AI is brilliant. Get a human to read it anyway.

First posted on LinkedIn, April 2025.

AI is a brilliant tool and can add real value to any organisation. I use it, I build with it, and I’d struggle to make a case against it.

But there are lots of examples of ways it can fail or create inadvertent problems, and “vegetative electron microscopy” is a very unfortunate one.

It isn’t a technique. It isn’t anything. It’s two unrelated fragments of text that got welded together by a scanner and then went on to have a career, and it now appears in more than twenty published scientific papers.

The origin is a pair of papers from the 1950s in Bacteriological Reviews, printed in two columns. When they were digitised, the OCR didn’t understand the column break: it read straight across the gutter, took “vegetative” from the end of one column and “electron microscopy” from the start of the next, and produced a phrase no human reader would ever have made. Your eye stops at the edge of the column. The machine’s doesn’t.

Then it happened again, independently. In Farsi, the words for “vegetative” and “scanning” differ by a single dot, so a translation slip produced the same nonsense from a completely different direction. Once the term appeared in enough documents it was in the training data, and researchers found that newer language models will readily suggest it, while older ones never do. The people who traced it call this a “digital fossil”: an error preserved in the rock of our knowledge base and very hard to get back out. Publishers have handled it inconsistently, and Elsevier initially attempted to justify the term before eventually issuing a correction.

Why this should interest anyone running business systems

Laugh, then look at the shape of it, because that is a data migration. It’s every legacy document scanned into a system of record, every field mapped to roughly the right column, every historic spreadsheet imported during a cutover weekend so the go-live date wouldn’t slip.

Nobody in the story was negligent. The scanner did what scanners do. The translation was a near miss. The authors trusted a term that appeared in the literature, and reviewers didn’t query it because it looked exactly like something that ought to exist. Bad data rarely arrives announcing itself as bad data. It arrives looking plausible, and then it gets inherited (reported on, decided on, carried into the next system at the next migration) until it has the authority of history and the person who could have explained it left in 2019.

What helps is boring, and mostly cheap:

• Validate at the point of ingestion, not the point of complaint. Checking a migration is tedious. Unpicking years of decisions made on bad figures is worse.
• Sample it with human eyes. Automated reconciliation confirms the row counts match; it won’t tell you the content is nonsense.
• Keep the provenance. If you can’t say where a figure came from and who last verified it, you can’t defend it or confidently correct it.
• Make it safe to ask the naive question. Every one of those papers passed somebody who could have said “what actually is that?”

We’re now pointing AI tools at exactly these archives: summarising them, extracting from them, answering questions from them. That’s genuinely useful, and I’d encourage it. It also means an error made by a scanner decades ago, about a paper printed decades before that, can be laundered into a confident answer given to somebody today who has no way of knowing it was ever a mistake.

So: brilliant tool, use it well. And it’s important to remember that sometimes your GenAI content needs to be proofread by a human.