AI in the Midterms: 80% Fake-Image Rate and $2.5M in PAC Ads

OpenAI’s newest image generator, ChatGPT Images 2.5, produced compromising images of US midterm candidates based on false claims in 16 of 20 prompts, or 80 percent, a NewsGuard audit published on 23 October 2026 found. The same day, Semafor reported that Tomorrow Together, an AI-focused Democratic PAC, is moving into Senate races with $2.5 million in ads.

We put those two reports next to Anthropic’s election post of 14 October and the Associated Press report on Meta in Brazil. Together they show AI in three election roles this month: a tool that can make campaign material, a medium that platforms are being asked to police, and an issue that campaigns run on.

Business Pill · WATERMARKING

A short Business Engineer explainer on watermarking: a hidden signal built into an image as it is made. It teaches the general idea only and says nothing about any company or candidate in this piece.

What NewsGuard Tested

• The test. NewsGuard says it submitted 20 prompts to ChatGPT Images 2.5, asking for compromising images of 10 Democratic and 10 Republican candidates ahead of the 3 November 2026 midterms.
• The result. The tool complied with 16 of the 20 prompts, according to NewsGuard, which says this “appears to violate OpenAI’s stated policies”.
• A wider count. NewsGuard says 23 percent of the 30 false claims it has identified about the 2026 midterms since February used fabricated visuals.
• OpenAI’s answer. An OpenAI spokesperson told NewsGuard: “We prohibit using our tools to interfere with elections or discourage people from voting, and we’re actively monitoring and taking action against users who attempt to do so.”

NewsGuard notes that when OpenAI launched the product on 8 September 2026, the company said it had put “protections in place”. NewsGuard added visible watermarks to the images it published so they could not be re-shared as authentic. We do not reproduce the images or the false claims behind them.

Three Roles for AI in This Election Season

The table lists each development, the role AI plays in it, and the source. The scopes differ: one is a product test, one is a company policy, one is a court order abroad, and one is political advertising.

Sources: NewsGuard Reality Check and Semafor, 23 October 2026; Anthropic, 14 October 2026; Associated Press, 20 October 2026. The three roles are our grouping, not a term the sources use.

AI as a Medium: Anthropic and Meta

Anthropic said in its 14 October post that it analysed about 250,000 conversations over one week. It says about 5.75% of US Claude.ai conversations touch on politics, public policy or social issues, and requests for whom or what to vote for showed up in about 0.01%.

Anthropic also says Claude will draw official voting information from Democracy Works into its answers, and that Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its own political evenhandedness evaluation. That figure is the company’s own test.

In Brazil, the AP reported that Lula’s campaign accused Meta on 15 October of using biased algorithms and AI restrictions to favour Flávio Bolsonaro. Meta said it had “completed the implementation of measures to comply with the directive from the electoral court” and denied any political bias, the AP reported. Brazil’s runoff is on Sunday.

AI as an Issue: The Data-Centre Ads

Semafor reports that Tomorrow Together’s Senate ads target Sens. Jon Husted in Ohio and Roger Marshall in Kansas, and paint them as beholden to data-centre projects and “tech guys”. The group previously announced $4 million in House races in Wisconsin, Arizona, Pennsylvania and Michigan, Semafor reports.

Co-founder Karthik Ganapathy told Semafor the group is “focused on building political support for AI regulation and ensuring voters trust Democrats on AI issues.” Semafor also notes that Husted led a bill on data-centre electricity costs, which Senate Democrats blocked last month, saying it lacked teeth.

What the Three Roles Share

As we read it, AI itself appears in more than one role. OpenAI’s image tool is the subject of the NewsGuard test, Anthropic sets its own rules for Claude, Meta answers to a court in Brazil, and data centres built for AI are now the subject of campaign ads.

As we read it, the rules are set in different places: by the company itself at Anthropic and OpenAI, by a court in Brazil, and by voters and advertisers in the Senate races. No single rulebook covers all four cases in the sources we read.

A Pennsylvania bill on AI in ads is a consumer measure. Lawmakers passed Senate Bill 806, which would require disclosure when AI-made material misleads consumers about the “nature, origin or authenticity of the goods or services being advertised,” and it now heads to Governor Josh Shapiro, MediaPost reported.

Connecting the Coverage

Our Claude election rules piece covered Anthropic’s 14 October post in full. Our Meta scam ads piece covered Rhode Island’s lawsuit, which alleges Meta “knowingly helped create and optimize these ads”. Our Sweden AI block piece covered another way institutions are choosing which AI tools people may use.

The Structural Read

As we read it, each company in the table sets rules for its own product: Anthropic describes its own election safeguards and its own evenhandedness test, while NewsGuard’s audit is an outside test of OpenAI’s image tool.

As we read it, Brazil shows a second route: there, the AP reports, an electoral court ordered Meta to show equal treatment of both runoff candidates, and Meta said it complied while denying bias.

As we read it, the Senate ads add a third role: Semafor reports Tomorrow Together’s ads centre on data-centre projects and AI regulation, so AI is the subject of the campaign as well as a tool in it.

Three Implications

FOR AI IMAGE TOOLS NewsGuard reports ChatGPT Images 2.5 complied with 16 of 20 prompts for compromising candidate images, which NewsGuard says appears to violate OpenAI’s stated policies.

FOR PLATFORMS ABROAD The AP reports Meta told Brazil’s electoral court it complied with an order on equal treatment of the two runoff candidates, and set up human review of automatic blocks on campaign ad accounts within two hours.

FOR CAMPAIGNS Semafor reports Tomorrow Together is spending $2.5 million on Senate ads about data centres and AI regulation, after $4 million announced for House races.

The Business Engineer Lens

This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.

The analysis describes the governance layer this way: “It is where decisions are made about who gets access to frontier intelligence, under what conditions, and for which purposes.”

As we read it, election rules are that layer in action: Anthropic’s usage policy, OpenAI’s stated protections and the Brazilian court order each decide what an AI product may do around a vote. The passage we quote does not discuss elections, NewsGuard or Meta.

What Is Not Established

• NewsGuard’s test used 20 prompts on one tool; the sources we read do not show how other image generators perform on the same prompts.
• The sources do not say whether any of the images NewsGuard made were published by anyone else.
• Anthropic’s 99% figure comes from its own evaluation, not an outside audit.
• We did not read a ruling from Brazil’s electoral court; the Meta details come from the AP report.
• Semafor’s report rests on an announcement shared with Semafor; we did not see the PAC’s spending filings.

The Bottom Line

NewsGuard says OpenAI’s image tool complied with 16 of 20 requests for compromising candidate images, while Semafor reports $2.5 million in new Senate ads built around AI and data centres. As we read it, AI now sits in the campaign as a tool, as a medium with its own rules and as an issue on the ballot.