What it is
A hand-curated public catalogue of datasets/archives that predate, or aim to exclude, post‑2022 machine-generated content. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
Gabriel’s notes
Low-background Steel is a public, hand-maintained catalogue of “pointer” links to data sources (archives, dumps, datasets, projects) whose contents either predate late 2022 or are meant to deliberately exclude the post‑2022 flood of generative AI output. Submissions are accepted via a simple web form and reviewed/added by hand. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
Quick take: If you’re trying to build (or audit) a training/eval corpus and you’re tired of playing “is this human-made or a blender smoothie,” this catalogue is a pragmatic starting point. It won’t solve licensing, bias, or provenance by itself—but it will save you time and sanity. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
The name is doing some clever work. In the physical world, “low-background steel” (a.k.a. pre-war / pre-atomic steel) refers to steel produced before atmospheric nuclear fallout became a contamination concern for ultra-sensitive radiation detectors. The site borrows that concept as a metaphor for “pre‑AI” (or AI-excluded) data sources. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
A few operational details that matter:
- It’s explicitly curated by hand, with submissions welcome. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
- It’s maintained by John Graham-Cumming (Cloudflare), and the site notes it’s hosted on Cloudflare. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
- It’s not an endorsement list; it’s a claim about plausible “uncontaminated” status—not truthfulness, quality, or rights clearance. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
I saved this under AI because… the moment you start training, evaluating, or even just researching with modern corpora, you’re forced to think about contamination (synthetic text, SEO spam, model-on-model feedback loops). This catalogue is basically a map of “places to dig” where that problem is smaller—or at least more bounded. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
Good fit if you want to:
- Assemble a “pre‑2022-ish” reference corpus without reinventing the scavenger hunt.
- Find candidate sources for evaluation sets where you care about synthetic contamination.
- Build a provenance story you can explain to a client, auditor, or your future self.
- Get inspiration for what kinds of archives/dumps exist (and how people package them).
Pricing snapshot (auto-enriched)
Unknown / not confirmed as a formal policy page—but it presents as a free, public catalogue with a free public submission form (no paid tiers mentioned). ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
Work-use / compliance snapshot (auto-enriched)
This site is a directory of links, not a license umbrella. Practically: each linked dataset/archive will have its own terms (copyright, scraping rules, database rights, attribution requirements, downstream distribution limits, and possibly privacy constraints). The catalogue also emphasizes that inclusion isn’t an endorsement—so you still need to do your own diligence on legality, provenance, and suitability for your use case. Unknown / not confirmed whether the curator performs systematic license verification for every entry. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
Alternatives (auto-enriched)
- Common Crawl (Aug 2022 crawl): If you want “one big web snapshot” around the pre‑LLM-release era, this is more of a bulk raw ingredient than a curated list. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
- Chronicling America (Library of Congress newspapers): If you want historically bounded, high-signal text (1730–1969), this is narrower than a general catalogue—but extremely clean from a “post‑2022 contamination” standpoint. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/))
Before you adopt it:
- Write down your cutoff rule. “Pre‑2022” can mean “created before,” “finalized before,” or “captured before.” Pick one and be consistent. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/submit))
- Keep evidence. Save the specific dump IDs/snapshots you used (and dates), because “the internet” is not a stable artifact.
- Don’t confuse ‘not AI-generated’ with ‘safe to use.’ Rights, privacy, and contractual restrictions don’t disappear just because the content is human-made. ([lowbackgroundsteel.ai](https://lowbackgroundsteel.ai/about))
Sources
- https://lowbackgroundsteel.ai/
- https://lowbackgroundsteel.ai/about
- https://lowbackgroundsteel.ai/submit
- https://en.wikipedia.org/wiki/Low-background_steel
- https://apnews.com/article/72fd64095c789741e9ff14500c3121ae