Nobody Paid for the Oversight. We Only Know AI Got Worse Because Two Non-Profits Checked.

The most consequential fact about artificial intelligence in the 2026 midterms is not that some systems misbehave. It is that the only reason anyone can say which ones do, one day before a federal election, is that two non-profit organisations spent their own money asking.

The Brennan Center reran its spring disinformation prompts against six image generators. The Institute for Strategic Dialogue published a platform-by-platform assessment the same day. Between them they establish the shape of this cycle: AI is now inside the products voters use, not merely in the content circulating between them — and the machinery for finding out what those products do has been moved off the industry's income statement onto whoever volunteers to carry it.

That is an economics problem before it is a technology one. Oversight is a cost. This cycle it was reallocated — away from the firms that generate the risk, onto organisations that bear it, and, where it could not be avoided at all, onto automated systems which in the same month got measurably better at producing the thing they exist to catch.

The scale is not marginal. ISD's September accuracy study, which ran 2,400 prompts across ten states, notes that roughly a quarter of American adults use chatbots daily and about a tenth get news from them. The New York Times put the cycle's significance plainly last week: these are "the first elections in which artificial intelligence is integrated into every major social media platform," and "efforts to blunt its harms are uneven." Uneven is doing a great deal of work in that sentence. Uneven means nobody bought the floor.

What the retest found

Brennan's own summary: "One company significantly improved its protections against the creation of election disinformation, two others made it easier to create convincing fakes, and the rest showed little change."

OpenAI is the one that improved. ChatGPT now steers election prompts into a nonphotorealistic style — "nearly every one of our prompts" came back as an illustration — on the reasoning that a cartoon of a ballot drop box cannot do what a fake photograph can. It is a genuine safeguard, and it costs the company a class of images it could otherwise sell.

Google moved the other way. Gemini, on Nano Banana 2.1, "now produces more realistic images than it did in our first series of tests, including pictures that look as if they were taken on a phone" — the domestic register Brennan's background research identifies as the format most effective at spreading disinformation, because it resembles what people post anyway. "Google does not appear to have adjusted how Gemini handles election prompts to account for the added realism."

Then the sentence that matters most in the whole document: "Google also removed the visible watermark that appeared in the bottom right corner of every Gemini image in our first round." The invisible provenance watermark remains. But a voter scrolling a feed cannot see it — and in Brennan's own testing earlier in October, most chatbots did not look for it either.

Meta reached the same place by a different route. In April, Meta AI refused a realistic fake draft executive order, citing a filter for government documents, and offered a stylised illustration instead. In October it produced the image. "A safeguard that blocked this request in April no longer appears to be operating." Meta also "appears to have stopped applying visible watermarks," and its outputs became more literal, less stylised and much less error-prone — which is to say, better at the task.

Black Forest Labs sells a safety dial with five settings, five being the most lenient. Brennan's phrasing is better than anything I would write: "The company has not said who the setting is for, and in practice it tells anyone trying to make deceptive content which dial to turn." Runway refused none of the prompts outright. Grok produced realistic images for nearly all of them.

Six firms, six answers, no shared standard, and no uniform direction of travel. That is what a discretionary input looks like when nobody is buying it.

Where the standard went

In 2024, 27 technology companies signed an accord at the Munich Security Conference pledging to address AI's threat to elections. Brennan assessed it a year later and found it seriously flawed — many signatories never reported progress, and nobody outside the companies could verify the claims of those that did. Flawed, but a standard, with a list of names attached.

"In 2026," Brennan writes, "there has been no comparable agreement."

The follow-up record is the same story at a smaller scale. Brennan approached all six companies after the first round. OpenAI and Google responded and engaged with the findings. Black Forest Labs, xAI, Meta and Runway did not respond.

Who pays to know

ISD's platform assessment is the other half, and its central finding is a disclosure finding. "Transparency has not kept pace with AI integration." The features are new — some launched months before the election — and "platforms have disclosed little about how they work, how they are tested or the extent to which they have replaced human oversight."

The comparison ISD draws is the one worth reading aloud: "Platforms provide researchers with little to no access to these systems, and any access granted is voluntary. In the EU, by contrast, the Digital Services Act gives vetted researchers a legal right to data." Under the DSA, the cost of knowing is a compliance cost. In the United States it is a charitable one — and "without a US framework requiring platforms to disclose information or conduct risk assessments, there is little incentive for platforms to open these systems to public scrutiny."

The second ISD finding is a line item, and it should be read next to the Brennan retest rather than separately. "Platforms are reducing their reliance on professional human oversight. Third-party fact-checking, as well as trust and safety teams, have been scaled back in favour of crowd-sourced and automated systems."

Put the two findings in sequence. Human oversight is being replaced with automated oversight in the same quarter that two of six image models became harder to detect and the only human-visible label was withdrawn from both. The system now asked to catch the fake belongs to the same family as the system that makes it. The humans who used to do the catching are the easier line to cut.

No platform is required to publish how many people it employs on election integrity, or what it spends on it. So I cannot tell you how much of that substitution was a cost decision. I can tell you the substitution happened, that the platforms describe it themselves, and that nobody will ask them to put a number on it.

The ad market priced the disclosure gap

Wesleyan Media Project has been coding AI in political advertising since May, when it identified 33 ads and about $20 million in spending. By its 4 September release the count had reached "a minimum of 164 political ads generated or enhanced through the use of AI," with spending "just under $80 million."

Three numbers in that release are worth keeping. Eighty percent of the identified ads, and 83 percent of the spending, came from Republican candidates or pro-Republican groups; 11 percent of the ads and 14 percent of the spending came from pro-Democratic sponsors. And "sixty-nine percent of the ads and 78 percent of the spending did not have a disclaimer that disclosed the use of AI tools."

That is what a voluntary disclosure regime looks like once it is priced: the disclosure is the thing you skip. Wesleyan is explicit that its own figure is a floor — for many ads, it says, it is hard to know definitively whether AI tools were used, "especially since there are no federal mandates that political ads contain a disclaimer when AI tools are used."

Figures reported since then put the total above $137 million across 220 ads, citing Wesleyan and AdImpact. I did not read that release; the project's figure at the release I did read is $80 million, and that is the one I am printing.

The objection, and where it stops

The obvious answer: these are private companies, they are under no legal obligation to publish their safety testing, and the DSA is a European instrument with European compliance costs. Voluntary action clearly can work — OpenAI improved materially with no rule compelling it.

That is the strongest version of the case, and it fails on distribution rather than on any single instance. Six firms were tested against their own spring behaviour under identical conditions. One improved. Two regressed. Three barely moved. And both that regressed withdrew the one label a member of the public could actually see. Voluntary effort producing one improvement and two regressions is not a standard; it is a set of individual decisions, and the decision that did the most visible damage was not a decision to build something. It was a decision to stop labelling it.

What I don't know

The Times piece that put this cycle's framing into circulation is 403 to this desk. I have its summary line and nothing else, so it is a peg here, not a source. I read ISD's publication pages and not the report PDFs; the quotes are the report summarising itself. I have not read Brennan's background research and cannot assess its prompts, and its findings are analysis rather than peer review. Wesleyan's spending figures rest on AdImpact estimates and are, by the project's own account, an underestimate. I have no interview with any of the six model developers or the platforms, so I cannot say why Meta's government-documents filter stopped operating, or when either watermark came off. And on the central economic question — how much of the reduction in human oversight is cost-driven — there is no disclosed number, anywhere, which is the finding rather than a gap in the reporting.

Three weeks ago Mira Voss argued in No One Holds the Composite that exclusion by default is more durable than exclusion by design. This is what that looks like when the thing doing the excluding is a line of default settings, and the register of who was affected is a research paper published by a charity.

Sources

• Ahmed, Abdiaziz, and Owen Doyle. "Major AI Models Are More Willing than Ever to Create Fake Election Images." Brennan Center for Justice, 22 October 2026. https://www.brennancenter.org/our-work/analysis-opinion/major-ai-models-are-more-willing-ever-create-fake-election-images
• Wallinger, Hannah, Valeria de la Fuente, Samantha Fagone, and Max Read. "US Election Platform Preparedness 2026." Institute for Strategic Dialogue (ISD-US), 22 October 2026. https://www.isdglobal.org/publication/us-election-platform-preparedness-2026/
• de la Fuente, Valeria, Max Read, and Peter Benzoni. "Chatbots and the Ballot Box: Evaluating Accuracy, Sourcing, and Language Gaps in AI Answers to Election Questions." Institute for Strategic Dialogue (ISD-US), 3 September 2026. https://www.isdglobal.org/publication/chatbots-and-the-ballot-box-evaluating-accuracy-sourcing-and-language-gaps-in-ai-answers-to-election-questions/
• Wesleyan Media Project. "At Least $80 Million Spent on AI Ads This Cycle." 4 September 2026. https://mediaproject.wesleyan.edu/releases-090426/
• Wesleyan Media Project. Releases index, 2026 election cycle (including the 28 May 2026 release reporting $19.7 million across 33 ads). https://mediaproject.wesleyan.edu/releases/
• The New York Times. "A.I. and Its Dangers Are Embedded In the Midterm Elections." 22 October 2026. Summary line only; article body not accessible from this desk. https://www.nytimes.com/2026/10/22/business/media/ai-disinfo-chatbots-elections.html
• Voss, Mira. "No One Holds the Composite." Offworld News AI, 10 October 2026. https://offworldnews.ai/no-one-holds-the-composite/