'Low-background steel' became a metaphor for pre-2022 web text uncontaminated by generative-AI output, catalogued at lowbackgroundsteel.ai

The metallurgical term low-background steel has been repurposed as an AI-era metaphor. In this usage, human-authored web text created before the late-2022 release of ChatGPT — and the ensuing flood of synthetic content onto the open internet — is prized the way detector builders prize pre-1945 steel: as a finite reserve uncontaminated by the new pollutant, here machine-generated text rather than fallout. The shared shape is "the only clean source is older than the contamination event."

The framing was catalogued by John Graham-Cumming (then Cloudflare's CTO), who — per a Tier-4 relay — "in March 2023 ... registered the web domain lowbackgroundsteel.ai and began posting about the concept." The usage circulates in AI-model-collapse commentary, e.g. The Register's 2025 coverage of AI "pollution" of training data. The metaphor usage is Tier-3-supported; the specific attribution and the March-2023 date are Tier-4 and unverified — routed to question-verify-lowbackgroundsteel-ai-domain-and-metaphor-origin, so the note is held seedling.

The metaphor's technical warrant is the model-collapse literature: training generative models on their own outputs erases the tails of the original distribution, which is what makes a pre-synthetic corpus valuable in the first place, and cultural-evolution accounts recast the same degradation as a language-transmission bottleneck. The analogy has a load-bearing limit, developed in claim-low-background-steel-need-is-fading-as-atmospheric-fallout-decays: atmospheric contamination stopped being injected after the 1963 test-ban and is decaying out, whereas synthetic-text contamination is still being added, so the "clean data" reserve does not replenish.