Fact-Checking Chapter 23: Chemotherapy Errors & Self-Auditing in Medical AI

Chapter 23 of the Medhavy cancer textbook covers chemotherapy, and it is the densest chapter the reviewer had worked on. The video reports what the audit found in the textbook, and then what it found in the reviewer's own work. Its conclusion is that checking AI-generated text is not enough, because the same failure can happen to the person doing the checking.

The scope of the audit

The chapter spans seven sections: an introduction to chemotherapy, principles of treatment, classes of chemotherapeutic agents, pharmacology, drug resistance, side effects and a summary. It contains 474 sentences, and 136 of them were flagged for verification. Almost every paragraph contains a drug name, a trial result, a dosing formula or a regulatory fact.

The reviewer examined 86 rows in total. Of those, 52 needed revision and 34 were approved as written.

The patient safety finding

The most serious problem involves patient safety. The textbook presents venetoclax for multiple myeloma as established practice. According to the video, that use is not FDA approved, and the FDA label carries a mortality warning against that exact combination. A phase 3 trial published in January 2026 missed its primary endpoint and showed more treatment-emergent deaths in the venetoclax arm. The video calls this the kind of error that could directly affect patient care.

Historical and dosing errors

The audit also found mistakes in the history of chemotherapy and in drug facts:

• The textbook dates arsenic-based treatments to the 1900s, but Heinrich Lissauer treated leukemia with arsenic in 1865, about 35 years earlier.
• It attributes nitrogen mustard research to World War I, when the work came from World War II.
• The half-life of methotrexate is stated backwards relative to the FDA label, which matters because the timing of rescue treatment depends on it.

Auditing the audit

What made this chapter different, according to the reviewer, was the quality control. The reviewer ran three full audits: one on the science, one that refetched every source and confirmed the quoted text was on the page, and one that compared the deliverables against the instructions.

Those passes caught five errors in the reviewer's own work, including two quotations attributed to the wrong papers, and one verdict the reviewer reversed after reading the 2025 evidence on dexrazoxane. The video draws a direct parallel: the textbook cites sources that do not say what it claims, and the same failure can happen to a reviewer.

A practical file warning

The video adds one technical note. Files two through five have a character encoding defect that corrupts Greek letters. It needs to be fixed before any text edits are made, because otherwise a find-and-replace will silently fail.

Key takeaways

• Chapter 23 is a dense chapter: 474 sentences, 136 flagged, 86 rows reviewed.
• Of the 86 rows, 52 needed revision and 34 were approved as written.
• The textbook presents venetoclax for multiple myeloma as established practice, although the use is not FDA approved and the label carries a mortality warning.
• History was wrong in places: arsenic treatment dated 35 years too late, and nitrogen mustard research tied to the wrong world war.
• A backwards methotrexate half-life matters because rescue timing depends on it.
• Three separate audits caught five errors in the reviewer's own findings.
• Checking AI-generated text also means checking yourself.
• Fix encoding defects such as corrupted Greek letters before editing text.

Try it yourself

Take something you have already fact-checked and audit your own audit. Open the original source and confirm that the exact sentence you quoted is really on the page. The reviewer's five errors were hiding in the checking, so that is the place to look.