Human-Only Review Misses Errors Too, and AI Can Catch More

Keeping AI out of GxP review does not keep errors out of GxP records. FDA cites a failure in records, or in their review, in 301 of the 540 drug manufacturing warning letters it has posted since January 2021. In the few studies that have compared people working alone with people checking AI, the pair made fewer errors when the AI was usually right, though none of them looked at a GxP document.
The real choice is between two error rates, and the company answers for its records under either. Studies from cancer screening and medical research show that the comparison can be run. They also show that a wrong AI suggestion can drag a reviewer down, so each company has to run it on its own process.
FDA cites record and review failures in more than half of drug GMP warning letters
Most of FDA's recent drug manufacturing warning letters, its formal notices of serious violations, fault how a company keeps or checks its records. We read the 540 letters dated 22 January 2021 to 10 November 2026 that cite CGMP, FDA's drug manufacturing rules. Then we counted those citing any of four failures in records or their review.
540 drug CGMP warning letters, 22 January 2021 to 10 November 2026301 cite at least one of these failures; a letter can sit in several rows60 of 60 letters read by hand, 15 per row, that match their row408 of 408 quotes found word for word in the letter text
Open a row to see the letters behind its count. The highlight marks FDA's words about the failed record or review.
195 of 540 letters, newest first
FDAYour firm failed to thoroughly investigate any unexplained discrepancy or failure of a batch or any of its components to meet any of its specifications, whether or not the batch has already been distributed (21 CFR 211.192).
Read the letterFDAYour firm failed to thoroughly investigate any unexplained discrepancy or failure of a batch or any of its components to meet any of its specifications, whether or not the batch has already been distributed (21 CFR 211.192).
Read the letterFDAYour firm failed to thoroughly investigate any unexplained discrepancy or failure of a batch or any of its components to meet any of its specifications, whether or not the batch has already been distributed (21 CFR 211.192).
Read the letter
Every card is in the data (JSON), and the counts by year are in the CSV below.
The failure FDA cites most is the one review exists to prevent. 21 CFR 211.192 requires a company to review production records and investigate any unexplained discrepancy before releasing a batch. It appears in 195 of the 540 letters. Investigators also cite it on the Form FDA 483, the list of observations handed over at the end of an inspection. In fiscal 2025 it appeared at 209 of the 638 drug inspections with posted citations.
These letters do not measure an error rate. They come from the worst inspections, and some describe deliberate falsification that no reviewer would stop. They do establish that review by people fails often enough for FDA to cite it in most of these letters. That process, not a flawless one, is what AI should be compared with.
Measured directly, manual record work goes wrong in a few percent of entries
Where researchers have counted errors in manual record work, the rates sit well above zero. In one clinic, staff typed blood glucose results into the health record by hand while the same results arrived from the device. Of 6,930 typed results, 260 (3.7%) did not match the device (Mays and Mathias, JAMIA 2019).
The rate depends on the task, and a second person lowers it without removing it. A review of 93 studies of clinical trial data found that copying data from medical records by hand had an average error rate of 6.57%. Typing data from paper forms had a rate of 0.29%, or 0.14% when typed twice and compared (Garza et al., International Journal of Medical Informatics 2025). Checking misses things too. When reviewers sorted research summaries in a randomized trial, one reviewer alone missed 13% of the relevant studies and two working independently missed 3% (Gartlehner et al., Journal of Clinical Epidemiology 2020).
GxP rules require a second person for this reason. The question for AI is whether it lowers the errors that survive that second check.
In three studies outside GxP, people working with AI made fewer errors than people alone
The strongest evidence comes from breast cancer screening, where a missed cancer is a countable error. Sweden's MASAI trial assigned 105,934 women at random to one of two kinds of reading. In the standard group, two radiologists read every mammogram. In the AI group, software scored each exam, and radiologists read it with the score in view: one radiologist for most exams, two for the highest scores.
| Study | People alone | AI plus people |
|---|---|---|
| MASAI, Lancet 2026: 105,934 women, randomized | Caught 73.8% of cancers | Caught 80.5%, same false alarms |
| MASAI, Lancet Digital Health 2025: same trial | 5.0 cancers per 1,000 women; 109,692 readings | 6.4 per 1,000; 61,248 readings (44% fewer) |
| PRAIM, Nature Medicine 2025: 463,094 women in Germany, not randomized | 5.7 cancers per 1,000 women | 6.7 per 1,000; slightly fewer called back |
| Gartlehner et al., Annals of Internal Medicine 2025: 9,341 data items from 63 studies | 11.0% of items wrong | 9.0% wrong; 41 minutes saved per study |
In screening, AI helped radiologists find more cancers, but it has not yet been shown to cut the cancers screening misses. MASAI counted the cancers that surfaced between screenings, the ones screening missed, and the AI group had slightly fewer, a difference small enough to be chance. PRAIM is weaker evidence, because radiologists chose when to use the software, which its authors say may bias the result.
The Gartlehner study is closest to document work, and its gain was modest. A language model drafted each data extraction, and a person checked it against the source article. Serious errors were about equal in the two groups, at 2.5% and 2.7% of items, so the gain came in minor errors and in time.
None of these studies tested GxP records, so none says what AI would do to batch record review or a deviation write-up. They show something narrower: a person checking AI output against a source can beat people alone, by a countable margin. A company has to count that margin in its own process, for example with a planted-error test, before relying on it.
A wrong AI suggestion can leave reviewers worse off than working alone
People tend to follow a tool's suggestion even when it is wrong, a habit researchers call automation bias. In a Radiology 2023 study by Dratsch et al., 27 radiologists read mammograms alongside a supposed AI that was deliberately wrong on 12 of 40. The least experienced readers rated 79.7% of mammograms correctly when the suggestion was right and 19.8% when it was wrong. Even the most experienced readers fell from 82.3% to 45.5%.
A JAMA 2023 trial by Jabbour et al. measured that cost against people working alone. Hospital clinicians diagnosed 73.0% of cases correctly without AI and 75.9% with an accurate model. With a model built to be wrong in a consistent way, they fell to 61.7%, well below their score alone. Explanations of the model's reasoning recovered little of that loss.
So AI with a reviewer beats a reviewer alone only while the AI is usually right and the reviewer checks it. Each condition needs a control. The tool is tested on the company's own records, and a qualified person checks each point against the source and records that check (what the record should show). A signature alone does not show the check happened, which is the gap FDA found in its first warning letter citing AI.
FDA checks its own AI output and expects methods to stay current
FDA uses AI itself and keeps experts checking its output. It launched its internal AI tool, Elsa, in June 2025, with staff already using it to "accelerate clinical protocol reviews, shorten the time needed for scientific evaluations, and identify high-priority inspection targets" (FDA, 2 June 2025). With the May 2026 upgrade, it said "human subject matter experts verify all inputs, analytic processes, and output implementation" (FDA, 6 May 2026).
FDA has not said companies must use AI, but its CGMP rules do expect methods to keep up. FDA says the "C" stands for "current", "requiring companies to use technologies and systems that are up-to-date" (FDA). So the case for AI rests on evidence rather than on the rule. Where a company has shown that AI with review makes fewer errors in a process, human-only review of that process becomes harder to defend.
Questions and decisions to bring to a steering committee:
- What error rate does our second-person review find today, by record type?
- Which record type will we test AI-assisted review on first, side by side with current review?
- What error rate and time per record must the pilot beat before we rely on it?
- How will reviewers record which sources they checked and what they corrected?
- How will we test whether reviewers catch planted AI errors?
- Who owns the tool's change control when the vendor updates the model?
Data and method
We used Arca's copy of FDA's warning letter pages. A drug CGMP letter is one from a drug office that cites a section of 21 CFR Part 211 other than 211.34. There are 540. A letter counts in a row if it cites 211.192, 211.188 or 211.194, or uses the words "data integrity". A letter that describes a record failure without either is missed, so the counts are a floor. FDA also removes some letters from its pages. We left out 211.100(b), because letters usually cite it alongside process validation failures. A hand check of 60 letters, 15 per row, found all 60 correctly placed. The Form 483 count comes from FDA Data Dashboard inspection citations. Study figures are from each journal article's abstract or full text.
Downloads: warning letter counts by year (CSV).
Frequently asked questions
Isn't it safer to just not use AI?
Only if your current review is close to error-free, and FDA cites record or review failures in more than half of its drug manufacturing warning letters. The defensible position is to measure your current error rate and adopt AI only where it lowers that rate, with documented review.
Who is accountable if the AI is wrong?
The company, as for any GxP record. The qualified person who reviews and approves the record answers for its content, whoever or whatever drafted it.
Do the mammography trials prove AI works for our documents?
No. In those trials AI with a reviewer beat reviewers alone on a task where errors can be counted, and in other studies a wrong AI made reviewers worse. Run the same comparison on your own records before relying on AI.