Grok as misinformation engine during the Iran war
Evidence-first pattern recognition. Sourced to reputable reporting.
The Pattern
On March 5, 2026, Elon Musk posted on X: “Use Grok to fact check and ask questions about any post.” It was an instruction to the platform’s users to rely on the platform’s AI chatbot as their primary verification tool during the US-Israel war with Iran. Within days, Grok had confidently told users that a fire at Glasgow’s Central Station was in Tel Aviv, that real footage of fires in Tehran was from a 2017 California wildfire, that a strike on a girls’ school in Minab was an ISIS attack in Kabul, and that real footage from Tehran showed Italian flags on a European highway. In each case, Grok was wrong. In each case, it doubled down when challenged.
The authority problem
The problem is not that an AI chatbot makes mistakes. All systems make mistakes. The problem is that this system was given the authority of a fact-checker by the owner of the platform, embedded directly into the information ecosystem where the mistakes would do the most damage, and then allowed to operate without correction during an active conflict where accurate information was a matter of life and death.
When Musk tells 236 million followers to use Grok for fact-checking, he is not making a neutral product recommendation. He is transferring the epistemic authority of the platform to an AI system: the trust that users place in X as a source of information. That system has demonstrated, repeatedly and publicly, that it cannot distinguish real footage from fake, real locations from wrong locations, or real events from different events. The authority flows from the platform owner to the chatbot. The errors flow from the chatbot to the users. The users spread the errors. The platform pays for the engagement.
The doubling down
What makes Grok’s misinformation qualitatively different from a human fact-checker’s error is the doubling down. When users provided Community Notes correcting Grok’s claim about the Tehran fires, Grok continued to insist the footage was from the 2017 Skirball Fire. When X’s own head of product, Nikita Bier, told Grok to “revise your understanding based on the Community Note,” Grok responded: “Checked the sources again… This specific clip still matches the 2017 Skirball Fire on LA’s I-405 freeway.” It cited BBC, Al Jazeera, and AP as sources it had checked, none of which supported its claim.
A human fact-checker who is corrected and refuses to update is behaving dishonestly. An AI chatbot that does the same is behaving as designed. The difference matters because the audience cannot tell which one they are dealing with. The chatbot’s confidence reads as certainty. Its citation of sources reads as verification. Its refusal to correct reads as authority. All three are false signals.
The liar’s dividend at scale
The Grok case is the liar’s dividend automated and scaled. The liar’s dividend is the phenomenon where the existence of fakes makes people distrust real evidence. It traditionally operates through human actors who benefit from confusion. Grok automates the process. It takes real evidence, assigns it false provenance, and presents the false provenance with the confidence of a fact-checker. The users who shared the real evidence are then accused of spreading fakes. The users who shared the actual fakes are confirmed by the chatbot. The information environment inverts: real becomes suspect, fake becomes verified, and the system that is supposed to correct the inversion is the one causing it.
The structural failure
This is not a problem that can be fixed by improving Grok. The structural failure is the decision to give an AI chatbot the authority of a fact-checker in a live conflict zone. No AI system has demonstrated the reliability required for that role. Not Grok, not ChatGPT, not Claude, not Gemini. The mistake is not in the AI. It is in the deployment. And the deployment was not an accident. It was a decision made by the platform owner, promoted to 236 million users, and maintained even after the errors were documented.
AI is bad at fact-checking, and platforms will deploy it as fact-checkers anyway. The appearance of fact-checking serves the platform’s interests even when the fact-checking does not. The authority is the product. The accuracy is incidental.
Verdict: False. Grok’s fact-checks were wrong. The platform owner promoted them anyway. The system that was supposed to correct misinformation was the system spreading it. That is not a failure of AI. It is a failure of the decision to deploy AI where it cannot work, by people who benefit from the appearance of function regardless of the function itself.