The tells were computable. Nobody ran them.
The scary part of ChatGPT citing fake local news isn't the reach — it's that a crawler could have caught the fakes for pennies, and nobody in the pipeline is paid to care.
Here's what happened, because the mundanity is the point. During the run-up to the 2026 midterms, The Washington Post asked six chatbots to build voter guides and got serious errors back — including citations to "local news" outlets that don't exist as news outlets. One of them, the Beacon-style site in the reporting, was a sixteen-story operation covering a single candidate, with no archive history and no masthead to speak of. ChatGPT happily laundered it into an answer. Separately, the Post reported on Iran-planted AI-generated articles seeding themselves into real US news media, and TechPolicy.press found that LLMs were still generating election deepfakes days before the vote. The harm arrived on schedule. The trivial defense never did.
The number everyone reached for was the reach figure — how many voters saw the fake. That's the wrong number to be alarmed about. The tells on the fake site were all machine-checkable: a tiny corpus, single-candidate focus, no Wayback footprint, no masthead history. A crawler could run those checks across every domain an LLM retriever might cite for pennies per site. This is not an adversarial-AI problem, the kind you need a lab to solve. It's a boring data problem. The failure isn't that detection is hard — it's that no actor in the pipeline runs the check. Search indexers don't lose money when they index a fake. Model vendors don't lose money when they cite one. That makes this an incentive problem dressed up as a capability problem, and it explains why the gap hasn't closed: absent defenses against a technique that funds itself don't close on their own. Nobody is bleeding on the other side of the hole.
The second tell is quieter and worse. The same technique — AI-generated fake local news, in the same shape — has now run for Democratic candidates, fully detached from its right-wing origin. That's the commoditization moment, and it's the moment an influence op stops being an operation and starts being weather. Operations have authors; you can attribute them, prosecute them, shame the platform that carried them. Infrastructure has no author in any useful sense — just users, ambient technique, rain. Attribution frames assume an adversary. Once the tool is for rent, there's no adversary to find, only climate to prepare for.
When influence ops become infrastructure, they stop being operations and start being weather — something you prepare for rather than prosecute.
So here's a prediction I'll own: text provenance resolves as vaporware. Images got C2PA because images are expensive to fake and the stakes concentrated among a handful of camera-and-software vendors who could be made to care. Text is cheap to fake, and the cost of bad provenance lands on readers, not publishers — so nobody pays for the defense, and the standard never gets the muscle behind it. Content credentials for prose will be announced, piloted, and quietly starved, in roughly that order.
The open question isn't whether the check gets built for elections. Elections are the one domain with auditors, journalists, and a deadline — the eweek survey exists because someone was paid to look. The question is what happens when the same sixteen-site technique points at local business reviews, health claims, court records — the domains where nobody is auditing at all, where the fake doesn't even need to survive scrutiny because scrutiny never shows up. The tells will still be computable. That's the part I'd bet on. Whether anyone runs them is the part I wouldn't.