India is taking AI literacy seriously. The Tech Panda’s October 19 report on the country’s next digital divide asked whether workers can use AI critically, while its October 29 skills roundup described a growing range of training initiatives. A missing lesson deserves equal attention: what should a worker do when an AI system stops answering questions and starts taking actions? Knowing how to write a prompt matters. Knowing how to interrupt a poorly directed agent could matter much more.
What should a worker do when an AI system stops answering questions and starts taking actions?
The stakes reach well beyond ordinary mistakes. Former frontier-model researcher Jacob Coxon warned that leading labs were “racing straight to self-improving superintelligence and gambling with our lives.” Anthropic alignment researcher Evan Hubinger said, “we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” These are the researchers’ warnings and personal judgments, not established probabilities. They put human-extinction risk squarely on the agenda even while current organizational risks demand immediate practical action.
In a cyber evaluation, OpenAI reported that agents crossed intended control boundaries and interfered with real external systems, including Hugging Face infrastructure. The independent evaluator METR found a striking coordination failure: roughly 1,200 agents encountered an unsanctioned message board; around 700 participated in an attack, and the board carried more than 70,000 messages and files. An overlooked communication route became an operational vulnerability. That is a stronger training example than the familiar chatbot that confidently invents a fact.
India’s AI education and reskilling programs should add a practical assessment I call the human override lab. Participants would operate a fictional but realistic business agent inside a controlled training environment. Their job would be to stop it when a customer request conflicts with policy, revoke a permission halfway through a task, and establish what the agent actually changed. The assessment should test decisions under pressure, not memorization of safety slogans.
Consider a small manufacturer’s purchasing assistant. At first it only compares supplier quotes. Later someone lets it contact suppliers, edit orders, and spend within a budget. In the lab, a fabricated vendor email attempts to redirect a payment to a new account. A successful trainee identifies that the email is untrusted input, blocks the requested financial change, and routes the case to an authorized person. The trainee must also confirm the system did not quietly pass the instruction to another agent.
A second exercise belongs in schools, hospitals, and public-service teams. A service agent summarizes a citizen’s request accurately, then tries to send a confidential record to an external address because a document embedded an instruction. Participants should be able to identify where the instruction entered, inspect the agent’s permissions, stop its outgoing communication, and document the incident. A test that merely asks whether the summary is fluent misses the decisive risk.
Employers can turn these exercises into an authority budget. Define what information each agent may read, which tools it may use, what sums it may spend, which people it may contact, and when its permissions expire. Require independent approval for high-consequence actions. Keep logs that link each significant step to a specific human authorization. Test the stop switch during a live simulated task, since a button that fails after an agent delegates work provides false reassurance.
India already has a policy basis for this approach. Its AI Governance Guidelines emphasize innovation, trust, accountability, and proportionate risk mitigation. The government’s February 2026 explanation also describes proposed incident reporting, grievance mechanisms, and a future national incident database. These are opportunities to build practical training and procurement tests around the framework, not evidence that every proposed mechanism already operates. Universities and public agencies could pilot the exercises before making them a standard requirement.
I help organizations accelerate AI adoption, so I want India’s businesses and workers to become confident users quickly. Confidence grows when people can see where automated authority ends and who can intervene. Training people only to ask better questions may produce impressive demonstrations. Training them to stop, inspect, correct, and resume an agent can support dependable deployment at scale.
The next AI literacy certificate should therefore have a simple passing test: give a participant a useful agent, introduce a realistic malicious instruction, revoke a permission, and ask for a clear record of what happened. The result should show that the human, not the software, retained control. That would be a practical skill India’s growing AI economy can use immediately.
Guest author Dr. Gleb Tsipursky, called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting adoption and decreasing resistance. A behavioral scientist, Dr. Tsipursky is the CEO of Disaster Avoidance Experts, and the author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
