How governments worldwide are shaping what AI can say about politics

With the U.S. midterms approaching, about 1 in 5 voters has asked AI chatbots about elections and politics, a recent survey found.

But many voters may not know that governments around the world increasingly shape what chatbots can say. And because many tech companies apply the same policies worldwide, the answers American voters get may be shaped by governments at home and abroad.

I’m a researcher at a nonpartisan think tank at Vanderbilt University, where I study free speech policy. In two reports published in 2025 and 2026, my colleagues and I worked with global legal experts to examine how governments in 12 countries and regions regulate artificial intelligence. Our research shows that they often use existing rules, such as national security or anti-state propaganda laws, rather than new AI laws to control what AI models can say.

Censorship masked as safety

AI can cause real harm, from launching cyberattacks to generating sexualized images of children.

It’s important that governments tackle these harms. But, in our assessment, some governments use legitimate concerns like these as a pretext to control political discussion, particularly chatbot responses about those in power.

In 2024, an Indian journalist asked Google’s Gemini chatbot whether the policies of India’s prime minister, Narendra Modi, were fascist. Gemini responded that some experts say yes. The government reacted by requiring official permission before “unreliable” AI models could be released. Although authorities later dropped this requirement, the case shows how terms like “under-testing” or “unreliable” can be used for political purposes.

India is not alone. China requires AI models to uphold “socialist core values.”

And in Turkey, a court blocked posts by Grok, the chatbot built into the social platform X, that allegedly insulted President Recep Tayyip Erdoğan, founding father Mustafa Kemal Atatürk and religious values. Research by Turkish law scholar Yaman Akdeniz identifies the Turkish rulings as among the earliest documented cases of state censorship of chatbots.

Vietnam shows what such content controls can mean for dissidents, including those in exile. A Vietnamese court sentenced two activists, who live in Germany, to 17 years each in prison in December 2025 for spreading anti-state material. Prosecutors accused them of using AI to fake voices and images of Vietnamese leaders, illustrating how prosecutions can target political criticism.

Foreign rules, domestic impact

Some of these government measures only affect people residing in their countries. Other statutes cross borders.

The European Union’s AI Act, adopted in 2024, requires any company offering AI tools in Europe, wherever it is based, to mark AI-generated content. Anthropic, for one, applies the marks globally because it has no reliable way to apply them by region. This is an example of the Brussels Effect, in which rules adopted in the EU end up applying elsewhere.

European rules also affect what chatbots say about politics.

In August 2026, the EU placed ChatGPT under the strictest tier of its Digital Services Act, Europe’s online safety rulebook. OpenAI, creator of ChatGPT, now must assess and reduce the “systemic risks” of its service, including risks to elections, such as the spread of disinformation.

And the Dutch Data Protection Authority tested four chatbots before the 2025 national election and found that the political middle ground “virtually disappears” in their voting advice. It urged providers to “prevent their systems from being used for voting advice.”

These rules can be used to pressure platforms.

In 2024, Thierry Breton, then the EU’s top digital rules enforcer, reminded Elon Musk of his legal obligations before Musk’s live interview with then-presidential candidate Donald Trump on X, warning against the spread of “harmful content.”

The United States is no exception

Threats to Americans’ access to information do not come only from abroad.

In 2025, President Trump issued an executive order on “Preventing Woke AI in the Federal Government.” It requires the AI models the government buys to be “truth-seeking,” ideologically neutral and not favoring ideas such as “diversity, equity and inclusion.”

“Truth-seeking” may sound like an objective, neutral term, but it can mean very different things to different people. Take media reporting: The president recently banned three news outlets from the White House for reporting “FICTION and LIES” and dismissed media polls as “fake.”

More recently, the Federal Trade Commission has suggested that AI companies that steer answers toward undisclosed ideological or biased objectives might break consumer protection law.

A 2025 Pew Research Center survey found that 8 in 10 Americans believe Republican and Democratic voters cannot agree on basic facts. In other words, as our report notes, government mandates to seek truth are often driven by ideological belief systems, particularly on political issues.

Meanwhile, dozens of states have passed laws restricting AI-generated political deepfakes. California and Hawaii, for example, are two states that have banned sharing deceptive AI content about candidates in the months around an election. U.S. courts have blocked both of those laws, citing the First Amendment.

What does this mean for voters? When you ask your chatbot about the elections, treat the answer as a starting point. Keep in mind that governments, both foreign and domestic, may try to influence the answer. AI responses should not replace your critical thinking at the ballot box.