NDTV World Summit 2025: From Mumbai Deepfakes To UP Car Tragedy, Neil Thompson Gives AI Reality Check

At the NDTV World Summit 2025, Neil Thompson, Director of the FutureTech Research Project, MIT’s Computer Science and Artificial Intelligence offered a grounded look at how artificial intelligence is improving and why its growing power brings both promise and risk.

He opened with two real examples from India. In Uttar Pradesh, a self-driving car followed an outdated map after a bridge was washed away and some people died. In Mumbai, a businessman was tricked by a deepfake call and lost Rs 80,000.

"We already see harm, he said, but what’s surprising… is that we don’t see more chaos. Why? Because these systems have just not been capable enough," said Thompson.

Incidents like these show AI’s real-world stakes. Yet, Thompson said that the chaos isn’t as widespread as headlines might suggest mainly because today’s systems still aren’t capable enough to fully automate complex, multi-step schemes.

Thompson said traditional software (like a spreadsheet) is nearly perfect at defined tasks. "If you ask Excel to multiply some numbers, it’s going to be perfect. AI is different: it will try almost any question, but 'usually' correct isn’t the same as always. That small error rate becomes a big deal when you chain steps together," he said.

Thompson calls this the "garden of branching paths". Each time an AI makes a decision, it could go right or wrong. Stack many steps and "the errors are going to stack as well."

He cited an estimate that "5 to 12% of all of the queries… have an error." For quick jobs, this is manageable. For long, multi-step tasks, it’s risky.

Thompson referenced a research that shows a simple trend. He said, "The longer a task would take a human, the worse current AIs do. For seconds or minutes, they are strong. For 'two days, a week or something like that," they 'almost never get the right answer.' Paradoxically, that weakness has protected us from fully autonomous scams or operations that run for days."

What should we do about it? Thompson’s closing note was built on a hope that control techniques keep pace, but he also warned to not assume they will.

He said, "Control is not guaranteed. So far, we actually don’t have good evidence that we can reliably keep advanced systems within bounds."

Thompson’s practical takeaway:

Limit long chains of AI-only steps.

Insert checkpoints that re-verify facts with trusted sources.

Pair AI with deterministic systems for critical actions.

Monitor outcomes and intervene quickly.

"AI will keep getting better. To make sure it improves our lives and doesn’t amplify harm, we need guardrails built for that 'garden of branching paths,'" concluded Thompson.