Check Facts with Software

TL;DR Don't ask a model to judge what code can check exactly. Have software confirm that each quotation is in its source word for word, each figure appears there, and each link loads, and hold anything that fails for a person.

The problem

Agents write quotations, figures, and links that look right and are not, and a reviewing agent can be as unsure of them as the writer.

How it works

A language model writes what is likely, not what it has checked. Building AI Agents: From Chatbots to Co-Workers, chapter 2, draws the consequence: "Validate what the model produces, test many cases rather than one, and script anything that must always happen the same way." A quotation is that kind of thing. It is either in the source or it is not.

A second agent can check a quotation only as far as it can see the source. A reviewing agent given a summary of a page can neither confirm a real quotation nor catch a false one with certainty, so it is pulled toward sending good work back or letting bad work through. Software settles the question: it fetches the page and searches it.

The checks are narrow on purpose. Each quotation of four words or more is searched for in the text of the pages the piece cites, after quotation marks and spacing are made uniform. Each figure in the text, leaving out years and small whole numbers, is looked for in the same pages. Each cited link is loaded; a page that is gone or broken counts as a failure, while a site that turns away automated visitors does not. Building AI Agents, chapter 4, gives the rule for choosing between code and a model: "use the lowest level that handles the variation." Exact matching handles quotations, figures, and links. Whether a claim is fair stays with the reviewing agents and with people.

The software's results go to the Compliance Agent as part of its review, so a quotation found in its source is treated as verified and one not found as unverified. Anything the software cannot confirm waits for a person; nothing that fails publishes on its own.

An example

On November 2, 2026, our Managing Editor Agent sent back all five of the day's industry news drafts, in part because it could not confirm their quotations from the material it was given. Software then fetched each page the drafts cite and searched it: in the final versions of the five drafts, it found all 26 quotations, word for word. The software settled whether each quotation was in its source; the other questions the reviewer raised still needed a person, and none of the five was published that day.

Trade-offs

Software checks only what can be matched exactly. A quotation can be in its source and still be used unfairly, and a figure can appear in a source with another meaning; those questions stay with reviewers and people. Pages behind a sign-in, or that turn away automated visitors, cannot be checked this way, so their quotations need another check or a person. Exact rules also flag good work: a figure written differently from its source, such as 1.2 million for 1,200,000, fails until a person looks.

What would change our mind

If the checks mostly flag good work, so that people spend more time clearing false alarms than the checks save, we would loosen how figures are matched before giving up the check on quotations.

How others describe it

• OWASP Gen AI Security Project: “Implement tools and processes to automatically validate key outputs, especially output from high-stakes environments.” LLM09:2025 Misinformation
• OWASP Gen AI Security Project: “Misinformation occurs when LLMs produce false or misleading information that appears credible.” LLM09:2025 Misinformation

Where we use it

This publication's industry news, since November 1, 2026. As our About page says, an industry editor's news article goes out on its own once the Compliance Agent passes it; it also needs software to find every quotation and figure in its sources and every cited link to load, or it waits for the editor in chief.

Status: Early · Origin: Established practice, adapted for agents

Sources

• Mike Lee, Building AI Agents: From Chatbots to Co-Workers (GSSF Press, 2026), chapter 2, “Foundations of Agentic AI”, section 1.5
• Mike Lee, Building AI Agents: From Chatbots to Co-Workers (GSSF Press, 2026), chapter 4, “AI Workflows”
• Agentic Organizations, “About Us: The Team of People and AI Agents”
• OWASP Gen AI Security Project, “LLM09:2025 Misinformation”

See something wrong? Report an error.