A little over a year ago we publish Here in LXA an article about hallucinations of the Artificial IntelligenceThey are a danger, if we understand this as wasted time or the fact that we may be left with false information. When we ask them something that they cannot obtain from a database quickly, they can answer anything, and that has not improved much in 2025, which is already well into the future.
It has recently become popular DeepSeek, an artificial intelligence that comes to us from China and that can offer results similar to those of ChatGPT or even improve them. But the best of these AIs is obtained if we activate the options of think more deeply, reasoning or whatever they have called it in the model in question. Otherwise, they all tend to fail at the same thing: responding quickly regardless of whether what they say is correct or not.
New hallucinations of artificial intelligence
From time to time, I spend some time playing with My Steam DeckI recently finished the base game of Borderlands The Presequel, and I didn’t quite understand the ending (spoiler alert): Lilith punches Handsome Jack… where did Lilith come from? I asked ChatGPT, and I don’t even remember what he said. I do remember, for the sake of redundancy, that I was chasing our player using his invisibility. Then I continued, Handsome Jack had hinges on his face and he couldn’t tell me it was a mask.
But the worst part was when I asked him where I had to go to play the Claptrap Voyage expansion. What do you see here, what do you see there, what… I found out that I had to go to floor 13/2 or something like that (I can't remember the name now), something we can do via fast travel.
To give me more of a basis for this article, I thought about asking ChatGPT the same thing, but with reasoning, something I also did in DeepSeek. DeepSeek got it right the first time, while ChatGPT gave me two options, one correct and the other denying that there is an expansion in The Presequel related to Claptrap. I chose that option to see if it would fix it… but no.
Do you want precision?
ChatGPT is the most popular right now, and the one most people use. If we want it to reason, we have to face two problems: the first is that it takes 20-40s to start offering the answer, and the second is that free reasoning is only possible a few times a day; when a limit is passed, it no longer works and we are invited to use a paid model. On the other hand, DeepSeek He can always reason, but sometimes he gets stuck and doesn't work.
So either we pay, or we wait, or we get lucky.
Even so, hallucinations about Artificial Intelligence are still the order of the day, and I think it is worth remembering this to prevent anyone from taking the information they give us for granted.
What if we search?
What is perhaps worthwhile is not letting the default model answer us. It is the one that fails the most, the most "cuñao", that is, the one that answers the most whether or not it knows the answer. These models give the answers about what they have learned, and if they have not learned it, they try to resolve our doubts by creating relationships that do not usually give good results. Things change if we press the search or reasoning buttons.
If we ask it to search, it will do just that. It will analyze what it finds and show us a result that we could see as a summary of a search we would do. The option to reason will prevent it from saying the first nonsense that comes to mind, although sometimes ChatGPT reasons to make things worse.
Artificial Intelligence continues to improve, but if we do not make good use of it, it will also continue to confuse us more than it helps us.