6 min readNew DelhiSep 15, 2026 10:23 AM IST Earlier this week, Anthropic CEO Dario Amodei wrote a lengthy essay about the dire need to slow down the progress of frontier artificial intelligence (AI). His view was backed by rivals like OpenAI’s Sam Altman and SpaceX CEO Elon Musk, all three of whom have been involved in public disagreements in the past. Considering the influence these frontier AI companies wield on the broader industry, the bigger question is: is there a kill switch for AI? And can these companies actually exercise that if the situation calls for it?
In a recent interview with the BBC, Anthropic co-founder Jack Clark said that an AI kill switch may need to be mandatory for companies developing the technology at the frontier. Clark believes that a way of shutting off AI systems completely if they get too dangerous is something society may want to eventually pass rules around.
Clark explained that most AI labs have different ways of pulling the plug on their AI systems. However, he is of the opinion that lawmakers would need to make such safeguards mandatory. The fears around AI’s rapid development and the risks it poses to humanity have been in the spotlight for the past few days after a spate of warnings from executives from AI companies. When it comes to the kill switch, Clark told the BBC that the specifics around the kill switch, such as the technical requirements and how it would be verified and enforced, should be part of the larger policy debate around AI.
The debate gained momentum after Anthropic’s former AI researcher Jacob Coxon claimed that top AI labs were racing straight to self-improving superintelligence and gambling with lives. He went on to claim that if unchecked, the technology could eventually eliminate humanity by the end of the decade.
Close on the heels of the revelation, Amodei called for the pace of AI development to slow down and be closely monitored. He sought to establish independent third-party evaluators working inside AI companies with employee-level access to the frontier models, coordination among democracies to establish safety standards, and global cooperation between nations, including major rivals like China.
Transparency over self-policing
Even as Amodei calls for the industry to slow down AI development, a veteran security researcher has argued that the real issue is not speed but its honesty. According to Gary McGraw, CEO of the Berryville Institute of Machine Learning, the current alarm over AI risks comes from a fundamental misunderstanding of what these systems actually are.
In an interview with DW, McGraw said that AI companies often misleadingly describe their models with human-like terms, as though they escape or lie, when in reality any harm goes back to how humans built and deployed them. He applies the same reasoning to the recent security incidents involving frontier models, including the OpenAI-Hugging Face incident. Instead of treating it as evidence of runaway AI, McGraw called it a case of poor engineering, arguing that the sandbox was ‘very badly engineered’ and not necessarily a system acting on its own.Story continues below this ad
When it comes to self-regulation, McGraw is wary; he dismissed the idea of including third-party evaluators inside AI companies. He compared this to expecting cult members to objectively report on their own cult. McGraw wants regulators to make transparency, such as disclosures about underlying architecture, training data and testing regimens, transparent, instead of dictating what companies can or cannot build. When asked if the AI companies have truly acknowledged the dangers posed by some of their models, or if it is a bit of PR, McGraw was quick to say that it is indeed a lot of PR.
This scepticism is reflected in the growing divide over calls for the industry to collectively slow down AI development. Following the tacit agreement between Altman, Musk, and Amodei, two factions emerged – the frontier labs, which are in support, and critics who see it as a smokescreen. The latter argues that the AI companies can slow things down if they want to, and the successive claims about the dangers of AI amount to little more than theatrics.
Venture capitalist David Sacks, in his post on X, slammed Anthropic and OpenAI. He said that since both companies are at the frontier of AI development, they do not really need permission. Sacks said that he would support the decision to slow down if the unreleased models are genuinely scary. However, he asked the AI companies to stop pretending that the motivation to slow down is purely altruistic.
AI researcher Eli David, who has been critical of Anthropic for their recent publications, offered another explanation. He argued that Anthropic and OpenAI are delaying their IPO as their S-1 (mandatory filings) would reveal that they are incurring losses and have no apparent path to profitability. David claims that the ‘AI slowdown’ appears to be a solution to these companies, as they can cut costs for training new models. “It has everything to do with IPO, and nothing to do with safety,” read his post on X.Story continues below this ad
This disbelief persists as warnings about AI’s potential impact continue to rise. Last week, a post from a former AI researcher who quit Anthropic about concerns that AI could wipe out humanity went viral. In another instance, Anthropic scientist Evan Hubinger said that he personally thought the possibility of human extinction from AI was greater than 10 per cent within the next decade. The result is an increasingly polarised debate, with some researchers and executives calling for slower development, stronger oversight, and mandatory mechanisms to shut down dangerous systems, while others argue that the existential threat is overstated, or warnings are being used to generate hype. The question may not simply be whether AI needs a kill switch. It would be who should control it, who gets to decide when it’s activated, and whether the companies building the technology can be trusted to pull the plug themselves.