I have an idea that could make AI fact-checking significantly more reliable.
I call it Source-Lineage-Aware Fact Checking. when AI searches the internet to verify something, it shouldn’t simply count how many websites agree. Hundreds of websites can repeat the exact same information because they all copied one original article. That can make one source look like hundreds.
Instead, AI could use plagiarism and citation analysis to identify which sources are actually independent and group copied information together. For example, if 500 websites all trace back to the same original article, the AI could treat that as one source rather than 500.
Then it could compare the genuinely independent sources, give more weight to primary sources, and explain when an apparent “consensus” is really just thousands of websites repeating the same information.
This could become especially important as more AI-generated content fills the internet. Otherwise, future AI systems could accidentally learn something false simply because an earlier AI-generated claim was copied thousands of times.
The basic idea is simple: don’t count how many websites say something—count how many independent sources actually support it.
I think this could be a useful addition to how AI systems research and fact-check information.
I agree with you. I ran into this problem when trying to use some open-source software to perform fact-checking tasks for me. Here’s what it looks like.
It’s relatively easy when you’re checking a false claim that has no real presence online: ideally, you shouldn’t find any credible evidence supporting it, and that’s it. But when a false claim is already circulating online, bad things start to happen:
Such a claim has supporting sources, so your retrieval system will find them.
If you’re using search engines and a standard context-building approach, you’re basically falling into a similarity trap. For these tools, relevance is often measured through similarity, so your pipeline will naturally tend to prefer sources that support or repeat the claim.
When I tried a popular tool like GPT Researcher and looked into its internals, I was surprised to find that, even though it has a source-curation phase, that phase happens only after the context has already been collected, chunked, embedded, and finally compressed. For a fact-checking task, this makes the process potentially unreliable: curating sources that late may already miss the point, because you’re no longer really curating (actually bumping up by relevance/similarity) sources, but rather highly processed chunks of data.
All this may be effective for standard research, where gathering information closest to the topic works well. But for fact-checking, it can become a mess.
As for GPT Researcher, I made a PR that introduces an isolated source-assessment step, scoring sources against an admission policy before they enter the evidence context, because this is what I think may help.
I’ll leave a link to the work, as I think it’s relevant here and this whole subject is interesting: assafelovic/gpt-researcher/pull/2090 @ github
Search and deep research can help inspect cited sources: https://help.openai.com/en/articles/10500283. We’ll pass along the request to distinguish independent evidence from copied sources, including earlier source-quality checks. No timeline to share.