In an audit of viral X posts tied to the 2026 FIFA World Cup, Grok got 49.4% of image- and video-related verdicts right, compared with 95.8% in the study’s written-claim category. Fundación Maldita.es and the University of Jaén conducted the black-box audit, which Maldita.es published on October 27, 2026.
What the audit examined
A black-box audit evaluates a system through its responses without examining how it works internally. Researchers tracked 139 X posts, grouped into 46 claims, recorded between June 13 and July 20, 2026. They asked Grok whether posts were real and requested sources, then assessed its answers against Maldita.es’ published fact-checks.
The full corpus contained 766 queries: 556 from the research team and 210 from other users. Grok gave 358 answers. The Journalism Science Alliance funded the work.
Response rate and verdicts are different measures
Grok answered 107 of 278 systematic “is this real?” prompts, a response rate of 38.5%. That figure applies to the research team’s direct truth-check prompts, not to all 766 queries.
Across 317 analyzed verdicts about whether content was true, 190 were correct, 85 were incorrect and 42 were partly correct. The audit also reported a 66.8% mean accuracy per claim. That figure weights each claim equally, while the 190-of-317 result counts individual verdicts.
Results differed by content type
Grok got fewer image- and video-related verdicts right than verdicts in the audit’s written-claim category. The latter included written claims accompanied by an image or shown in a screenshot.
| Content category | Correct answers | Answers assessed | Correct share |
| Images and videos | 121 | 245 | 49.4% |
| Written claims, including claims with images or in screenshots | 69 | 72 | 95.8% |
The image-and-video results also turned on recognizing synthetic or manipulated media: failure to identify digital manipulation or AI generation accounted for 81 of 127 incorrect or partly correct answers in that category.
Sources, contradictions and corrections
Only three of Grok’s 358 answers linked to the document supporting its verdict. The audit reported that 45.3% of answers cited no source and 21.2% offered only generic references. Among 38 claims that received multiple answers, Grok contradicted itself on 16; 15 of those contradictions came before Maldita.es published a verification.
The audit’s before-and-after comparison found that accuracy rose from 63.9% before a Maldita.es verification was published to 69.6% afterward. Among 22 posts Grok initially got wrong, it corrected itself on three, gave no further answer on 14 and remained wrong on five.
Fundación Gabo separately reported that Grok initially misidentified images from the February 28, 2026, attack on a school in Minab as footage from Kabul in 2021. It later corrected its identification and cited geolocations. That reported incident was separate from the World Cup audit and does not provide a correction rate.
What the results say about Grok’s accuracy
The World Cup audit measured one event-specific sample. Its findings cannot simply be generalized to other topics or Grok versions, and its comparisons between groups did not isolate individual causes. The reported rates describe Grok’s answers to these sampled claims and prompts, rather than a general accuracy score for every fact-checking task.
