A study by Maldita.es found that Grok gave an incorrect or partly incorrect verdict in about four out of ten analysed answers about viral World Cup claims. The audit also found that the chatbot often contradicted itself and rarely named the sources behind its answers, a troublesome combination when people use it to check what they see online.
xAI Watch analysis
What happened
Maldita.es and the University of Jaén audited Grok’s responses to 46 claims that circulated on X during the 2026 FIFA World Cup. The researchers collected 766 queries and analysed 358 answers, comparing them with Maldita.es fact-checks. The report is based on observed responses, not access to Grok’s internal workings.
Of the analysed answers about whether claims were true, 59.9 per cent were correct, 26.8 per cent incorrect and 13.2 per cent partly correct. The average accuracy per claim was 66.8 per cent. Grok did better with written claims than with images and video: it was right in 49.4 per cent of answers about visual material, where manipulated or AI-generated content was a recurring source of error.
The audit also found inconsistency. For 16 of 38 claims that received multiple answers, Grok gave contradictory verdicts, including cases where its answer changed before a fact-check had been published. Only three of the 358 analysed responses linked to a document backing the verdict. The researchers say the findings describe this World Cup sample and should not be treated as a measure of every topic or version of the model.
Why it matters
A confident answer can look like a fact-check even when it is wrong, unstable or impossible to trace to a source. That matters most when users are trying to assess a fresh image or video, precisely where this audit found weaker results. Asking a chatbot “is this real?” may produce a verdict; it does not guarantee evidence.
Our read
The revealing number is not just how often Grok got the answer wrong. It is how often a reader could not see what the answer rested on. Treat chatbot verdicts as a lead to check, not a stamp of authenticity. Annoying, yes. Still cheaper than letting a fluent answer do the thinking for you.
What to watch
- Whether later audits find similar accuracy and citation patterns on other topics.
- Whether Grok becomes more willing to abstain when evidence is unclear.
- The University of Jaén’s planned research into how these answers affect belief and sharing.
Discussion spark: When a chatbot gives a verdict on a viral image, should it have to show an identifiable source or abstain, even if that means answering fewer questions?
Sources and evidence
not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor