Fed on falsehoods, AI chatbots generate more false information

The world’s top AI chatbots, such as ChatGPT, are providing answers containing false information twice as much as they did last year, a new study from disinformation-fighting watchdog NewsGuard has found.

The disinformation rates of the top ten leading chatbots have doubled, going from 18% in August 2024 to 35% a year later.

The tools now regularly reproduce false claims on topics such as health, politics, international affairs, companies, and business brands. The chatbots now provide false claims to news prompts more than one-third of the time.

“When it comes to providing reliable information about current affairs, the industry’s promises of safer, more reliable systems have not translated into real-world progress,” said NewsGuard.

The natural question is, why? After all, with the AI revolution gaining speed, the chatbots should be improving and becoming more reliable.

Malign actors polluting data

The problem, according to NewsGuard, is a “structural tradeoff.” The chatbots have by now adopted real-time web searchers and moved away from declining to answer questions: their non-response rates fell from 31% in August 2025 to exactly 0% a year later.

Instead of citing data cutoffs or refusing to weigh in on sensitive topics, the large language models now pull from a polluted online information ecosystem and treat unreliable sources as credible, NewsGuard explained.

Surprising no one, this ecosystem is sporadically deliberately seeded by networks of malign actors, including Russian disinformation operations.

“Malign actors are exploiting this new eagerness to answer news queries to launder falsehoods via low-engagement websites, social media posts, and AI-generated content farms that the models fail to distinguish from credible outlets,” said the watchdog.

“In short, the push to make chatbots more responsive and timely has inadvertently made them more likely to spread propaganda.”

At least three different organizations – the American Sunlight Project, NewsGuard, and Open Measures – warned this year that Russia is seeding AI models with Moscow-friendly falsehoods in support of its international and political ambitions.

The chatbots with the lowest fail rates were Claude (10%) and Google’s Gemini (16.67%).

Russian groups are essentially publishing millions of fake news content pieces and hoping that at least some of them would fall through the security filters of the large language models and be included in results.

Too much trust in LLMs

China’s probably also at it. In June, a report by the American Security Project surprisingly found that many AI models – not just Chinese ones – parrot Chinese Communist Party messaging, meaning that they might have been trained on party-state disinformation and conspiracy blogs.

The chatbots that most often produced false claims in their responses on topics in the news were Inflection (56.67%) and Perplexity (46.67%). ChatGPT and Meta spread falsehoods 40% of the time.

Meanwhile, the chatbots with the lowest fail rates were Claude (10%) and Google’s Gemini (16.67%).

Interestingly, even OpenAI CEO Sam Altman has already acknowledged the problem. In June, he said: “People have a very high degree of trust in ChatGPT, which is interesting, because AI hallucinates. It should be the tech that you don’t trust that much.”

Unfortunately, we tend to trust AI with almost everything, it seems. A recent ChatGPT leak revealed that users often share private, sensitive, and even risky information with the chatbot – one should assume they also believe whatever the bot is telling them.

Unlock more exclusive Cybernews content on YouTube