New Delhi, Sept. 10 -- An exchange with Google Gemini over the District Magistrate of Gonda in 1947 offered an unsettling lesson: AI can be remarkably persuasive even when it is wrong.
While preparing to analyse a District Agri-business Plan for Gonda district in Uttar Pradesh, I was doing something increasingly routine: checking facts through Artificial Intelligence. I happened to be using Google Gemini. Reminiscing about my own tenure as District Magistrate in Rajasthan, I asked: Who was the District Magistrate of Gonda on the eve of Independence?
The answer identified K. K. K. Nair, an Indian Civil Service officer, and added context about his subsequent role as District Magistrate of Faizabad in 1949 and association with the early Ram Janmabhoomi-Babri Masjid dispute. It sounded convincing. It was wrong.
I knew this because the District Magistrate of Gonda on August 14-15, 1947, was my maternal grandfather (Nana), Shri Radha Mohan. I challenged the answer. Gemini reversed itself, explaining that Nair had been transferred in July 1947, after which Radha Mohan had taken charge. It acknowledged that it had conflated an officer associated with the preceding period with the officer actually holding the post on that date.
Then I became curious. Growing up, I had heard many stories from my grandfather about Independence and his tenure at Gonda. I wondered what traces of those years might be available to AI. So I asked: "Tell me some more about Shri Radha Mohan."
Gemini described the tensions surrounding Partition and said Radha Mohan had taken over Gonda's administration in July 1947. It then suggested that his tenure was relatively brief and attributed to him the role of overseeing the transition of power and hoisting the first Tricolour at the Collectorate.
The narrative seemed plausible. It was again wrong. My grandfather remained District Magistrate there until September 1949. I knew this because, in an attempt to get a slice of my heritage, I had obtained a copy of the Incumbency Board from the Gonda DM's office some years earlier. I told Gemini to check again. It apologised and acknowledged that Radha Mohan had served as District Magistrate/Deputy Commissioner from 1947 to 1949.
What if I had not known? What if I asked AI 10-15 years from now, with fading memories? What if my daughter were asking it some years from now?
She could have gone away believing that K. K. K. Nair was the District Magistrate of Gonda on the eve of Independence. Why would she doubt it? The answer contained dates, historical context and citations. It sounded researched. And therein lies the problem with AI-generated information.
An incorrect answer does not necessarily look incorrect. The most dangerous AI errors are plausible answers, carefully worded and delivered with confidence, containing a small but consequential distortion. Future generations might regard such an answer as more authoritative than family recollection. Over time, the machine-generated version could become accepted history, allowing a real person to disappear from the past because AI had confidently filled a gap with a plausible reconstruction.
This is why I have begun thinking of AI as an AI Prism rather than an oracle. A prism refracts light; it does not create the light itself. AI similarly refracts an enormous body of information-documents, websites, books and databases-through statistical patterns and associations. What emerges is not automatically a verified historical record. It can be a synthesis, an inference, a conflation of people or periods, or a coherent story assembled around an incorrect premise.
This is commonly called AI hallucination. But in matters of history, public policy, law or governance, a better description might be confident reconstruction without adequate verification.
There is another lesson. We have been taught to trust information when it comes with citations. AI has complicated that instinct. A citation can be genuine while the conclusion drawn from it is wrong. A source showing that an officer served in a district in 1946-47 does not establish that he held the post on August 15, 1947. A few months-or days-can completely change a historical answer.
Ironically, an old family memory challenged Artificial Intelligence. Memory is not infallible, but AI and human knowledge have different strengths; responsible research requires the two to interrogate each other.
The answer is not to reject AI. But we should ask it to show uncertainty, provide precise dates and primary sources, identify contradictory evidence, and distinguish between what is established, inferred and uncertain. When the issue matters, return to the original record.
The most important skill in the age of AI may therefore be the willingness to ask one more question: "Are you sure?"
In my case, that question protected a small piece of family history. But if we increasingly use machines to remember our past, we must ensure that their confidence does not become a substitute for historical truth.
Views expressed are personal. The writer is a former IAS officer
Published by HT Digital Content Services with permission from Millennium Post.