A Fundación Maldita.es audit published Oct. 27, 2026 found that Grok correctly judged 121 of 245 responses about images or videos (49.4%) in a sample of World Cup-related posts on X. For written claims, it got 69 of 72 responses right (95.8%). Those figures describe the audit’s sample, not a general score for Grok.
What the two accuracy rates measure
The audit assessed 317 responses for truthfulness. The visual group covered responses about images or videos; the written-claim group included text alone, text paired with an image, and text reproduced in a screenshot.
| Response group | Correct responses | Assessed responses | Correct |
| Images or videos | 121 | 245 | 49.4% |
| Written claims | 69 | 72 | 95.8% |
What the audit sampled
Fundación Maldita.es examined 139 X posts grouped into 46 claims. The claims included 35 that were false or manipulated, six that needed context, two alerts, and three true claims. Researchers and other users submitted 766 questions; Grok returned 358 responses, of which 317 could be assessed for truthfulness.
The sample leaned heavily toward false or manipulated material, so its results describe that World Cup-related set of claims. Among 127 responses judged incorrect or partly correct, 81 failed to detect visual material that had been manipulated or generated with AI.
What happened after fact-checks appeared
Among 22 posts where Grok’s initial answer was incorrect or partly correct, it later corrected itself in three cases. It gave no further answer in 14 and remained wrong or incomplete in five.
NeoTeo’s earlier coverage of Grok 4.6’s benchmark results looks at a separate model-performance question.
