In the post-nuclear age, scientists noticed a peculiar problem: steel produced after 1945 was contaminated. Atomic bombs had infused the atmosphere with radioactivity, which contaminated the metal.
This made most steel useless for precise equipment such as Geiger counters and other highly accurate sensors. The solution? Salvage old steel from sunken pre-war battleships resting deep on the ocean floor, far away from the nuclear fallout. This material, known as low-background steel, became prized for its purity and rarity.
Fast forward to 2025, and a similar story is unfolding — not under the sea, but across the internet.
Since the launch of ChatGPT in late 2022, AI-generated content has exploded across blogs, search engines, and social media. The digital realm is increasingly infused with content not written by humans, but synthesized by models and chatbots. And just like radiation, this content is tricky for regular folks to detect, is pervasive, and it alters the environment in which it exists.
This phenomenon poses a particularly thorny problem for AI researchers and developers. Most AI models are trained on vast datasets collected from the web. Historically, that meant learning from human data: messy, insightful, biased, poetic, and occasionally brilliant. But if today’s AI is trained on yesterday’s AI-generated text, which was itself trained on last week’s AI content, then models risk folding in on themselves, diluting originality and nuance in what’s been dubbed “model collapse.”
Put another way: AI models are supposed to be trained to understand how humans think. If they’re trained mostly on their own outputs, they may end up just mimicking themselves. Like photocopying a photocopy, each generation becomes a little blurrier until nuance, outliers, and genuine novelty disappear.
This makes human-generated content, from before 2022, more valuable because it grounds AI models, and society in general, in a shared reality, according to Will Allen, a vice president at Cloudflare, which operates one of the largest networks on the internet.
This becomes especially important as AI models spread into technical fields, such as medicine, law, and tax. He wants his doctor to rely on content based on research written by human experts from real human trials, not AI-generated sources, for instance.
“The data that has that connection to reality has always been critically important and will be even more crucial in the future,” Allen said. “If you don’t have that foundational truth, it just becomes so much more complicated.”
Paul Graham’s problem
Joe Corrigan/Getty Images for AOL
This isn’t just theoretical. Problems are already cropping up in the real world.
Related stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
Almost a year after ChatGPT launched, venture capitalist Paul Graham described searching online for how hot to set a pizza oven. He found himself looking at the dates of the content to find older information that wasn’t “AI-generated SEO-bait,” he said in a post on X.
Malte Ubl, CTO of AI startup Vercel and a former Google Search engineer, replied, saying Graham was filtering the internet for content that was “pre-AI-contamination.”
“The analogy I’ve been using is low background steel, which was made before the first nuclear tests,” Ubl said.
Matt Rickard, another former Google engineer, concurred. In a blog post from June 2023, he wrote that modern datasets are getting contaminated.
“AI models are trained on the internet. More and more of that content is being generated by AI models,” Rickard explained. “Output from AI models is relatively undetectable. Finding training data unmodified by AI will be tougher and tougher.”
The digital version of low-background steel
Tyler Miller/Sportsfile for Web Summit via Getty Images
The answer, some argue, lies in preserving digital versions of low-background steel: human-generated data from before the AI boom. Think of it as the internet’s digital bedrock, created not by machines but by people with intent and context.
One such preservationist is John Graham-Cumming, a Cloudflare board member and the company’s CTO.
His project, LowBackgroundSteel.ai, catalogs datasets, websites, and media that existed before 2022, the year ChatGPT sparked the generative AI content explosion. For instance, there’s GitHub’s Arctic Code Vault, an archive of open-source software buried in a decommissioned coal mine in Norway. It was captured in February 2020, about a year before the AI-assisted coding boom got going.
Graham-Cumming’s initiative is an effort to archive content that reflects the web in its raw, human-authored form, uncontaminated by LLM-generated filler and SEO-optimized sludge.
Another source he lists is “wordfreq,” a project to track the frequency of words used online. Linguist Robyn Speer maintained this, but stopped in 2021.
“Generative AI has polluted the data,” she wrote in a 2024 update on coding platform GitHub.
This skews internet data to make it a less reliable guide to how humans write and think. Speer cited one example that showed how ChatGPT is obsessed with the word “delve” in a way that people never have been. This has caused the word to appear way more often online in recent years. (A more recent example is ChatGPT’s love of the em dash — don’t ask me why!)
Our shared reality
As Cloudflare’s Allen explained, AI models trained partly on synthetic content can accelerate productivity and remove tedium from creative work and other tasks. He’s a fan and regular user of ChatGPT, Google’s Gemini, and other chatbots such as Claude.
And just like human-generated data, the analogy to low-background steel is not perfect. Scientists have developed different ways to produce steel that use pure oxygen.
Still, Allen says, “you always want to be grounded in some level of truth.”
The stakes go beyond model performance. They reach into the fabric of our shared reality. Just as scientists trusted low-background steel for precise measurements, we may come to rely on carefully preserved pre-AI content to gauge the true state of the human mind — to understand how we think, reason, and communicate before the age of machines that mimic us.
The pure internet is gone. Thankfully, some people are saving copies. And like the divers salvaging steel from the ocean floor, they remind us: Preserving the past may be the only way to build a trustworthy future.
Sign up for Business Insider’s Tech Memo newsletter here. Reach out to me via email at abarr@busienssinsider.com.
