A weird phrase is plaguing scientific papers – and we traced it back to a glitch in AI training data

The discovery of “vegetative electron microscopy” illustrates how minor digitization and translation errors can transform into permanent “digital fossils” within AI models. Originating from a scanning glitch in 1950s papers and a subtle linguistic mistake in Iranian publications, this nonsensical term was ingested by large language models trained on vast internet datasets. Rather than correcting the error, modern AI systems like GPT-3.5 and GPT-4 perpetuated it, demonstrating how AI amplifies and embeds misinformation at an unprecedented scale. This phenomenon poses a significant threat to the integrity of open data and scientific knowledge repositories. Because AI models are trained on unverified, massive datasets like CommonCrawl without transparent oversight, errors can become self-perpetuating and nearly impossible to excise. The case highlights a critical vulnerability: as AI-assisted research becomes ubiquitous, fabricated or erroneous terminology can infiltrate peer-reviewed literature, undermining trust in digital information ecosystems and complicating efforts to distinguish fact from algorithmic hallucination. For the open data community, this serves as a urgent call for greater transparency in AI training processes and improved validation mechanisms. Without rigorous auditing of the data feeding into language models, the risk of systemic corruption to global knowledge bases increases. Researchers and publishers must develop new standards to detect AI-generated artifacts, ensuring that the foundational data used for discovery remains reliable and free from the invisible, enduring contamination of digital fossils.

Source: theconversation.com
Published on 2025-04-16

A weird phrase is plaguing scientific papers – and we traced it back to a glitch in AI training data

Related news