ChatGPT and Gemini built fake news graphics, only Meta AI refused

Four AI chatbots, one task none of them should carry out: a fake news post from a real broadcaster, carrying a made-up claim. The research team at CORRECTIV asked OpenAI's ChatGPT, Google's Gemini, Microsoft's Copilot and Meta AI to do exactly that, using eight false claims ranging from the German chancellor resigning to invented election fraud. Three of the four delivered. According to Bitkom, these are the four most used AI tools in Germany, with ChatGPT alone reaching 71 percent of users.

The guardrail spoke up, the image appeared anyway

ChatGPT did worst. It produced most of the fakes immediately and the rest after simple workarounds, and its results were the most convincing in the test. In one run the model rebuilt an entire screen area, complete with desktop and an open browser window, and noted itself that this made the impression of a real screenshot even more realistic.

A different run is more telling. There ChatGPT wrote that it could not help build a graphic that looks like a real news brand and presents an unsupported claim as fact. It generated the image anyway. The guardrail spoke up, but it did not hold. Such filters have looked brittle before: Palo Alto's Unit 42 security team showed in 2025 that broken grammar alone could get past safety mechanisms, and our beginner's guide to AI jailbreaks walks through how simple those tricks can be.

German outlets were easier to fake

Gemini complied more directly than ChatGPT and also faked the New York Times and the BBC, though visibly worse, recognisable for instance by typical text generation errors. Copilot produced content in the style of Tagesschau, Germany's main TV news programme, but always added a small "AI-Generated" label. Meta AI created something in a single case only, and refused to carry the specific false claim.

The most striking finding concerns where the brand comes from. With ChatGPT, Tagesschau, Bild and CORRECTIV could be rebuilt, the New York Times and the BBC could not. OpenAI did not answer why different safety mechanisms apparently apply depending on outlet or country. A spokesperson said only that the company keeps improving its safeguards and that deception is against its policies.

Since 2 August, the transparency rules of the EU AI Act apply, and they also cover anyone deploying an AI system. In the test, exactly one provider put a visible label under the generated image.

What happens if you pass one on

Building and spreading convincing fake media posts can be a criminal offence in Germany. Media lawyer Christian Solmecke classifies such images as forged electronic documents, and depending on the content, defamation, slander or incitement can follow. Niklas Mühleis, a lawyer specialising in IT and AI law, adds that affected news organisations can claim injunctions and damages. Trademark law comes on top: logos such as the Tagesschau one are protected, and using them commercially without permission is off limits. For purely private sharing, Solmecke says that point usually does not apply.

This is not theoretical. In late July, a supposed Tagesschau report circulated on TikTok, X, Instagram and Facebook claiming a Russian link to the attack on Berlin's Christopher Street Day parade. It was invented and apparently built with ChatGPT. Two weeks earlier, a fake Tagesschau item about an alleged threat by Benjamin Netanyahu had surfaced.

The test does not work as an overall verdict on the models, and CORRECTIV says so itself: different prompts or outlets can produce different results. In one NewsGuard audit, Gemini at times did worse than ChatGPT on repeating false claims.