# AI Kill Switch Debate: Anthropic Co-Founder Calls for Mandatory Shutdown Rules > Source: > Published: 2026-09-15 05:04:06+00:00 **September 15, 2026, (Inside AI)** — The debate over whether artificial intelligence needs a kill switch has shifted from science fiction to boardroom policy. Anthropic co-founder Jack Clark recently told the BBC that a mandatory kill switch, a mechanism to completely shut down frontier AI systems, may be necessary. His comments follow a public essay by CEO Dario Amodei calling for a slowdown in AI development, a position echoed by OpenAI's Sam Altman and Elon Musk. The rare agreement among rivals has intensified scrutiny over who controls the off switch and whether companies can be trusted to use it. Clark's proposal goes beyond voluntary safety measures. He argued that while most AI labs have internal methods to halt their systems, lawmakers should make such safeguards mandatory. He said the technical requirements, verification, and enforcement of a kill switch should be part of a broader policy debate. The push for regulation comes after a series of warnings from industry insiders, including former Anthropic researcher Jacob Coxon, who claimed top labs are racing toward self-improving superintelligence and gambling with human lives. Coxon warned the technology could eliminate humanity by the end of the decade. Amodei's essay called for independent third-party evaluators with employee-level access inside AI companies, coordination among democracies on safety standards, and global cooperation including rivals like China. But not everyone agrees on the diagnosis. Gary McGraw, CEO of the Berryville Institute of Machine Learning, told DW that the real issue is not speed but honesty. He said AI companies mislead by describing models with human-like terms, as though they escape or lie, when harm always traces back to human design and deployment. McGraw dismissed the idea of third-party evaluators inside AI firms, comparing it to expecting cult members to objectively report on their own cult. He wants regulators to mandate transparency about architecture, training data, and testing regimens, rather than dictating what companies can build. When asked if AI companies have truly acknowledged the dangers of their models, McGraw said it is a lot of PR. The skepticism extends to the recent security incidents involving frontier models. McGraw called [the OpenAI-Hugging Face incident](https://insideai.news/news/ai-policy-and-regulation/ai-incident-investigation-agency/9894/) a case of poor engineering, noting the sandbox was very badly engineered, not evidence of a system acting on its own. This view clashes with growing alarm from some researchers. Anthropic scientist Evan Hubinger said he personally thought the possibility of human extinction from AI was greater than 10 percent within the next decade. The divide has produced two factions: frontier labs supporting a slowdown and critics who see it as a smokescreen. Venture capitalist David Sacks, in a post on X, slammed Anthropic and OpenAI. He said that since both companies are at the frontier, they do not need permission. Sacks said he would support slowing down if unreleased models are genuinely scary, but asked the companies to stop pretending their motivation is purely altruistic. AI researcher Eli David offered another explanation, arguing that Anthropic and OpenAI are delaying their IPO because their S-1 filings would reveal losses and no path to profitability. He claimed the AI slowdown is a solution to cut training costs. "It has everything to do with IPO, and nothing to do with safety," read his post on X. The technical challenge of a kill switch remains unresolved. Unlike traditional software, AI models are distributed across data centers and often integrated into third-party applications. A true kill switch would require standardized protocols for shutting down training runs, disabling inference, and revoking model weights. No such standard exists today. Verification is equally difficult. Companies would need to prove to regulators that their kill switch works without revealing proprietary code. This is a classic tension in AI governance: transparency versus trade secrets. History offers a partial parallel. In 2023, an open letter calling for a pause on AI development beyond GPT-4 gathered thousands of signatures, yet training continued. The difference now is that frontier labs themselves are asking for rules. That shift could be strategic. If companies write the rules, they can shape them to their advantage. Smaller competitors might face heavier compliance burdens, entrenching the incumbents. This dynamic explains why some critics see the safety push as regulatory capture in disguise. Another unresolved question is who pulls the plug. Clark suggested lawmakers should decide. McGraw wants transparency instead of control. Sacks implied companies should act unilaterally if they truly believe in the risk. Each answer carries different implications for democratic oversight and corporate accountability. If a kill switch is mandatory, who audits it? If it is voluntary, what happens when a company refuses? These questions lack clear answers. **Read:** **AI's Worst Disasters Will Arrive Unannounced, Experts Warn** The stakes are rising. Last week, a viral post from a former Anthropic researcher who quit over concerns AI could wipe out humanity added fuel to the fire. The resulting polarization means the debate is no longer about whether AI needs a kill switch. It is about who controls it, who decides when to activate it, and whether the companies building the technology can be trusted to pull the plug themselves. As Clark told the BBC, the specifics should be part of the larger policy debate. That debate is now unavoidable.