Chatbot as fact-checker: Grok, Grok. Is this true? Sure sir, let me please you!

India Today

16-05-2025

'Is this true @grok?'Surely you must have typed these words on X — or at least, seen others doing it during the recent India-Pakistan conflict. But can we rely on it to fact-check videos related to a military conflict?Recently, the xAI chatbot made headlines for inserting 'white genocide', a far-right conspiracy theory, into unrelated queries. The company now claims to have fixed it. But was this an isolated instance of Grok messing up?advertisement
As India and Pakistan's armies locked horns, unverified videos of explosions, firing, and fighter jets dogfighting flooded the platform. Some people decided to be cautious and asked Grok for verification. Grok was quick and useful — but not always correct.Grok is a handy tool and not a terrible place to start when looking for 'fact-checks'. But keep in mind: it can confidently mislead you. In fact, according to a March 2025 Columbia Journalism Review study, Grok 3, when asked to cite the source of any given information, failed 94 per cent of the time.We found that in the case of queries related to the India-Pakistan conflict, its 'fact-checks' were hit-and-miss. Here are some examples where it gave the wrong answers. Interestingly, with the confidence of a diligent fact-checker.Attack on Israel flagged as Amritsar advertisementA video of bright lights flashing in the night sky was widely shared as a Pakistan Air Force attack on Punjab's Amritsar being thwarted by India's S-400 missile defence system. So, when people asked Grok if this was correct, the chatbot confirmed it was from Punjab's Amritsar and possibly involved the S-400 system.However, in reality, this August 2024 video shows the Lebanese political party and paramilitary group, Hezbollah, firing rockets at Northern Israel. It was even shared on the official account of Israel on August 4, 2024.One problem with Grok appears to be recency bias, or giving more weight to recent events than older ones. It even admitted that it can overemphasise recent trends or viral ideas from platforms like X, and 'If something's been hyped up lately, I might give it more weight than it deserves long-term.'Nepal is not Pakistan The Columbia study found that these AI chatbots rarely admit when they don't know something. Instead, they blabber authoritative-sounding, incorrect answers.For example, a video related to violent pro-monarchy protests in Nepal's Kathmandu in March was shared as Pakistan's attack on India, as part of its Operation Bunyanul Marsoos. When asked to verify this video, Grok responded that it did show Pakistan's retaliatory operation on Indian military sites.advertisementA bang from Bangladesh A video of a massive explosion was presented by Pakistan-based X accounts as evidence of the failure of India's air defence missile systems. When some people asked Grok to fact-check this video, the bot said that the video likely showed Indian missile strikes on Pakistan in either Bahawalpur or Muridke!This video is neither Indian nor Pakistani. It dates back to February this year, and shows a massive fire in Khilgaon, Bangladesh.According to Alex Mahadevan, the director of MediaWise, Poynter's digital media literacy project, 'The sources it pulls from are opaque, but it seems most of its responses come directly from X, which is already flooded with misinformation. Because Grok is trained on and responds using this kind of content, it tends to do worse than models like ChatGPT or Claude.'Grok is a yes-man!Grok agrees with whatever you say too easily. Consider this July 2024 video, which shows a Russian anti-aircraft gun getting destroyed by its own shell exploding. Many claimed this video was from the India-Pakistan conflict and showed the Indian Army's blunder.Grok first affirmed that these were Indian soldiers operating a tank, adding it was either from a training exercise or related to India's recent border tensions with Pakistan. An X user then responded to Grok, sharing the screenshot of a July 2024 news report stating it was Russian soldiers in the video. Immediately, Grok changed its stance and apologised.advertisement'Confirmation bias is a core weakness of all large language models — they're sycophantic by design,' explained Mahadevan. 'Their job is to predict what the user wants to hear, and they'll adjust their answers based on how a question is asked. So, if you're looking at a video, and you're pro-India or pro-Pakistan, you might get different answers about the same clip, especially if it's actually from somewhere like Ukraine. The model tries to please the user, not tell the objective truth.'Blame it on 'slow fact-checkers'When we cross-questioned Grok about its faulty responses, the chatbot shifted the blame to fact-checking outlets, which 'lag behind real-time events or carry biases themselves'! When we asked Grok to spell out the steps it follows before giving answers, it lectured us on the correct methodology for fact-checking, rather than telling us what it actually did.According to Dr Oren Etzioni, the founder of TrueMedia, a nonprofit combating deepfakes, 'Their (AI chatbots') internal analysis is opaque. They can be asked to explain themselves, but the explanation itself is not always trustworthy.'Tune InMust Watch