AI Contamination: The Permanent Embedding of the Nonsense Term 'Vegetative Electron Microscopy'

A study reveals how the nonsensical term 'vegetative electron microscopy' became permanently embedded in AI systems. Originating from errors during the digitization of 1950s papers and amplified by translation mistakes, this phrase was learned and generated by large language models. This highlights the lack of transparency in AI model training data, the difficulty of correcting errors, and challenges to knowledge integrity. Researchers call for greater transparency in AI training data, improved peer review processes, and new ways to evaluate information in the age of AI-generated misinformation.