AI FactScan
Curation Nation author Steven Rosenbaum published a new book about AI and truth. Five days later, a reporter asked him for the source of a quote attributed to an anonymous DARPA researcher.
That question opened a much larger problem. Rosenbaum's publisher brought in two human fact-checkers to review the manuscript line by line. They found 26 quotes that could not stand as written. Each was traced to a primary source and corrected, rewritten as a paraphrase, or removed.
The irony was obvious. A book called The Future of Truth contained language that AI had made more convincing than the evidence allowed.
A book about truth failed a source check
Rosenbaum says AI did not write the book. He used it to organize research, summarize documents and refine language. The manuscript still went through five editorial rounds and more than 4,000 tracked changes.
Those safeguards improved the writing, but they did not answer a separate question: did every quotation match something a real person had actually said?
The errors that looked broken were easy to catch. A false DOI, dead URL or missing source announces itself. A polished sentence attributed to a real person is harder. It reads like a clean quotation, especially when it sounds consistent with that person's public views.
Proofreading asks whether a sentence reads correctly. Source checking asks whether you can retrace it.
How do I check if an AI answer is based on the right data?
Start with the claims that would cause the most damage if they were wrong: direct quotes, statistics, legal authorities, scientific findings and descriptions of what a named person believes.
Then follow the source all the way back. A citation is useful only when it lets you find the original material and confirm that it supports the words beside it.
Why a working link still is not proof
AI does not need to invent an entire source to mislead you. It can cite a real article that never makes the claim, attach a genuine DOI to the wrong title, or turn a loose paraphrase into a direct quotation.
This distinction matters when choosing an AI fact-checking tool. Look for a tool that keeps the source citations visible and lets you open them. A score without the underlying URL gives you no way to inspect the evidence yourself.
A Columbia University School of Nursing team scanned about 2.5 million open-access biomedical papers and found 4,046 nonexistent references across 2,810 papers. The study was published in The Lancet. And that was only the easiest error to test: whether the cited paper existed at all. The audit did not count real papers that were misquoted or did not support the claim, and it could not prove that AI produced every fake reference. Rosenbaum's book passed through five editing rounds. These papers passed peer review. In both cases, bad sources still made it through.
When the AI provides no source
Ask again for the original URL and the exact passage supporting the claim. If the model still cannot provide one, treat the claim as unverified.
For a direct quotation, the rule can be even simpler: no retrievable original, no quotation marks.
The useful lesson is not to stop using AI
Rosenbaum did not respond by pretending there was a pre-AI world to return to. He corrected the book, disclosed what happened and published a revised edition.
The practical lesson is less dramatic. AI can help organize and edit research, but it cannot be the final authority for the material it helped produce. Before publication, someone must still open the source.
That is why AI FactScan organizes the sources attached to an AI answer: so you can open the original pages and make that final check before citing them.