More than a dozen people inside an OpenAI office in San Francisco are pulling around-the-clock shifts this week, trying to keep ChatGPT from being weaponized to disrupt Tuesday's midterm elections. Anthropic, the company behind the Claude chatbot, has stood up its own election war room for the first time, splitting analysts into teams based in San Francisco, New York and Washington. The moves come as OpenAI, Anthropic and other tech companies have pledged to protect the vote from efforts to undermine democracy.
The San Francisco-based effort is led by Josh Lawson, OpenAI's head of civic and strategic risk, according to The New York Times. The Times reports that what it calls the company's operations center represents a vast expansion from OpenAI's comparatively nascent efforts protecting the 2024 presidential election, when ChatGPT had only a fraction of today's roughly one billion users and the underlying AI tools were far less powerful. Lawson told the Times that in 2024, text was the most readily available technology for bad actors to exploit; now, he said, the threats extend well beyond text monitoring alone.
Separate reporting from The Hill describes OpenAI's Election Operations Center in broader terms, saying it is made up of hundreds of employees, including the company's top investigators and specialist teams. OpenAI says the center is meant to bring its teams together to respond quickly to emerging issues and enforce company policy. The two accounts describe different scopes of the same effort, and it is not clear from available reporting how the figures relate to one another.
What Threats Are These Teams Watching For
The Times frames the potential dangers broadly: deepfakes, cyberattacks, and foreign and domestic influence operations that can now be created and launched with little more than a series of prompts. Those concerns are not hypothetical. According to CNN, private firms tracking online activity reported that suspected Russian operatives recently posted a wave of AI-altered videos on X attacking Democratic candidates, and U.S. intelligence officials are tracking efforts by Russia, China and Iran to influence American voters ahead of the midterms.
Domestic political players have also turned to AI-generated content this cycle. Per Reuters, the National Republican Senatorial Committee ran an AI-generated ad depicting Texas Senate candidate James Talarico reciting old social media posts, and the Republican Committee for Loudoun County released three AI-generated ads attacking Virginia Governor Abigail Spanberger back in February.
OpenAI's Specific Safeguards
OpenAI says its Democracy Works integration is now live across languages in the United States, feeding ChatGPT users information on voter registration, deadlines and polling locations. The company also says it has offered U.S. voting-system manufacturers and state election authorities free, special access to its models for cyber defense through a program it calls Daybreak. For image requests that touch on election interference or that could discourage voting, OpenAI says it applies stylization so the output comes back as a cartoon-style illustration rather than a photorealistic image. The company also says it continues to prohibit political ads on its platform and bars use of its products to create or distribute scaled campaign messaging for or against candidates, parties or ballot measures, details the company laid out in its own account of its 2026 election safeguards.
Google and Meta have likewise pledged robust efforts to stop their AI tools from being used to undermine confidence in the democratic process, the Times reports. Even so, all of the companies have faced criticism for not doing enough to limit malicious content as AI becomes more deeply woven into online social and political life.
Independent Testing Finds Gaps
Outside researchers are not convinced the safeguards are airtight. The Brennan Center says AI tools have grown far more sophisticated and widely available since the 2024 presidential race, and its own testing found that chatbot models sometimes suggested modifications that would let users bypass the platforms' own restrictions on election-conspiracy content, according to the Brennan Center for Justice. In the group's test of four chatbots, only Grok produced the full target of 100 prompts for a potential election-misinformation campaign, though the Brennan Center found that ChatGPT, Claude, Gemini, Grok, Flux.2 and Runway all ultimately agreed to generate images from those prompts that could be used to spread election misinformation.
Accuracy on basic voting logistics is also uneven. Data collected in October by the Information Futures Policy Group, reported by NPR, found that 13.9% of chatbot answers to voting-logistics questions were outdated or inaccurate, with errors in 5.9% of answers severe enough that they could mislead a voter about deadlines, election dates or required steps. Broken down by platform, the group found Claude gave wrong or outdated information one in three times, Copilot one in four times, and ChatGPT one in 10 times.
The idea of a tech-platform election war room is not new. Semafor traces the concept back to 2018, when Meta stood up the first such operation amid concerns over election-related misinformation on its platforms. Nearly a decade later, AI companies are adapting that same model for a landscape where synthetic video, audio and images can be produced in seconds rather than assembled by hand.
