JayUtah
Penultimate Amazing
I agree. I'm not opposed to using AI for fact-checking in general, as long as you're not asking Grok to evaluate Elon Musk's latest claim. Previously, using a search engine meant you had to sift through a bunch of relatively unsophisticated textual matches to see how the referenced documents applied to the question at hand. It's time-consuming and intellectually difficult. Since the AI has previously digested the whole Internet, it can use mathemagical computation to organize the available documentary evidence. Depending on how you write your prompt, you may get a good summary on the merits of a controversial subject, or you may get a "balanced" view that tries to give some weight to both sides. That's why it always pays to dig deeper if you're using the AI to fact-check. It's tempting to offload the actual thinking to the AI.
As opposed to using the AI's engine to synthesize a reasonable, hopefully objective summary of the available facts, it's when you use that same AI to generate things that a separate fact-check becomes important. The AI may hallucinate. Or it may simply invent things as instructed in the prompt and render them realistically enough to convince the casual consumer. Intentionally fake content is the pressing danger.
Using AIs to write legal briefs is a good example. When a lawyer alleges a prior determination of law in favor of a point, a citation is universally required. The opposing party and the court must be able to see and test the application of that reasoning to the present case. Lately some lawyers have gotten into hot water for submitting AI-written briefs that rely on hallucinated cases that don't actually exist. That's sanctionable behavior. Even if you're using AI in good faith for this task, you need to separately fact-check your brief's citations against a properly stupid source such as Westlaw.
Imagine if we get into the habit of conversing with a generative AI in both its generative mode and its summary or conversational mode. Your newly-generated brief cites to Smith v. Gulf Coast Energy Corp., 742 F.3d 512 (5th Cir. 2014). You ask the same AI, "Does this case actually exist?" (It doesn't—I totally made it up, but the citation format looks official.) What will the AI say? How can you have confidence in it? If it says, "Yes, Dave, that's a real case," it's a lie that's nevertheless consistent with its previous work. If it says, "No, Dave, that's not a real case," then how can you depend on the legal validity of the brief it just generated? These aren't hypotheticals. This is actually happening.
So the question is not whether AI can or should be used for fact-checking. It's that AI can both generate and test content. However you use AI, it's prudent at this point not to get into a situation where the fox guards the henhouse.
