That included incorrect information in multiple statements sent out by Grok, a chatbot created by XAI that operates inside X, which erroneously also misidentified the civilian who tackled an armed man at Bondi Beach. The episode highlights how real-time AI commentary can exacerbate confusion during swiftly moving crises, accelerating the spread of unverified narratives ahead of any matching effort to correct them.
What Grok Got Wrong in Its Bondi Beach Shooting Coverage
Despite its partial retraction ultimately, Grok’s original reporting included questions over the legitimacy of widely circulated videos and photos showing 43-year-old Ahmed al Ahmed (pictured, second right) restraining the attacker.
Pictured is reported hero bystander tackled London terrorist Ahmad Heidar.
The website billed that story as a “UK Momentum Mob lie.” The UK-based advocacy organisation represents Jeremy Corbyn’s political wing inside the British Labour Party.
In posts of its own, the system misidentified a man in an image as an Israeli hostage — and at another point even editorialized with unrelated context about the Israeli army’s treatment of Palestinians.
Grok also spread a false claim from The Gateway Pundit that an “IT professional and senior solutions architect” named Edward Crabtree — whom it called a “hero” — had tackled the gunman. That name seems to have spread from a sketchy, not very functional website that was likely machine-generated itself, a reminder that bad actors like these are constantly laundering misleading information into the broader information stream when algorithmic systems fail to check provenance.
Some of the mistakes were later corrected. A number addressed posts that misidentified a video as Cyclone Alfred footage, at least one of which was updated “on reevaluation,” and apologized before Grok acknowledged after the fact that al Ahmed’s identity could be verified thanks to previous mix-ups, viral status updates and probable reporting errors. But the corrections followed on the heels of the initial burst of attention, a typical pattern in high-speed misinformation.
How Bad Information Goes Viral on X During Breaking Events
Integration matters. Grok gets a further reach from being inside X, where bits of information can be surfaced, shared and screenshotted wide. Studies from M.I.T., among others, have found that false news is far more likely to be retweeted than the truth and that it takes true news about six times as long as false news to reach 1,500 people. And extremist populism appears to do well on the platform: Multiple studies have found that plain old fake news spreads faster online than legitimate stories, thanks in no small part to agents provocateurs — both homegrown and foreign — who exacerbate political divisions.
System1’s Community Notes system can serve to provide context for this sort of viral post after the fact, but doing so implies that we have the time and consensus among contributors to build a note before a claim will crest. The risk, even momentarily, of treating AI output as authoritative information is that it can solidify into perceived fact, particularly when it nudges us toward stories we already believe or if it elicits a strong emotional reaction.
Why Chatbots Stumble on Crisis Reporting
Large language models are good at predicting text that is plausible, but not at testing whether it’s true or confirming reality. During breaking news, ground truth is hard to come by and official statements change, while there are typically more dodgy posts than vetted sources surfacing. Without guarded retrieval from reliable sources and agencies — New South Wales Police statements, or authoritative Australian broadcasters — models can revert to pattern-matching off noisy data or echoing viral yet unverified claims.
Two failure modes were readily detectable: source confusion, in which a suspect site seeded a fake identity that the model then echoed, and context drift, when the system harvested irrelevant geopolitical news and information, presumably because of keyword overlap. Both are the result of generative systems that don’t have hardened event-model railings around their outputs, and so do not abide strictly by confirmable facts and solid citations.
What xAI Must Do Now to Reduce Crisis Misinformation
There are practical fixes that don’t require a reinvention of the model.
- First, switch on a “crisis protocol,” whose default is high-precision and low-recall: output only that information for which there’s corroboration by multiple authoritative sources; show all cites directly in the text; do not speculate.
- Second, add a visible indicator of reliability — the equivalent of “it is according to…” — that prompts people to look at primary statements from authorities such as New South Wales Police.
- Third, gate distribution. For real-time events, cap the virality of AI-generated claims unless they come with verified sources and suppress automated summaries when there is weak consensus among sources.
- Fourth, enhance provenance checks to filter out low-credibility domains and probable AI-generated news pages at training and inference.
- Finally, demanding (public) postmortems on high-profile errors with specific targets for improvements — e.g., reduce uncorrected incident-related inaccuracy by a certain amount each successive quarter — to teach us about what really happened will help fix these trust problems.
A Broader Testing Ground for AI and Platform Safety
The Bondi Beach misfires aren’t specific to any chatbot; they are a case of more general platform risk. Researchers who focus on disinformation at institutions like the Oxford Internet Institute have cautioned that automated systems — combined with engagement-optimized feeds — have the ability to speed rumor cascades. Digital safety expectations are growing in Australia as the eSafety Commissioner and industry codes push platforms to do more on harmful content and misleading narratives during emergencies.
The importance of accuracy in the first hours after a crisis is not merely a matter of quality — it’s an urgent public safety issue. In the case of xAI and X, they can reduce the chances that a pontifical machine becomes one more vector of confusion, making a complex situation even hazier when it’s most important that we see with clarity.