Inside AI giants’ Election Day ‘war rooms’

Artificial intelligence firms are on high alert for Election Day, taking a page out of social media firms’ playbook to manage the flurry of information expected online in the coming days.

After a dramatic few months for the AI industry, companies like OpenAI and Anthropic are under pressure to ensure their products are feeding out accurate information to voters.

And in the face of old and new cybersecurity threats, firms say they are also more closely watching for election interference attempts carried out with AI.

Here’s everything you need to know about how AI firms are preparing for Tuesday’s chaos, including “war rooms.”

In the two years since the last election, talk around AI and the technology’s benefits and risks to society skyrocketed in Washington and on the campaign trail. From policy debates to industry campaign spending to AI-generated ads, AI played a bigger role this midterm cycle than ever before on multiple fronts.

Research indicates voters are increasingly turning to AI chatbots or AI-generated search results when looking for election information, and they could be trusting these answers.

Like the launch of social media in the 2000s, the use of AI is raising questions over how technology influences the way voters stay informed and whether the public can trust everything they read from an AI chatbot or search result.

Lawmakers, many of whom are running for reelection, have called on AI firms and social media companies to ensure platforms’ accuracy.

In a letter to tech CEOs last month, Sens. Mark Warner (D-Va.) and Alex Padilla (D-Calif.) and Reps. Jim Himes (D-Conn.) and Joe Morelle (D-N.Y.) wrote, “You are aware of the great responsibility that your companies hold in America’s elections.”

“Your platforms play a significant and important role in sharing, disseminating and in some cases generating information about those elections,” the lawmakers said.

Four of the country’s largest AI firms spoke publicly in recent months on how they plan to handle the influx of information and changes in real time.

AI’s accuracy has drastically improved over the past few years, but skepticism still remains about its error rates. A study from the Institute for Strategic Dialogue in September found nearly a third of AI chatbots’ replies to election questions were incomplete, outdated or inaccurate.

The AI labs say they are focused on distributing accurate information, rather than trying to persuade a voter on their ballot decisions amid speculation on model bias.

“People should be able to access reliable information, understand the origin of AI-generated media that they might encounter off-platform and to explore kind of political questions without being steered in one particular direction by the model,” Josh Lawson, OpenAI’s head of civic and strategic risk, told reporters last month.

Both OpenAI and Anthropic are partnering with Democracy Works, which compiles information on registration, deadlines, voting locations and updates throughout the election process.

This information is used to form ChatGPT’s responses and provides links to official election resources for voters, OpenAI said. The company is also partnering with The Associated Press for real-time voting results.

For election queries to Anthropic’s models, the company features a banner linking to TurboVote, Democracy Works’s nonprofit and nonpartisan resource for voter guidance.

Lawson said AI’s search function has “significantly improved” to provide more up-to-date information since 2024.

“Whenever you’re asking a question, if it’s about politics or any type of civic area, the model is almost always going to go search first,” Lawson said. “It’s not trying to answer you based on its baseline training … it is using what it understands about the world to then say, ‘Hey, this is what it seems like this person is really interested in understanding, and let’s go out and pull in sources of information that could be helpful to that person.’”

While Meta now offers AI models, this issue is nothing new for the Facebook and Instagram parent company. It reactivated its Elections Operations Center last week, which started in 2018 to quickly respond to threats of misinformation, fake accounts or harmful content leading up to and after an election. A “restriction” period also began last week, during which new political and social issue ads are prohibited given the limited time to verify new claims.

When voters ask Meta AI or Meta’s Muse model about elections, the models are supposed to give local information and direct to “authoritative government sources,” according to the company.

Google, which runs one of the most popular search engines along with AI models, also partnered with Democracy Works and the AP.

The company came under fire last week after its chatbot did not directly say who won the 2020 election, while responding normally to other election years. Google later said there was a “technical gap” with 2020 election queries, which was fixed.

“Our systems verify election responses using news articles or AP data for recent elections and historic records for older ones. 2020 was caught in between. The same was true for 2022 Senate races, which we also fixed,” the company said last Friday.

As AI advances at a rapid pace, the industry faced mounting scrutiny this year for how bad actors could utilize the technology to more easily hack into sensitive systems or infrastructure, create misleading imagery or language and eventually build powerful weapons.

Rather than tampering with a voting machine, cybersecurity experts say bad actors, including foreign influence operators, are more focused on confusing and misleading voters.

A report released earlier this year by cybersecurity firm Check Point Software found efforts to phish, impersonate and spread misinformation will be amplified by AI, giving malicious actors cheaper, broader and faster deception methods.

Taking a page out of Meta’s playbook, OpenAI and Anthropic are hosting their own “election war rooms” for the first time to monitor whether their tools are being used by bad actors.

According to OpenAI, its Election Operations Center is made up of hundreds of OpenAI employees, including the company’s top investigators and specialist teams. They will be managing “rapid response” for reviewing flagged abuse and enforcing policies or suspending and terminating accounts when needed.

Misuse will be the primary focus for OpenAI, though teams will also be managing product functionality and external requests. Politico first reported on OpenAI’s operations center.

“We’ve put considerable effort in ensuring we help users get reliable voting information, combat misuse of our tools, and support election defenders,” Lawson said in a statement shared Monday, adding the center “builds on that years-long work, bringing our teams together to respond quickly to emerging issues and enforce our policies as people head to the polls.”

Anthropic also has a dedicated war room to readily respond to technical and misuse incidents, including the use of Claude, the company confirmed to The Hill. The company defines “misuse” as using Claude to deceive voters or disrupt elections, such as by impersonating election officials or spreading misinformation. Semafor first reported Anthropic’s election war room.

Both AI firms continuously monitor the use of their models to mislead voters or sow chaos or confusion during the electoral process.

In a threat report last month, OpenAI revealed it banned several accounts associated with propaganda operations in Iran and Russia.

Aside from text posts containing misinformation, concerns have been raised for years over the potential for deepfakes or other AI-generated audio or images impersonating officials and misleading voters.

OpenAI launched a tool for users to upload images to check whether they were made by ChatGPT. Google also offers a similar tool.

The ChatGPT maker also introduced an “image stylization” or “cartoonification” feature to avoid users requesting ultrarealistic images for election-related content. With this tool, an OpenAI model will make a cartoon-looking image that voters will easily be able to identify as fake.

Asked what types of prompts ChatGPT may “cartoonify,” Lawson pointed to images depicting long lines at polls, natural disasters and other tactics bad actors have used to mislead voters.