Examining AI ‘hallucinations’ and Sunday’s Gaza blasts Thomas Copeland, BBC Verify Live journalist As we prepare to close this live coverage, here's a summary of today's key stories. The AI chatbot Grok inaccurately claimed on X that aerial footage of an anti-Trump protest in Boston from Saturday was from 2017. We analyzed why Grok made this error, how the misinformation spread online, and detailed how reverse image searches helped debunk the claim. An expert explained why AI chatbots sometimes “hallucinate” or produce false information. Meanwhile, Israel’s military reported conducting strikes against Hamas in Gaza on Sunday, with accusations of ceasefire breaches from both sides. Our team reviewed videos circulating of a man being beaten and shot by masked men in Gaza. Despite recent postings, reverse image searches revealed the footage is over a year old. Additionally, we verified footage of a massive explosion and fire at a chemical plant in Jiangyin, eastern China. The video shows intense flames and smoke; local authorities confirmed the incident involved phosphorus overflow but was contained within about 50 minutes, with no environmental contamination detected. BBC Verify Live remains logged in to verify imagery and monitor ongoing stories, returning tomorrow. Video of Chinese chemical plant blast verified Yi Ma and Kumar Malhotra, BBC Verify We authenticated video showing a large fire caused by an explosion at Chengxing chemical plant in Jiangyin, Jiangsu province. The structure and surrounding greenery in the footage were matched with Google Earth satellite imagery. Reverse image searches confirmed the video’s recent origin. Local officials reported the fire stemmed from phosphorus overflow and was quickly controlled without environmental harm. Join BBC Verify Rob Corp, BBC Verify Live editor BBC Verify investigates the truth behind claims, political statements, social media videos, and war zone images. We welcome suggestions—if you see something suspicious online or wonder if content is AI-generated or a deepfake, contact us here to request investigation. Context: ‘No Kings’ protests in the US Thomas Copeland, BBC Verify Live journalist We covered how AI chatbot Grok wrongly identified aerial footage of Boston’s “No Kings” anti-Trump protest as from 2017. On Saturday, millions gathered across US cities including Boston, New York, Washington DC, and Miami to protest Trump’s policies. The No Kings coalition, comprising around 300 groups, first gained attention with large demonstrations on Trump’s birthday in June. Conservative politicians criticized the events while organizers reported about seven million peaceful participants nationwide. What is reverse image search? Thomas Copeland, BBC Verify Live journalist Reverse image search is a key verification method where a photo or video frame is uploaded to a search engine to trace prior instances online, helping confirm when and where content first appeared. We use tools like Google Lens, Bing, Yandex, and others, each excelling in different regions or contexts. For videos, we examine multiple frames. Though effective, reverse image searches are not foolproof—for instance, genuinely new footage cannot be found this way. Thus, we conduct multiple checks to ensure authenticity. Old Gaza video showing man beaten and shot Kumar Malhotra, BBC Verify senior journalist Last week, we verified videos of Hamas public executions in Gaza City. On Sunday, another graphic video surfaced showing a shirtless man with bound hands being dragged, beaten, and shot by masked men. The clip, which amassed over 700, 000 views, was claimed as recent and linked to Hamas. However, reverse image search dated it to October 3, 2024.

We cannot confirm identities, though BBC Arabic colleagues suggest the men speak with Gaza accents. Old videos often resurface amid conflicts, so verifying the timing is crucial. Why AI chatbots sometimes fabricate information Thomas Copeland, BBC Verify Live journalist The AI chatbot Grok falsely dated Boston protest footage to 2017. xAI, Grok’s developer, was asked about these claims. Such errors are not new; a similar issue arose after protests in London. Dr. Peter Bentley, a computer scientist from UCL, explains that AIs aim to please by giving plausible answers using limited data—mostly training data and internet content. With novel information like fresh protest footage, data scarcity or inaccuracies cause AI to generate wrong, confident answers known as “hallucinations. ” Fire at Ukraine coal mine after reported Russian strike Fridon Kiria and Thomas Copeland, BBC Monitoring and BBC Verify We verified footage on Telegram showing a fire at a coal mine in Ternivka, Dnipropetrovsk region, Ukraine, following a reported Russian drone attack. Satellite images matched features seen in the video. Reverse image searches confirmed the footage was new. The Ukrainian energy firm DTEK confirmed the strike, noting 192 miners were underground but all were safely evacuated without injuries. This marked the fourth major enemy attack on DTEK mines in two months. Verified footage of Israeli strikes and Gaza aftermath Benedict Garman, BBC Verify senior journalist Following a temporary ceasefire break, Israel conducted strikes in Gaza in retaliation for an alleged Hamas attack. We verified graphic videos showing damage and casualties at multiple locations, including a beachside cafe in Az-Zawayda, a sports field in Nuseirat, and areas east of Khan Younis and behind al-Shifa hospital in Gaza City. Footage shows injured civilians, including children, and smoke plumes. Verification continues on additional videos from Bureij, Nuseirat, Az-Zawayda, and Sheikh Radwan. AI chatbot Grok’s false claim about ‘No Kings’ protest footage Thomas Copeland, BBC Verify Live journalist Grok wrongly asserted on X that aerial footage of Boston’s “No Kings” protest was from 2017. Reverse image searches of multiple video frames confirmed the footage only appeared online after the protest date. Local media showed similar footage confirming event details. Grok’s error stemmed from an unapproved “Community Note” citing unrelated 2017 protest links. Screenshots of Grok’s claim were widely circulated by pro-Trump influencers to question protest legitimacy. After a BBC Verify post on X, Grok corrected its statement and even cited BBC Verify in its update.