Professor B Ravindran of IIT Madras argues that fears about AI ending humanity often skip the crucial middle steps: access, permissions, testing decisions and human control. His interview offers a useful counterweight to the current slowdown debate, without pretending frontier AI is harmless.
Watch Desk analysis
What happened
In an interview with Rediff, Ravindran, who leads several data science and responsible AI centres at IIT Madras, challenged the idea that catastrophic AI outcomes follow automatically from more capable models. He said current systems are computer programmes optimised for objectives people specify, not sentient actors with an innate desire to harm people.
His central point is practical rather than philosophical. For an AI system to cause real-world catastrophe, it would need access to consequential systems, inadequate safeguards and a chain of human decisions that allowed those conditions to persist. Ravindran said companies should improve technical controls, sandboxing and kill switches, while governments should set clear liability rules rather than simply demand a blanket slowdown.
Why it matters
The AI safety argument is increasingly framed as a choice between racing ahead and stopping altogether. Ravindran says that is a false binary. He supports stronger controls and more careful testing, but argues that a global slowdown could burden smaller developers and open-source projects while protecting the lead of the largest frontier companies.
He also questioned dramatic claims about extinction probabilities when the proposed pathway is not clearly specified. That is an important distinction: a risk estimate is only as useful as the assumptions beneath it, and percentages can acquire an impressive sheen rather faster than they acquire a sturdy foundation.
Ravindran attributed several claims about recent AI incidents and safety failures to accounts discussed in the interview. Those details are not independently established by this report. His broader argument is clear: companies should not test powerful systems in poorly isolated environments, then treat the resulting behaviour as evidence that the machines acted without human involvement.
Our read
This is a valuable intervention because it moves the conversation from cinematic intent to engineering responsibility. “The model might become dangerous” is a serious concern, but it is not a safety plan. Access controls, robust sandboxes, independent evaluation and clear liability are the less glamorous machinery that determines whether a system can do damage.
Ravindran may be too confident that catastrophic pathways can always be interrupted. But his challenge to vague probability claims is fair, and his proposed focus on purpose-built models and stronger testing gives the debate something more useful than another round of existential fog.
What to watch
- Whether frontier labs publish enough detail about sandboxing failures and corrective actions to support independent scrutiny.
- Whether governments pursue enforceable liability and auditing rules rather than broad calls for restraint alone.
- Whether independent evaluators can be genuinely independent when expertise remains concentrated inside major AI companies.
- Whether smaller and open-source developers receive workable safety requirements rather than the bill for frontier labs’ race. Sources & Evidence: Rediff interview with Professor B Ravindran, published 24 September 2026. The interview supports Ravindran’s attributed views on AI risk, human control, guardrails, regulation, testing and the case against a blanket global slowdown. Claims about incidents mentioned within the interview remain attributed and were not independently verified here.
Discussion spark: Should AI policy focus first on controlling access, testing environments and liability, or is a slowdown still necessary before those safeguards can be trusted?
Sources and evidence
- 'When You Say AI Is Going To End Humanity…' – Rediff (24 September 2026, 06:58 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.