I have to roll my eyes a little bit every time I hear someone talking about AIs hallucinating and how this is such a problem and we’ve got to do something about it. This is such a misguided take.
Mountainpeaks Semien Ethiopia [commons]
It reminds me very much about an anecdote I caught on one of Joscha Bach’s podcast appearances. He mentions the interaction he had with someone and that someone spoke about how we “see” faces in clouds and wallplugs and so on; to which Bach replied basically, “and we see even faces in faces.” The cognitive science take here is that nothing about a “face” says it is a face by itself without the complex neural processing on the other side, which is specifically wired for perceiving face-patterns. When a tree falls in the forest and no-one hears it, does it make a sound? Same story.
What I’m trying to say. AI doesn’t hallucinate. Not any more than us. Or said differently, we all hallucinate, all the time. There is reality, and there are our models of it, which predict reality more or less accurately. Sometimes we get things wrong. The AI gets more things wrong. But no qualitative difference, only quantitative.
The whole thing drives me nuts because it is connected to a worldview which I’m trying to fight for 20 years or so. It is that of the naïve perception that such a thing as a fact exists, to boil it down to its simplest. For today, because I can’t dream of convincing you in short time, an appeal to authority must suffice, and I allow myself to simply point to the book What Is This Thing Called Science? by Alan Chalmers, which I have right next to my bedside table in its fourth edition.
As well [sic] shall see, the idea that the distinctive feature of scientific knowledge is that it is derived from the facts of experience can only be sanctioned in a carefully and highly qualified form, if it is to be sanctioned at all. - from the Introduction (p.xx)
It is a spectacularly good introduction into the “problem” of science, or let’s call it more respectfully into the Philosophy of Science. In any case, it should be possible for the sufficiently motivated and honest layman to come out on the other side of reading it[1] that the story of “facts” is not “solved,” or trivial (as hinted above). The stubborn enough reader will probably insist that this is because philosophers have the tendency to make simple things complicated. Urm, … f.. fair, enough… (that was hard; I have to admit there is not nothing to it).
Anyways, if there was such a thing as a fact, we wouldn’t have science, the whole business of which is to determine what we actually see, to interpret and re-interpret supposedly sense-data, but really measurement data, through constantly more capable conceptual lenses - yielding ever more powerful explanations. But where is that thing called “fact”?
Did person A murder person B? C says so, but person D says it can’t be. That is why we have judges. What story do the Roman coins found in Tomares, a suburb of Sevilla, tell us? Archeologists are there to tell a convincing story. Are we to believe they all agree? In 50 years’ time?
Was Thomas Aquinas’s birth-year 1224, or 1225? May it be our best guess at some point that it was 1225, only to be revealed later, by us coming into possession of new sources, that it was 1224? Are we sure then?
Did it rain yesterday? Well, three raindrops fell. Does that count? By now you probably can already guess how much respect I have towards the notion of fact-checking … But that’s maybe a story for another day.
Related article, On “Grok is this true?” from a friend of the blog, Greg Slepak, can be found here.
Footnotes
To my slight embarrassment, I have to admit that I haven’t read it fully through yet. Call me hypocrite. I only say in my defense that after 20 years of reading philosophy of science, even if only cursorily, I claim to have somewhat of a grasp of the matter. Enough to extrapolate from the first third of the book.