Experimenting with AI

[Tuur]

Not Physical or Life Science; more like technology. In any event, with AI becoming ubiquitous, here's an easy experiment: Ask an AI about something concrete but obscure that you're 100% certain of without any shadow of a doubt. Not anything "gotcha" like 11 + 11 (22 decimal; 110 binary). Just something straightforward. Which AI you ask doesn't matter.

I tried this, and the results were less than impressive. The only thing it got correct was the name of the event and when it happened. A supposed participant wasn't there. The specific event that triggered it wasn't mentioned, nor was the aftermath. Most of the response was fluff worded to pass for something authoritative. It didn't work. It read for all the world like a grade school student trying to bluff his way through an essay question.

In picking something concrete but obscure, you're asking the AI for something there's not a lot of info online to train it. No telling how many words are written about, oh, Washington crossing the Delaware, or calculating the value of Pi, and what AI regurgitates might be passable. Might. Something obscure isn't going to have many words online about it, and what AI comes up with can be questionable. It might be questionable even if there's a lot of words written about it, because it has no reasoning to fact check itself. This experiment is sort of looking at boundary conditions that evaluates the accuracy of AI.

Add to the mix topics that people debate. Add to it things that some people consider as not concrete. And consider that AI has no reasoning to evaluate its own responses.

Now consider the posts that show up on CF that are essentially "AI says." Then consider that if AI can't be relied on when it comes to the obscure, how can we rely on it for anything where it has to imitate "general knowledge?"

 

[partinobodycular]

Then consider that if AI can't be relied on when it comes to the obscure, how can we rely on it for anything where it has to imitate "general knowledge?"

Collectively the group tends to be more accurate than any random individual on their own, therefore information about obscure events will tend to have a greater margin of error due to the limited number of available sources. Information about general knowledge subjects on the other hand tend to be more widely available, with accurate information more likely to be cited within the group, and misinformation more likely to be described as such.

Although AI might not be able to 'reason' on its own, given sufficient sources it will tend to reflect the reasoning done by others.

 

[Ophiolite]

Now consider the posts that show up on CF that are essentially "AI says." Then consider that if AI can't be relied on when it comes to the obscure, how can we rely on it for anything where it has to imitate "general knowledge?"

Anyone unaware of the limitations of AI should not be using it. You don't put a loaded gun, with the safety off, into the hands of a six year old child.

My own experience of AI, when having it generate material, is that worthwhile content produced runs between 40% and 80% per response. This can be increased, especially at the low end, by use of better crafted prompts. Almost no response is error free. The significant thing is that the errors will generally only be recognised as such by someone who is well versed on the topic. When seeking material I am not familiar with I challenge aspects of the posts on the specifics, or note "That seems to contain several errors or ambiguties, please comment." Accuracy is extractable, but one must be alert and put in the effort.

 

[RDKirk]

Collectively the group tends to be more accurate than any random individual on their own, therefore information about obscure events will tend to have a greater margin of error due to the limited number of available sources. Information about general knowledge subjects on the other hand tend to be more widely available, with accurate information more likely to be cited within the group, and misinformation more likely to be described as such.

Although AI might not be able to 'reason' on its own, given sufficient sources it will tend to reflect the reasoning done by others.

It does try to evaluate sources, but that can become difficult for subjects where misinformation can come even from sources that ought to be reputable.

Sometimes you have to tell it go back and check more sources.