brick wall - since the output is not always correct, how can we trust it?

We know of many startups that have raised money on false promises - it is enough to say you use AI, have a killer presentation, and you can get millions in investment. After a while, they turn out to be garbage - they either hired hundreds of indians to do the real work, or don’t actually have a meaningful product, just the AI. With a future version of ChatGPT or an alternative AI, their product is obsolete. It begs to question how much we can trust the AI.

Then there are people who overuse AI for everything. Been there. Writing, code, images, essays, homework, therapy (maybe really avoid that). This could be an entire topic of it’s own, but that is not the point of the current article. While AI is capable, it has limits. It could be good for math related or text related tasks, and even then it makes up many things. But for human relationships, probably not the best tool.

Given many made up responses, the question is, what about the hallucinated parts? We have Sam Altman promoting his AI by saying AGI is near, and on the other side we have the real responses that are made up, so called “hallucinated”. So where is the truth? Probably somewhere in the middle.

Technology has so far appeared mostly truthful. Computers can either be right or wrong. With the addition of social media, things have changed: people can lie on the internet as well, as in the real world.

AI will give out answers based on most probable outcomes - that is how it is build. By design. It’s not a single source of truth. Or better, not an absolute truth. But truth-like? It will give out a reply that it thinks is the closes answer. While some other AI-s like Deepseek or Grok, might be manipulated to give replies based on what the Chinese Goverment or Elon Musk wants you to think.

AI is basically a huge collection of data, and many models on top of that data, that do the guesswork of what would be the best answer to your question. It is constantly evolving, so it’s probably never going to be complete. We are using that in situations and jobs, where it can outperform humans - because humans are also not perfect and make mistakes.

So is it hallucinated? Well, yes and no. The so-called hallucination is for filling in the blanks. If it doesn’t know the correct answer, it will show what it thinks is statistically most correct. And in some cases, it is straight out bullshitting. So we have to be aware of what we ask and how truthful the responses. So yeah, handle with care, and caution.

Will it reach AGI or superintelligence? Nobody really knows. There are many guesses, and as things improve, we can only watch it progress. And maybe use it to our advantage, taken if we can trust its responses enough for the tasks we want it to perform.

So does this conclusion help us when considering our future AI use cases? I definitely feel more relieved and hopeful. At first I expected AI to be able to do everything. But after the initial euphoria, I came to a sudden and somewhat disappointing realization - it’s a mirage, smoke and mirrors. That it’s not that good as advertised. And sure, it’s not, but it’s pretty powerful, always evolving and getting better by day. We just have to know it’s not perfect and that it can make mistakes, and occasionally hallucinate. Maybe we need to focus on asking better questions. So it’s useful to keep that in mind next time you ask it, what is the meaning of life, and getting replied with 42.1