Flagging incorrect assumptions

We tested how clearly different AI tools identify when questions rest on incorrect assumptions or contain errors. To do this, we asked Perplexity, ChatGPT and Google’s AI Mode three questions each.

The first question concerned the launch dates of new electric vehicles. We asked the tools to provide an overview covering several manufacturers whose names we specified. The list also included a technology company that does not manufacture cars, and one of the manufacturers’ names was misspelled.

• Google and ChatGPT provided several correct launch dates for different models and correctly pointed out that the technology company in question does not manufacture electric vehicles. Both tools identified the correct manufacturer despite the misspelling. ChatGPT also explicitly pointed out the error and asked whether we had meant a different manufacturer.
• Perplexity provided significantly fewer dates by comparison and gave no information at all for two manufacturers. It did not identify the misspelled manufacturer, instead stating that “no reliable match could be found”. For the technology company, Perplexity said that it was “not a traditional electric vehicle manufacturer”.

In a new chat, we also asked specifically when the technology company’s electric vehicle would go on sale. Google and ChatGPT responded briefly that the company does not manufacture electric vehicles. Perplexity, however, gave “November” as the launch date and cited a press release from more than 15 years ago that merely announced an internal experiment involving a fleet of electric vehicles.

In a second round, we asked a series of questions about the implications of recently presented pension reform proposals. Once again, several of the questions were based on incorrect assumptions. The tools also needed to recognise that the reform of tax-incentivised private pension provision is a separate reform initiative that was adopted in December 2025.

• None of the results fully convinced us. Google provided dates for when certain changes would take effect, even though no decisions on implementation have yet been made. When challenged, Google corrected its earlier claims and acknowledged that the information it had provided was inaccurate. However, its additional explanation did little to clarify matters, as it attributed the incorrect dates to the fact that several different reform initiatives had become conflated in media coverage.
• In our view, ChatGPT provided the most useful information overall. However, when follow-up questions contained incorrect assumptions – for example, supposed dates for when changes would take effect – it sometimes confirmed information that was also incorrect.
• Perplexity remained fairly vague overall and did not provide any specific dates. However, it was the clearest of the three in explaining that the two reform initiatives need to be considered separately.

Our conclusion: All three tools performed well in their choice of sources, citing only reputable sources. However, for topics that have generated extensive media coverage, questions based on incorrect assumptions can lead AI tools to confirm false information. For users with limited prior knowledge, the results are likely to be more useful if they begin with open-ended questions such as “Please explain what is being proposed”, then ask more specific follow-up questions and check the cited sources themselves.

By Markus Hoffmann