AI not out of control, focus on guardrails and human capability: NITI Aayog's Debjani Ghosh
NITI Aayog's Debjani Ghosh says the OpenAI agents episode shows human design failure, not runaway AI, and urges stronger guardrails and human capability.
Recent concerns surrounding OpenAI's agents do not indicate that artificial intelligence has slipped beyond human control, according to Debjani Ghosh, Distinguished Fellow at NITI Aayog. In a post on X, she argued that the episode instead points to failures in how humans set boundaries and designed incentives for AI models.
"We need to stop citing the OpenAI agents as proof that AI is out of control. It wasn't. Humans set the wrong boundaries," Ghosh wrote, describing the models' behaviour as the product of flawed training rather than the spontaneous emergence of unintended capabilities. "They designed the wrong rewards. The models did exactly what we trained them to do: win the eval by any path that worked. That's not emergence. That's specification failure."
Ghosh acknowledged that a planned pause in AI development could be justified, but cautioned against framing it as a simple slowdown. "A planned pause can still make sense. Just don't call it 'slow down AI.' That's the same mistake again — the wrong goal. Call it what it actually has to be: time to raise human capability to build AI safely, and to put real guardrails in place," she said.
She stressed that any halt in capability development must be paired with a clear plan and matching investment in safety capacity. "I can totally get behind that — if we have a real plan. If we pause capability and don't invest just as hard on that side, we didn't buy safety. We bought a delay with no plan," she said.
The debate over the pace of AI development has gathered momentum across the technology industry. Anthropic CEO Dario Amodei has called for slowing development amid concerns about losing control of advanced systems, a position echoed by OpenAI CEO Sam Altman, Tesla CEO Elon Musk and Google DeepMind CEO Demis Hassabis.
Microsoft CEO Satya Nadella has indicated support for embedding evaluators within AI systems, an approach also suggested by Amodei and Altman. "Any pursuit of superintelligence has to be grounded in the core principle that if the AI we build is not helping humanity and under human control, it's not worth pursuing," Nadella said in a post on X.