AI Hallucination…

I don’t have the habit of laughing out loud when I watch videos. I tend to watch videos silently. Videos are usually too slow for me, so I need to fidget a bit, and mostly listen to videos and watch them occasionally. And since I wander off easily, I play the bubble shooter game on a different monitor; the turn based version of the game is what I play, so that I keep listening and at the same time, not wandering off if I am also doing other things that require more attention.

Just a few minutes ago, I was watching a video on Linux, and the narrator was talking about the lawsuit against OpenAI about copyright infringement by some newspapers. One of the complaints in the lawsuit is that ChatGPT, I assume that is what they are referring to, damages the reputation of the journalists of the newspapers when ChatGPT uses the names of the journalists when it hallucinates. That is when I laughed out loud. That struck me as extremely funny.

Is “hallucination” the right word when an AI system “invents” things out of thin air? I guess I instinctively thought it is not. Maybe “imagination” is better? When people do the same thing, what do we call it? The phrase “false inference” is probably the most appropriate? But inference is not what LLMs do. They don’t do guessing either. They just do “statistically this is probably the most likely”.

Of course, the problem with statistics is that it does statistics very well, that is, given enough scenarios, it can infer the likelihood of something happening or something being true. Or false, or whatever. But it is always about likelihood. Not certainties. When you are dealing with individuals, on the other hand, likelihood is no good. It is, or it is not, for factual stuff. So, either the journalist did say something, or they did not, in a particular article, or ever. We don’t talk about the likelihood of a journalist saying something in a particular article.

So, this is the crucial thing about all statistics based systems: they are wonderful in handling likelihood. They are not very good in handling individuals, or facts, for that matter.

That brings us to the joke Douglas Adams had in his “Hitchhiker’s Guide to the Galaxy” trilogy of four (or is it five) parts. Statistically, the human population in the galaxy is zero.

That is what we must always bear in mind when we use any tools that is based on statistics. It is always about likelihood. Never individuals or certainties. Your chance of winning a lottery is some small number. Before the lottery expires. After it is expired, there is no more probability. You have either lost or won. Perhaps nobody wins. There is no more likelihood at that point, only facts.