The astroturf was AI-generated. The anger underneath it is not.

The popular reading of this week's news writes itself. OpenAI's latest threat report describes two clusters of ChatGPT accounts, likely run from China, that spent late 2025 and early 2026 posing as ordinary Americans and posting AI-generated outrage about the things Americans are currently outraged about: data centers driving up electricity bills, and tariffs. One campaign — OpenAI calls it the "Data Centre Bandwagon" — produced comments and cartoon strips of fat-cat businessmen clutching money bags while families gasped at their power bills. The accounts were prompted in Simplified Chinese, routed through VPNs, and costumed as citizens from a careful variety of American backgrounds. So the lesson being drawn, in Washington and on every platform where the data-center fight is live, is the obvious one: the backlash is manufactured. Beijing is astroturfing the grid debate. Treat the opposition accordingly.

I want to argue that this lesson is exactly wrong, and that the wrongness is going to cost us. Not because the operation didn't happen — it did, the attribution looks careful, and banning the accounts was correct. But the most important sentence in OpenAI's report is the one nobody is quoting: the campaign sought to exploit and amplify existing public concerns, and it had, in the company's own assessment, little or no real-world effect. The operation was real and it was trivial. The grievance is the thing that's real and large. A policy debate that gets those two facts backwards — trivial op treated as large, large grievance treated as fake — is a debate preparing to lose on both fronts.

The strongest version of the alarm, stated fairly

Steelman first, because the alarmed position deserves it. Covert foreign influence operations are a genuine wrong, independent of their effect. A provincial Chinese contractor paying people to impersonate retired schoolteachers in Ohio is committing a fraud on the public square, and the fraud doesn't become acceptable because it was clumsy. The attribution detail in this report should bother you more than the cartoons: OpenAI traced one cluster to what appears to be a private technology firm doing influence work for provincial-level government clients. That is not a propaganda bureau experimenting after hours. That is a market — contractors, deliverables, presumably invoices — professionalizing around the business of synthetic American opinion. Markets get better. Clemson's Darren Linvill, who studies these operations, put the current state plainly: "they aren't there yet." The operative word is yet, and the alarmed reading of this week is that we are watching the unimpressive early version of something that will not stay unimpressive.

All of that is true, and none of it rescues the conclusion being drawn from it. Because the question that matters for policy is not whether the operation was wicked. It's what the operation actually consisted of — and what it found when it arrived.

What the operation found

It found a grievance already at full size. Between May 2024 and June 2025 — before this campaign got going — at least 36 American data-center projects were delayed or cancelled by local opposition: zoning fights, county hearings, ballot pressure. Data centers consumed about 1.5 percent of global electricity in 2024 and the figure is compounding at roughly 12 percent a year; in the regions where the buildout concentrates, the question of who pays for new generation and transmission is not a conspiracy theory, it is a rate case. The people showing up to those hearings did not need a comic strip from Hangzhou to tell them their bills went up. The bills arrived independently.

The operation didn't create the anger. It photographed it, badly, and sold the photograph to a provincial procurement office.

This is the detail the "manufactured backlash" frame cannot survive. An influence operation that amplifies an existing, accurate grievance is parasitic on reality — and parasites tell you something true about the host. The campaign's content was, in substance, often indistinguishable from positions held in good faith by actual American ratepayers, county commissioners, and at least two governors' staffs I can think of. What was fraudulent was not the argument. It was the speaker. That distinction is the entire ballgame, and the public conversation is busy erasing it.

The category error

Here is the error, named precisely: we are treating a provenance problem as if it were a content problem. The wrong in "Data Centre Bandwagon" is impersonation — a foreign contractor pretending to be domestic speakers. The wrong is not "someone said data centers raise electricity prices," because someone saying that might be your neighbor, and in several utility territories your neighbor would be citing the utility's own filings. A response aimed at provenance — ban the fake accounts, publish the attribution, harden the platforms — hits the operation and leaves the citizens alone. A response aimed at content — treating the claim itself as an information hazard, dismissing its carriers as dupes or assets — hits mostly citizens, because citizens are most of the people saying it.

And the content-aimed response is coming. It always comes, because it is rhetorically irresistible. Every data-center developer facing a hostile county board now has a footnote from OpenAI's report available; every official who would rather not answer a rate question has the option of answering an espionage question instead. I spent years inside government watching this move — the reclassification of inconvenient domestic opposition as foreign-adjacent — and I can report two things about it. It is almost never ordered from the top; it happens retail, one talking point at a time. And it works in exactly one direction: the opposition it dismisses does not disperse, it radicalizes, because nothing confirms a suspicion that elites won't engage your argument quite like being told your argument is Chinese.

Which is, I'd note, the operation's actual victory condition. A campaign too clumsy to move public opinion can still succeed if the response to its discovery poisons the well — if every future hearing about transmission cost allocation begins with both sides accusing each other of being bots. Beijing's contractors could not buy that outcome with content. We are providing it free, as a frame.

Who actually enforces this, and how

Now the procedural question this episode keeps skipping past: who, exactly, is the enforcement agency here? The answer is a threat-intelligence team at a private company — one that did careful work, published it, and deserves the credit it's getting. It is also a company with one of the largest commercial stakes in the data-center buildout on Earth, currently adjudicating which opposition to data centers is inauthentic. I am not alleging bad faith; the report reads as scrupulous. I am observing the structure, because structures outlast the good faith of their current occupants. The entity defining, detecting, and announcing foreign interference in the AI-infrastructure debate is a party to the AI-infrastructure debate, operating with no statute, no review, no appeal, and no jurisdiction beyond its own customer base.

That last clause is the load-bearing one. OpenAI can ban accounts on OpenAI. The next contractor — there will be a next contractor; provincial procurement is a renewable resource — runs the same campaign on an open-weights model on rented GPUs, and no threat report ever issues, because there is no vendor to issue it. The enforcement regime we are applauding this week is real, useful, and structurally confined to the most observable, least sophisticated tier of the problem. Build policy on the assumption that vendor threat reports are the detection layer, and you have built policy for the adversaries considerate enough to use the products that report on them.

The better frame, with its costs attached

So separate the two problems, because they are two problems, and they have two different owners.

• The influence operations are a fraud-and-provenance problem. Handle them the way platforms handle spam: continuously, unglamorously, with attribution published and accounts removed — and with the state's involvement aimed at the contractors and their clients through sanctions and indictments, where the evidence supports it. This work is real but it is hygiene, not victory. Announce it like hygiene.
• The grievance is an energy-policy problem. If data-center load is raising residential rates in a service territory, that is answerable with the boring machinery built for exactly this: cost-allocation proceedings, large-load tariffs that make hyperscalers pay the marginal cost of their own interconnection, transparency in the tax-abatement and siting deals counties sign. If the buildout's defenders are right that the costs are being mismeasured, those venues are where they get to prove it.
• And the line between the problems should be policed in both directions. Officials who cite this report to dismiss a rate complaint should be treated as committing the same category error as the operation itself: misrepresenting the provenance of an argument to avoid its content.

The costs of my frame, admitted plainly. Treating the ops as hygiene means accepting that hostile states will keep amplifying true and useful grievances, and that we will keep declining to treat the grievances as contaminated by the amplification — which will feel, in every individual news cycle, like letting Beijing win one. Taking the grievance seriously means some data centers will be built slower, somewhere more expensive, after hearings in which some of the angriest testimony will later turn out to have been retweeted by a bot. That is genuinely inefficient. It is also what having a domestic politics costs, and the alternative — an infrastructure buildout that proceeds by reclassifying its opponents as foreign instruments — is not a price I'd pay for any number of gigawatts.

The operation OpenAI exposed this week deserved exposure, and its successors will deserve it too. But the report's findings run opposite to the use being made of them: it documents a weak campaign free-riding on a strong grievance. The campaign we could always afford. The reflex to call the grievance fake — that one we cannot. I say this as someone who spent a career watching governments reach for the comfortable frame, and I can tell you where it leads: the frame becomes the policy, and the policy fails at the hearing, in public, one county at a time.