We know how this story goes because we’ve already lived it once.

A “breakthrough” product gets marketed as a miracle: cheap, infinitely moldable, endlessly useful. It gets inserted into everything — utensils, appliances, cars, toys, clothing, packaging — until “everything” stops being a figure of speech and becomes a literal description of the material world. Decades later we discover our blood carries fragments of it. Our oceans have currents that function as trash-sorting machines. And still — even now, even having named the poison — we cannot stop producing it, because the thing that made it profitable in the first place, its insertion into everything, is the same thing that makes it impossible to remove.

That is the story of plastic. I think it is also, right now, in real time, the story of AI.

The insertion is the business model

Plastic didn’t win because it was better than wood or metal or bamboo at any one task. It won because a small number of petrochemical companies, sitting on a captive input, could make it cheaper than anything else and then sell it into every possible application at once. The generalization was the product. Once plastic was in your dashboard, your carpet, your food packaging, and your child’s toys simultaneously, no single ban — no paper straw, no tote bag — could touch the underlying architecture. You were fighting the applications one at a time while the source kept flowing.

AI is following the identical playbook, except the timeline has been compressed from decades to quarters. A handful of companies — OpenAI, Anthropic, Google — sit on the compute and the models the way oil majors sat on the wells, and the strategy is the same: get inserted into everything, simultaneously, before anyone can evaluate any single insertion on its merits. Customer service. Classrooms. Radiology suites. Unemployment offices. State government contracts. University partnerships. Not because AI is proven best-in-class at each of these, but because ubiquity is the moat. By the time we can measure the damage in any one domain, the technology is already load-bearing in the other twenty, and pulling it out looks less like a policy adjustment and more like a controlled demolition of institutions that have quietly rebuilt themselves around it.

We already have the receipts on the damage — and we’re still not stopping

With plastic, we didn’t lack knowledge. We had the Exxon Valdez, we had the Great Pacific Garbage Patch, we had micro plastics turning up in human blood, and we kept manufacturing anyway, because the dependency had already outrun the willingness to act on the evidence.

We are assembling the same kind of evidence base on AI right now, in real time, and treating it the same way. A 2025 multicenter study found that endoscopists who had grown reliant on AI-assisted polyp detection performed measurably worse once the AI was taken away — their adenoma detection rate had actually dropped in absolute terms after routine AI exposure, in what researchers called the first hard clinical evidence that AI reliance erodes a skill doctors need when the machine isn’t in the room. The same research found that the erosion is invisible in real time and will only become fully visible years from now, once a generation of clinicians discovers the foundation underneath their training was never actually built. Radiology is watching the identical pattern take shape, with controlled studies showing that flawed AI prompts push experienced breast-imaging radiologists toward more false positives, not fewer.

Meanwhile, the extraction problem — the equivalent of drilling — has its own paper trail. Court filings unsealed this year revealed that Anthropic ran an internal program, deliberately soft-codenamed to stay hidden, whose stated goal was to “destructively scan all the books in the world.” The company bought physical volumes in bulk from used-book sellers, ran them through a hydraulic cutting machine to remove the pages, scanned them, and then discarded the originals, at a pace that internal vendor proposals estimated at half a million to two million books over six months. A federal judge ruled the practice legal. Legal is not the same as consequence-free: those particular physical copies, some of them rare, are gone, in the same way a barrel of oil is gone once it’s been cracked into resin. The knowledge inside them survives only as compressed weights inside a handful of companies’ models — available to us now only through the very products built on their extraction.

This is the part that keeps me up. When the library at Alexandria burned, or when Spanish colonizers destroyed the codices of Mesoamerica, what was lost wasn’t just objects — it was the possibility of ever fully reconstructing what those objects contained, because the people who could have told you what was in them were also gone. We are doing something structurally similar, voluntarily, to ourselves, in the name of efficiency. And the generations coming up behind this transition will have no memory of the “before” to even miss it. There will be no contrast point. That’s not a hypothetical harm. That’s what makes it permanent.

The part of the plastic story we conveniently forget

Here’s where the metaphor should scare us and also, I think, where it should steady us, because plastic’s story isn’t only a story of failure. It’s also, eventually, a story of organized refusal — bag bans, straw bans, extended producer responsibility laws, all of it slow, all of it partial, but all of it real, and all of it built by people who decided that ubiquity was not the same thing as inevitability.

The organizing against AI’s physical infrastructure is following that same arc, faster than I expected. In the first quarter of 2026 alone, local opposition blocked or delayed roughly seventy-five data center projects worth an estimated $130 billion — matching the total for all of 2025 in a single quarter. That opposition is explicitly, deliberately cross-partisan: Republican officials raising alarms about grid strain and tax giveaways, Democrats focused on water use and environmental damage, and lawmakers on both sides warning that a handful of companies are accumulating an unprecedented concentration of capital and political power. Polling from earlier this year found that only about a third of Americans approve of the current pace of data center construction, and barely one in seven would welcome one in their own community — a lower tolerance than Americans show for nuclear power plants. And this isn’t confined to a few flashpoint towns: the number of active opposition groups has grown from under 400 to more than 800 in a matter of months, spreading to 49 states, with state legislators introducing more than 300 bills in the first half of the year to regulate the build-out.

That’s the fight actually being won right now — not because everyone opposing a data center agrees about AI’s promise or peril, but because the physical footprint is the one part of this system you can still stand in front of. You cannot organize a picket line outside a training run. You can organize one outside a substation.

So what do we do with a technology that’s already load-bearing

I don’t have a tidy ending, because plastic doesn’t have one either — we are still, right now, negotiating global treaties to try to slow production of a material we identified as catastrophic fifty years ago, and mostly losing. I don’t think AI politics resolves faster than that just because the harm is more legible sooner.

But I don’t think “already inserted everywhere” and “permanently inevitable” are the same claim, even though they get treated as synonyms by the industry that benefits from the conflation. Plastic bans that seemed symbolic — the straw, the bag — turned out to be load-bearing for a much larger regulatory architecture that came later: extended producer responsibility, recycled-content mandates, and now international treaty negotiations that would have been unthinkable in 1990. The bamboo fork was never going to solve microplastics. It was never supposed to. It was a foothold.

Data center moratoriums are today’s bamboo fork. They will not, on their own, un-insert AI from radiology, from classrooms, from the unemployment office. But they are proof, running in real time in forty-nine states, that ubiquity is not the same thing as consent, and that the people being asked to absorb the water use, the electricity bills, and the surveillance can still say no to the next substation even after they’ve already lost the last one. That is not nothing. In a fight where the other side is betting entirely on speed outrunning scrutiny, the fact that scrutiny is catching up at all is the most hopeful data point I’ve got.

Published by LA Progressive.