AI Election Safeguards: What Four Labs Published Before the November 3 Vote

Anthropic sampled 250,000 US Claude.ai conversations in late September and found that about 0.51% were about elections or voting. Three weeks before polls opened, it published that number alongside its strengthened AI election safeguards. OpenAI, Google and Meta published their own rules earlier in 2026. Read side by side, the policies agree on far more than they differ. What none of them offers is a way for you to check that the rules were enforced.

The short answer. For the 2026 US midterms, the four largest AI and platform companies published broadly similar AI election safeguards. They share three elements: official voting data from Democracy Works or election officials, bans on impersonating candidates and suppressing turnout, and disclosure rules for AI-made political ads. They differ most on images of candidates and on political advertising. None publishes verifiable enforcement numbers.

How big the AI election problem actually is on one chatbot

The most useful number published this cycle came from Anthropic, and it is small. Its 14 October post on preparing for the midterms reported a random sample of 250,000 US Claude.ai consumer conversations from 21 to 27 September 2026.

About 5.75% of those conversations touched politics, public policy or social issues. About 0.51% were specifically about elections or voting. About 0.01% asked Claude whom or what to vote for. A separate review of 250,000 conversations against the election policy found that about 0.002% violated it.

Applied to a 250,000 sample, 0.51% is roughly 1,275 conversations about elections. The 0.01% share is about 25 people asking for a voting recommendation. The 0.002% violation rate is about 5 conversations in the separate sample of 250,000.

Those counts are our arithmetic on Anthropic's rounded shares, not figures Anthropic printed.

A defender of the labs reads this as proof the risk is overstated. A critic reads it as risk hiding in volume, and we lean that way.

At 5 violations in 250,000, a classifier that is 99% accurate would still flag about 2,500 innocent conversations for every real one. That is a property of rare events, not a claim about Anthropic's classifier, whose accuracy is unpublished. It is the main reason enforcement claims cannot be judged from policy text alone.

AI election safeguards compared, cell by cell

Every cell below comes from a page the company published itself. Where a policy is silent, the cell says so instead of guessing. Silence is itself a finding, because nobody can be held to an unwritten rule.

Sources: Anthropic, 14 October 2026 and 8 October 2026. OpenAI election page, May 2026, and ChatGPT Images 2.5 system card, 8 September 2026. Google, 9 September 2026. Meta, 19 February 2026, updated 1 June and 27 October 2026. Full links are in the references. Read 3 November 2026.

Read the last two rows first. They describe how you could check whether each company kept its word, and they are nearly empty.

Voting information: the one place the labs converged

On how to vote, the four companies landed in almost the same place: official or near-official data instead of a model improvising a polling place.

Anthropic said that from late October Claude would draw "official voting information from Democracy Works into its responses on Claude.ai for users in the US". Google's 9 September midterm post names the same nonprofit alongside state and local government data. OpenAI's 2026 election page also uses Democracy Works for registration, deadlines and polling locations.

That convergence helps voters and creates a single dependency nobody voted for. If one nonprofit's dataset carries an error for one county, three assistants repeat it. We still think the trade is right, because the alternative is three models guessing independently.

OpenAI and Google both name The Associated Press as the source of live vote counts. Anthropic says Claude will show "up-to-date results on election night and in the days that follow" but does not name its provider. It is a small omission that matters on a night when counts move hourly.

Meta's February election post says state and local officials have sent more than 1 billion notifications through Voting Alerts on Facebook. That is a feed feature, not a chatbot answer, and the post is silent on how Meta AI handles a where-to-vote question.

Candidate imagery: where published policy and tested behaviour split

Images of candidates are where AI election safeguards are easiest to write and hardest to keep. Every company forbids deceptive depictions in some form. Only one was tested publicly against its own rule this autumn, and it failed more often than it passed.

OpenAI wrote the rule and NewsGuard tested it

OpenAI's system card for ChatGPT Images 2.5 is dated 8 September 2026. It says the model's realism could "allow more convincing deepfakes, including political, sexual, or otherwise sensitive imagery of real people". It adds that such images violate OpenAI's usage policies and that safeguards sit at the prompt and image layers.

On 23 October, NewsGuard reported that the same tool fulfilled 16 of 20 prompts asking for compromising images of 10 Democratic and 10 Republican candidates. It refused 3 and drew the wrong person once. NewsGuard did not publish its test dates.

OpenAI's response, quoted by NewsGuard, was that it prohibits election interference, is "actively monitoring and taking action," and that images carry SynthID watermarks that can be checked publicly. That answer moves the safeguard from refusal at creation to detection after the fact.

OpenAI's card reports that, in its adversarial "Political" test category, its production model was safe on about 90% of prompts. OpenAI warns those prompts are not representative of real traffic, and neither are NewsGuard's 20. Neither figure tells you how often a real candidate deepfake leaves ChatGPT.

Google's post promises "additional protections to limit the creation of misleading election-related images" and gives no rule you could test. Anthropic's pages do not discuss images, and Tech Policy Press noted in June that Claude does not generate images, video or audio. Anthropic's exposure here is text that impersonates a candidate, which its Usage Policy bans.

How platforms label what does get made is covered in the comparison of AI content label rules across platforms.

AI political ads rules and campaign use

Here the four companies differ in kind. Two sell political ads at scale, one chose not to sell them in its chatbot, and one has no ad product in this context.

Google and Meta: disclose, label, archive

Google requires election advertisers to pass verification and to disclose synthetic or digitally altered content. Its July 2026 ads policy change log says election advertisers should keep ticking the "Altered or synthetic content" box for every such ad. Election ads carry a "paid for by" line and appear in Google's transparency report.

Meta takes the same disclosure approach and adds detection. Since a 1 June update, ads it detects were made with third-party AI tools carry extra information in an "About this ad" panel. Meta also says it has more than 18 million US entries in its Ad Library, kept for seven years.

On 27 October at midnight Pacific time, Meta started its restriction period, blocking new political, electoral and social issue ads for the final week. Ads that had run at least once before then could keep running. That exemption stops new, unanswerable claims while letting everything already in market continue.

OpenAI: no political ads in ChatGPT this cycle

OpenAI's election page says ChatGPT will carry no political advertising during the 2026 election period. Gigazine reported this on 28 May and Tech Policy Press in June. The page also prohibits scaled campaign messaging for or against a candidate. Opting out is the cleanest policy in the table, and it says nothing about ads elsewhere that ChatGPT helped write.

Anthropic: rules on use, not on ads

Anthropic's 8 October Usage Policy update renamed its election section "Do Not Undermine Democratic Processes". It also removed a blanket ban on personalised vote and campaign targeting. The company cites nonprofits translating voter information and officials sending ballot cure notices as the legitimate uses the old rule blocked.

We think that change is right and underreported, because a rule that stops a county clerk drafting a ballot cure notice protects nobody. The replacement prohibitions on fake personas, hidden sponsorship and voter suppression target the conduct that actually does damage.

The enforcement gap in every AI election policy

Every policy in the table depends on enforcement inside the company. That is unavoidable. Publishing no number an outsider can check is not.

Anthropic describes classifiers for multi-step abuse, fake accounts and news sites, money-making schemes and foreign state actors. It describes one fake news site run by foreign operators that pushed Russian, Iranian and North Korean narratives. "We banned the accounts involved in these cases," the post says. It does not say how many accounts.

Meta gives the longest-running figure: 200 coordinated inauthentic networks removed since 2017. That covers nine years and every country, so it says little about AI misuse in this cycle. Google points to its quarterly Influence Operations Bulletin. OpenAI points to an Election Operations Center.

Red-teaming is not an audit

Anthropic went furthest on outside testing. The Institute for Strategic Dialogue and the McCain Institute tested Claude across the US, Israel and Brazil. They found that Claude "reliably refused a single harmful request, but could miss a malicious campaign" broken into harmless-looking pieces.

That is a candid finding about the model, not a check on enforcement. An audit would ask: of the accounts the classifiers flagged, how many were reviewed, how many were banned, and how many real campaigns were missed? No lab has published a framework that would let anyone answer it.

There are honest reasons for the silence. A precise ban count tells adversaries how much they were caught. A prevalence rate invites a comparison with rivals who measure differently. A third-party audit means giving outsiders access to user conversations, a privacy problem of its own.

None of those reasons rules out a delayed, aggregated report. Anthropic says it hopes "to share a more comprehensive look" at election use after the cycle. That is the right promise, and we would hold it to a date. The same disclosure gap shows up across the industry's safety reporting, as the scorecard of responsible AI reports found.

Where this comparison is weakest

This post compares published documents, a method with three limits, each of which could change a cell.

Policy pages are not product behaviour

A policy says what a company intends. NewsGuard showed how far behaviour can drift from intent. We ran no prompts of our own, so every cell describes the rule, not what the product does today.

We could not open one primary page directly

OpenAI's election page refused our automated retrieval. Its cells rely on the page as described by Gigazine in May and by Tech Policy Press in June, plus OpenAI's own system card. Secondary reports mention an October update to that page, which we could not read, so the OpenAI column may be out of date.

Silence is not the same as absence

Where we wrote "not addressed", we mean the election post does not say. Meta may have rules for Meta AI images of candidates in a help page we did not find. Anthropic may have an ad policy that does not apply to a product with no ads. The objection that we penalise companies for filing rules in a different drawer has real force, and we accept it for those two cells.

The audit row is different. If any of the four had committed to an outside audit of election enforcement, it would have said so in the post written to reassure voters.

What AI election safeguards mean for teams that deploy these models

If your team uses these models for campaigns, advocacy or civic communication, the vendor's election policy is part of your compliance surface. Most teams, in our view, have not read it. Your use is governed by their rule, not your intent.

Anthropic's updated policy takes effect on 12 November. It bans content that falsely appears to come from a candidate, election authority or news outlet, and it bans hidden sponsorship of persuasive messaging. A public affairs team using Claude for advocacy copy needs to know which side of that line its templates sit on. The broader contract terms are compared in the side-by-side of model provider terms.

Enforcement against you will be just as opaque. If a classifier flags your account during a campaign, you will learn about it from a suspension, not a report. Plan for it as a single point of failure. The governance controls each assistant exposes to admins are mapped in the enterprise governance comparison for ChatGPT, Claude and Gemini.

Synthetic media risk does not end with elections. Candidate imagery is the executive impersonation problem aimed at a different target. If NewsGuard could get compromising candidate images 16 times in 20, your finance team should assume a fabricated video of your CFO is easy to make too. The controls for that are set out in the payment controls guide for deepfake CFO fraud.

Frequently asked questions

Where to start if your team touches political content

Open the Usage Policy of every model your team uses for public-facing copy and find the election section. For Claude, it is "Do Not Undermine Democratic Processes", effective 12 November. Read it against your three most common templates and mark any that rely on undisclosed sponsorship or a persona.

Then put January 2027 in your calendar, when each lab will have had two months to publish what happened. A company that delivered its promised retrospective has earned some trust. One that did not has answered this post's question.

Related reading