AI FactScan
AI-generated writing now occupies a visible part of the web. Pew Research Center analyzed nearly half a million English-language webpages collected through Common Crawl. In its July 2026 sample, 10% of all pages showed significant signs of AI authorship or editing. Among pages published after ChatGPT was released, the share exceeded one-third.
Detection models sometimes misclassify individual documents, so Pew used the results to measure a broad trend. Its findings show how quickly AI-assisted text has spread across recently published pages.
AI answers have also moved into ordinary search. Another Pew survey found that 60% of U.S. adults read AI summaries at the top of search results. For many people, an AI answer now appears before the first website they open.
When that answer includes citations, how much confidence should those links add?
The question we asked Copilot
We gave Copilot a reasonable research prompt:
How do you KNOW when a post is AI/Bot-written? Include the related research.
The answer cited six sources. The links and paper-like references made it look well researched.
We selected the complete answer and scanned it with AI FactScan.
A 30-second recording of the question, scan and source checks. Watch on YouTube.
What failed in the source check
Two citations opened Bing search pages with no useful result. They led to a list of possible sources instead of the article, study or report behind the claim.
Two other citations used DOI links that returned DOI NOT FOUND. A DOI, short for Digital Object Identifier, is a permanent ID for a paper, dataset or other research item. It should keep leading readers to that item even if the publisher changes its webpage. When a DOI cannot be found, it may be mistyped, inactive or fabricated, so the cited paper still needs to be located and checked another way.
This does not prove that every sentence attached to those citations was false. It shows that four of the six references failed before we even reached the harder task of reading the evidence.
Try the same source check with AI FactScan for Chrome.
What the citation count leaves out
A citation marker gives an answer a traceable appearance. The count says little about what sits behind each marker. A useful first pass checks whether the source exists, what kind of source it is and whether the answer repeats the same article.
Claim verification comes next. That requires opening the source and reading the relevant passage.
Use the scan to question the answer
AI FactScan gives you a quick map of the evidence behind an AI answer. Missing papers, weak sources and repeated articles stand out immediately, so you can see which claims deserve attention first.
Take those results back to the AI. Ask it to replace a missing paper, provide the original study, show which source supports a specific claim or revise the answer using only sources you have verified.
A source grade starts the next question. Your follow-up is where critical thinking enters.
So how much should you trust the answer?
The original prompt asked how to identify AI-written posts. That is a difficult question, and writing style alone is weak evidence. Detector scores also need context. An answer to that question should be especially careful about the research it cites.
A polished explanation can sound more certain than its sources allow. Check the trail before repeating the conclusion, and keep your confidence within the limits of the evidence.
Treat this as a case study of one Copilot answer. Prompts, sources and results can change from one conversation to the next.