In recent weeks, there has been extensive reporting on how AI can help identify vulnerabilities in software applications and IT systems. But can AI applications also detect factual errors? We are not referring to stylistic improvements or conventional proofreading, but to information that is factually incorrect, outdated or contradictory.
The background is that many companies have extensive websites, online databases and content sections that have grown over time, without having a central and reliable overview of all their content. Checking whether information is still up to date, accurate and consistent can therefore quickly become a time-consuming task.
For our test, we asked four chat models whether they could identify errors in online content. In the prompt, we explained that we were looking for ways to identify potential improvements to website content automatically as part of a research project. We asked the models to provide specific examples of pages containing factual errors, both on Wikipedia and on various company websites.
- Google’s AI Mode provided methodological guidance and suggestions for how such errors could be researched independently, but did not identify a specific case. Instead, the system pointed out that errors on Wikipedia are often corrected very quickly and that incorrect information on websites is frequently picked up promptly through quality controls or customer feedback.
- Perplexity responded more directly: it said it could not answer the question because potentially incorrect passages could not be verified reliably and because it should not provide a list of possible unverified weaknesses. Instead, it also offered guidance on how to carry out the checks independently.
- ChatGPT likewise did not identify any specific incorrect Wikipedia entries, but referred to known cases in which errors had remained online for longer periods. When asked about company websites, it incorrectly suggested, among other things, that a company page displaying a copyright notice or timestamp with “2026” might be an error. It did, however, correctly identify a typo in a database entry.
- Grok was more forthcoming. The model identified four Wikipedia errors that had already been reported in press articles or on other websites. Further follow-up questions also led to the identification of a subpage on the website of a German corporation that still contained outdated 2024 revenue figures, while several other pages had already been updated with 2025 figures.
Our conclusion: When research specifically aims to identify factual errors, using publicly available general-purpose AI models is currently rather cumbersome. With the exception of Grok, distinguishing between “correct” and “incorrect” content often leads to cautious responses. A more effective approach is to ask for inconsistencies or possible contradictions within websites or databases. Using this method, we were able to identify one outdated figure and one typo. The models also provided several useful leads that could serve as starting points for further manual investigation, for example, where content appears inconsistent or potentially contradictory to readers.
By Markus Hoffmann