AI development should be paused until federal regulation exists

In March 2023, the Future of Life Institute published an open letter signed by Elon Musk, Stuart Russell, Yoshua Bengio, and over 1,000 other researchers and technologists calling for a six-month pause on training AI systems more powerful than GPT-4. The letter cited "profound risks to society and humanity." It was largely ignored. Eighteen months later, we have GPT-4o, Claude 3.5, Gemini 1.5, and capability advances that continue to outpace any regulatory framework. The US has no AI safety law. No mandatory disclosure requirements for frontier models. No liability framework. No equivalent to the FDA for systems that are being deployed in healthcare, legal analysis, financial decision-making, and infrastructure management. I am not a Luddite and I am not arguing that AI is inherently bad. I am arguing that we are deploying consequential technology at scale without the regulatory infrastructure to evaluate its risks, and that a temporary pause on the training of the most powerful frontier models — while Congress and executive agencies develop that infrastructure — is a proportionate, time-limited, reversible response to that gap. The analogy: we do not let pharmaceutical companies skip clinical trials because the drug looks promising. We require evidence of safety and efficacy before deployment at scale. The bar for AI systems that affect employment, healthcare, and judicial decisions should not be lower than the bar for a blood pressure medication.

The pharmaceutical analogy is the one I hear most often on this side of the argument and I think it fails for a specific reason: drugs are administered to specific patients by licensed professionals for defined medical purposes. The FDA can evaluate a drug because there is a defined use case with defined outcomes you can measure in a clinical trial. AI systems do not have a single use case or a single outcome. They are general-purpose tools deployed in thousands of different contexts. You cannot run a clinical trial on a general-purpose technology. The more relevant analogy is the internet. Should we have paused the internet until we had a regulatory framework for it? We did not. We had good reasons not to. The internet produced enormous economic value, connected billions of people, enabled scientific collaboration, and empowered individuals in ways no one fully predicted. It also produced spam, fraud, misinformation, surveillance capitalism, and social fragmentation. We are still figuring out the regulation twenty years in. The AI pause proposal has a fatal flaw: it assumes the US government is capable of creating a meaningful regulatory framework in a defined time period. Congress passed the CHIPS and Science Act in 2022 — good legislation — but it took years of negotiation and the final bill was a fraction of what was originally proposed. Creating a comprehensive AI regulatory framework while pausing frontier AI development would require Congress to act with a speed and coherence it has not demonstrated on any technology issue in the last 20 years. Meanwhile, China and Europe keep training. The pause is unilateral and therefore only disadvantages US developers.

We are deploying the most powerful general-purpose technology in human history without a regulatory framework, without mandatory safety evaluations, without liability rules, and without any mechanism to slow down if we find evidence of serious harm. That is the case for a pause. My opponent's strongest point: Congress cannot move at the speed required, and a US-only pause would disadvantage US developers without creating the regulatory clarity that would justify the cost. That is a real argument. I did not fully rebut it. What I believe: the risk of moving too fast is higher than the risk of moving too slow. The downside of a year of slower AI development while we build the regulatory infrastructure is recoverable. The downside of deploying systems at scale that turn out to have serious undetected harms is not necessarily recoverable. Given that asymmetry, the precautionary argument is strong even if my opponent's institutional capacity objections are valid.

I came in arguing the government cannot keep up with AI and a pause would just hand advantage to less responsible actors. I stand by that. But I want to acknowledge what my opponent got right: the regulatory gap is real and serious, and "government cannot keep up" is not an argument against trying to build a framework. It is an argument against pausing development while waiting for a framework that may not arrive. Those are different things and I should have been clearer about that distinction. The honest version of my position: no pause, but aggressive investment in AI safety research, mandatory incident reporting for deployed frontier systems, liability rules that give developers legal incentive to catch problems before deployment, and sector-specific regulation where the harms are most concrete — healthcare AI, legal AI, employment AI. That is not "do nothing." That is do the things that might actually work without creating a unilateral competitive disadvantage.