[mysterydip]

What do you use as “ground truth”? The page says “independent sources”, and I’m sure there’s too many to list, but my question is how are they vetted as being truthful and how are two sources with opposite viewpoints reconciled?

[skorniienko]

There is no whitelist source. It can't rate sources for truthfulness. When sources conflicts - skill drops verdict to misleading or unverifiable and both are linked. Can't pick a winner at the moment. That's probably the weakest part and still a judgement call for a model.

[gutechh]

Maybe you read them already, but wikipedia has a bunch of interesting pages about sourcing and truth like: https://en.wikipedia.org/wiki/Wikipedia:Verifiability,_not_t... and https://en.wikipedia.org/wiki/Wikipedia:Reliable_sources

[skorniienko]

I'll look at reliable-sources, and you are right "verifiability, not truth" is basically what I landed on too. Model doesn't get to decide what's true, it just needs to cite something.

Will read reliable-sources page properly.

[simonw]

Something I've found surprisingly effective is telling ChatGPT to "use credible sources" - you can then watch its thinking trace and see it do things like ruling out random blogs, considering media publications with a good reputation for fact checking, and double-checking information that seems unlikely.

[vunderba]

For any LLM with web tool usage I’ll often suggest that it pull relevant papers from arXiv, with an expectation that it cite sources. I find that this helps improve accuracy and lets me go deeper into the material if I want to.

[westoque]

I don't think random blogs should be ruled out. Just because a blog is random doesn't mean its untrue.

[simonw]

The reputation of the writer matters a lot. Recent models seem to be pretty good at evaluating author credibility as well.

What's dangerous is treating information from an anonymous author as factual. You gotta check that in other ways.

[bulbar]

It should not be seen a credible source though.

It can be, but that's hard to correctly judge on at scale.

[skorniienko]

Love it, I'll run tests for "credible sources" and will compare the results. Could be a first feature-request implemented!

[skorniienko]

Added this as issue so I don't forget https://github.com/SerhiiKorniienko/bullshit-detector/issues...

[autonomousErwin]

I could imagine running multiple different AIs seeing if they agree if something is true.

I imagine this is similar to how Twitter's/X's community notes work. Something along the lines of when you have two accounts that disagree traditionally and they agree on something. That's how you know it's likely to be true.

So from that approach, you actually want to have two AIs that are on the other side of the spectrum of whatever you're trying to find out if it's true e.g. if a conservative AI and a liberal AI both agree that something is false, it's highly likely to be false.

[mtweak]

This is extremely problematic. Much of news is highly syndicated, so what looks like 10 credible sources are actually just 2. I think the best you can aim for is empirical sources: receipts, videos, photo evidence, public disclosures directly at the source. Archiving is ok if disclosed.

I don't think LLMs can discern truth when the majority of news sources are incentivized primarily for views.

[skorniienko]

Filed both issues from this: https://github.com/SerhiiKorniienko/bullshit-detector/issues... - for syndication https://github.com/SerhiiKorniienko/bullshit-detector/issues... - empirical sources

[skorniienko]

Empirical sources - would be a first candidate to be implemented in a skill. And you are right about 10 credible sources are actually 2. It will look like 10 independent sources to it, real gap. Skill does prefer primary source over secondary. For example it'll prefer NASA article over blog post who cited NASA article.

I'll work on empirical sources improvement in a next releases.

[skinfaxi]

You may have overlooked this part of the parent's comment which addresses your syndication issue:

> Something along the lines of when you have two accounts that disagree traditionally and they agree on something. That's how you know it's likely to be true.

Outfits that blindly parrot talking points would not be at odds and thus not be good candidates.

[hunterpayne]

It also doesn't help that many news sources will knowingly fudge the truth. There are topics about which they are constant sources of misinformation. Seems like depending on the topic (anything science related) you wouldn't want to use news sources at all. Other topics, you probably only want to use news sources. And arranging all of this weighting of sources is going to be difficult and controversial as well. Doesn't mean it can't be done. Doesn't mean it wouldn't be valuable. But there is a lot to it and most of it isn't so much about technology or math as it is about understanding who should and shouldn't be considered authoritative about what topics.

Sounds challenging to me...