Saving the Pre-AI Internet
What sunken battleships and the em dash have in common
There's a new type of preservationist, only this time they're not out to protect untouched forests or historic buildings. These conservators are out to save the internet.
This isn't the first time in history we've faced a before-and-after tipping point. After World War II, much of the world's new steel contained trace amounts of radioactive isotopes from nuclear testing. It posed no risk to structural integrity, but it was enough to interfere with sensitive scientific and medical instruments. The search for pure steel was on, and it was found at the bottom of the ocean. Pre-war shipwrecks were pulled from the deep to salvage steel made before the atomic age.
As Alistair Barr writes in Business Insider, we're facing something similar today. Our threat may not seem as extreme. After all, what's a few extra uses of the word "delve" or the em dash? Nothing, unless you realize both have become telltale signs that what you're reading (or publishing) was written by AI. They're curious fascinations of artificial intelligence and not nearly as common in natural human writing. (Don't lie. You don't really know how to use an em dash properly either.)
AI models are increasingly trained on content that includes AI-generated text. The problem is that when AI learns from AI, we risk what researchers call "model collapse." It's no different than photocopying a photocopy until the original becomes unrecognizable. We're already at the point where it's hard to tell whether a source reflects original human thought or a machine's.
Again, this may seem like a non-threat. But when it starts to seep into medical research, treatment approaches or the engineering standards that determine whether bridges stand or fall, it becomes very real, very fast. In the mental health world I've worked in, the risk is already here. AI is increasingly being used in therapy and crisis support, often without a human in the loop to catch errors or provide genuine human connection.
That's why preservationists are searching for pre-2022 internet content. Projects like LowBackgroundSteel.ai are cataloging datasets, code repositories and websites from before we relied on ChatGPT for everything.
May be a good time to dust off your ol' Myspace page and remember how you used to write.