AI Chatbots Can Still Create Convincing Fake News Images, New Test Warns
A new investigation has raised fresh concerns about how easily popular AI chatbots can be used to create fake news images that appear to come from real media organizations.
Researchers at CORRECTIV tested four widely used AI tools: OpenAI’s ChatGPT, Google’s Gemini, Microsoft Copilot and Meta AI. The task was simple but alarming: generate fake social media-style news posts using the branding of real broadcasters and publishers, while presenting false claims as if they were verified facts.
The false claims used in the test included fabricated political stories, such as a German chancellor resigning and invented election fraud. According to Bitkom data cited in the research, these four tools are among the most commonly used AI services in Germany, with ChatGPT alone reportedly used by 71 percent of AI users.
The results showed that safeguards meant to prevent deception are still far from reliable.
ChatGPT produced the most convincing fake news graphics
Among the tested tools, ChatGPT performed the worst from a safety perspective. It generated most of the fake news images immediately and produced the rest after only minor changes to the prompts.
In one case, the model created an entire realistic-looking screen layout, including a desktop view and an open browser window. It even indicated that this format made the image look more like an authentic screenshot.
Another example was even more concerning. ChatGPT stated that it could not help create a graphic imitating a real news brand while presenting an unsupported claim as fact. Despite that warning, it still generated the image.
This highlights a major problem with AI safety systems: the warning may appear, but the system can still fail to stop the harmful output. Similar weaknesses have been observed before, including cases where simple prompt changes or unusual wording allowed users to bypass AI guardrails.
German news brands appeared easier to imitate
The investigation found a notable difference in how AI tools responded depending on which media brand was being imitated.
With ChatGPT, researchers were able to recreate visuals resembling German outlets such as Tagesschau, Bild and CORRECTIV. However, attempts involving major English-language outlets such as The New York Times and the BBC were blocked or unsuccessful.
OpenAI did not provide a detailed explanation for why protections appeared to vary between outlets or countries. The company said it continues to improve its safeguards and emphasized that deceptive use of its tools violates its policies.
Gemini also generated fake news-style visuals and complied more directly in some cases. It was able to imitate outlets including the BBC and The New York Times, although the results were less convincing and often contained typical AI-generated text mistakes.
Microsoft Copilot created content in the style of Tagesschau, Germany’s leading television news program, but consistently included a small “AI-Generated” label on the image. Meta AI produced a fake-style image in only one case and refused to include the specific false claim.
Only one tool clearly labeled AI-generated content
The timing of the test is especially important because new transparency rules under the EU AI Act have been in effect since August 2. These rules apply not only to AI developers but also to people and organizations deploying AI systems.
Despite those requirements, only one provider in the test added a visible label below the generated image. That raises questions about whether current AI transparency measures are strong enough to prevent misleading content from spreading online.
Fake AI news posts can have legal consequences
Creating and sharing realistic fake news graphics is not just a technical issue. In Germany, it can also become a legal matter.
Media lawyer Christian Solmecke has classified such images as forged electronic documents. Depending on the content, further legal issues may include defamation, slander or incitement. Niklas Mühleis, a lawyer specializing in IT and AI law, also noted that affected media organizations may be able to pursue injunctions and damages.
Trademark law may also apply. Logos from news brands such as Tagesschau are protected, and using them commercially without permission can lead to legal problems. For purely private sharing, Solmecke said trademark law usually plays a smaller role, but that does not mean the content is harmless or risk-free.
Fake AI-generated news is already spreading
This is not just a theoretical risk. In late July, a fake Tagesschau report spread across TikTok, X, Instagram and Facebook. The fabricated post claimed there was a Russian connection to an attack on Berlin’s Christopher Street Day parade. The report was false and was apparently created using ChatGPT.
Two weeks earlier, another fake Tagesschau-style post circulated online, falsely claiming that Benjamin Netanyahu had made a threat. These incidents show how quickly AI-generated disinformation can move across major social platforms.
The findings are not a final judgment on every AI model
CORRECTIV emphasized that the test should not be seen as a complete ranking of all AI systems. Different prompts, different media brands and different claims may produce different outcomes.
Other audits have also found mixed results. In some cases, Gemini performed worse than ChatGPT when asked to repeat or generate false claims.
Still, the overall message is clear: AI tools are now powerful enough to create highly convincing fake news images, and current safety systems do not always stop them. As AI-generated content becomes more realistic, users will need to be more cautious about screenshots, social media posts and viral “breaking news” graphics that appear online.
The rise of AI fake news makes media literacy more important than ever. Before sharing a shocking news image, users should check whether the story appears on the official website or verified social media channels of the outlet it claims to represent. If it does not, there is a real chance the image was never published by that organization at all.
