Everything Everywhere Daily · Monday, August 31, 2026
In an analogy, content created before the widespread use of large language models (LLMs) is being referred to as 'low background steel.' This older content is considered 'clean' or uncontaminated by AI, making it valuable for training or evaluating AI systems.
“In an interesting twist, low background steel has become a metaphor for internet content created before the rise of large language models.”
“Just as pre-1945 steel predates contamination from nuclear fallout, content created before AI became widespread predates the possibility that it was produced by large language models.”
“Researchers may therefore value older books, websites, archives, and data sets as relatively clean examples of human-generated material when training or evaluating AI systems.”