Earlier this year, scientists discovered a peculiar term appearing in published papers: "vegetative electron microscopy". This phrase, which sounds technical but is actually nonsense, has become a "digital fossil" -- an error preserved and reinforced in artificial intelligence (AI) systems that is nearly impossible to remove from our knowledge repositories. [...] The case of vegetative electron microscopy offers a troubling glimpse into how AI systems can perpetuate and amplify errors throughout our collective knowledge.
Curated from theconversation.com · 15 April 2025 →
Two papers from the 1950s were scanned into a digital archive, and the software running across two columns of text joined vegetative in one column to electron microscopy in the next. That phrase, which means nothing, sat in the scanned corpus, went into Common Crawl, went into the training data, and now appears in more than twenty published papers. GPT-3 completes it; GPT-2 and BERT do not, which is how the researchers dated the contamination. Elsevier initially defended one instance of it. This is the case for keeping the word hallucination at arm's length: nothing was invented here. An error entered a corpus, and the corpus is not something anyone can now edit.
