Fact-Checking the Medavy Cancer Textbook: Chapter 2 Audit Findings

Chapter 1 of the Medhavy cancer textbook was mostly right. Chapter 2 was not, and the video explains that the errors were not scattered at random. They came in clusters, because a whole run of numbers had been lifted from one stale source. The reviewer describes the audit and what it suggests about AI-generated text.

What Chapter 2 covers

Chapter 2 has five major sections: defining cancer and uncontrolled growth, the history of cancer research, the hallmarks of cancer, oncology terminology, and cancer epidemiology. It is one of the largest chapters in the textbook, and the video says it took a careful, systematic review to get through.

The size and result of the audit

The reviewer checked 46 claims in the main tab, 134 sentences in the AI-only tab, and eight editorial issues. Of the 46 main claims, 25 were confirmed true, 18 were false and three were flagged as ambiguous. That is a much higher error rate than Chapter 1, where most claims checked out.

The stale data problem

The biggest pattern was stale data. Six claims in the history of cancer research section all came from the same 2021 commemorative article and had not been updated for a 2026 textbook. Two examples from the video:

• The text says there are 71 NCI-designated cancer centers, but the current count is 74.
• It says about 100 targeted therapies have been approved, but the real number is now over 157.

The epidemiology section had the same problem. Three claims were pulled word for word from a 2007 WHO publication, which makes them nearly 20 years out of date. The numbers on obesity, occupational exposure and hepatitis B have all changed significantly since then.

Errors in content the AI flagged as safe

Even in the AI-only tab, where sentences had been flagged as not needing human review, the reviewer caught three errors. One reversed a well-known biological pathway. Another placed a 12th-century scholar in the 13th century. A third garbled one of the six hallmarks of cancer described by Hanahan and Weinberg. The video points out that these are exactly the kind of mistakes that slip through without human review.

What the pattern shows

The reviewer's conclusion is that AI-generated text tends to grab numbers from whatever source it drew on, without checking whether those numbers are still current. For a medical textbook, the reviewer says, that is not good enough, and every claim needs verification.

Key takeaways

• Errors in Chapter 2 came in clusters tied to a few stale sources, not as isolated slips.
• Of 46 main claims, 25 were true, 18 were false and three were ambiguous.
• Six history claims all traced to one 2021 article and were out of date for a 2026 book.
• Three epidemiology claims were copied from a 2007 WHO publication.
• Examples of drift: 71 versus 74 NCI-designated cancer centers, and 100 versus over 157 targeted therapies.
• Sentences flagged by AI as not needing review still contained a reversed pathway, a misdated scholar and a garbled hallmark.
• A stale number always sounds plausible, so plausibility is not a test.

Try it yourself

Take any document that quotes statistics and pick five of them. For each one, find the original source and check the data on it, not just whether the source exists. Sort the numbers into three groups: current, true once but since changed, and impossible to date. Do not judge a number by whether it sounds plausible.