There’s every reason to slow AI development. But here’s why it probably won’t happen.
The rational case for easing off the accelerator is strong. Over the last few years, AI has moved from a niche research field into a general-purpose capability with unpredictable consequences. It can generate persuasive disinformation at scale, help amateurs carry out sophisticated cyber intrusions, turbocharge scams, and degrade public discourse. It will restructure labor markets faster than education and policy can adapt, amplifying inequality and regional shocks. It tilts power toward those who control data centers and supply chains. It consumes growing amounts of energy and semiconductors. And at the frontier, systems exhibit emergent, poorly understood behaviors, raising questions about reliability, control, and the risk of misuse in domains like bio, cyber, and the information sphere.
Slowing would buy time: time to harden institutions, develop better safety techniques, create economic buffers for displaced workers, sort out liability and auditing frameworks, and build international norms. A cautious society would stagger deployment, gate access to dangerous capabilities, and demand proof of safety before scale.
Yet a broad slowdown is unlikely. Not because the risks are imaginary, but because the incentives, technologies, and geopolitics make restraint fragile and coordination hard.
Why slowing is rational but improbable
– Competition creates a race dynamic. If one firm pauses, another will leap ahead. In winner-take-most markets with strong network effects, executives feel that slowing down risks losing the whole game. Shareholders can even sue leadership for forgoing obvious growth opportunities.
– Geopolitics rewards speed. National security establishments view advanced AI as a decisive edge across intelligence, cyber defense, logistics, and weapons systems. Governments are reluctant to accept constraints that could leave them behind rivals. Even if leading countries agreed to slow, holdouts and fast followers would be tempted to defect.
– Diffusion is hard to bottle. Research papers, code, and model weights spread quickly. Once training techniques and architectures are public, thousands of teams can replicate and incrementally extend them. Even when companies keep weights closed, open-source ecosystems eventually produce competitive alternatives.
– Regulation lags and struggles to bite. Lawmakers can write rules, but enforceability is the crux. It is difficult to verify what compute is used where, to define “frontier” systems with legal precision, to audit opaque models for dangerous capabilities, and to police jurisdictional arbitrage when cloud resources and talent are global.
– Hardware trends compound the problem. Compute becomes cheaper and more abundant, specialized chips proliferate, and algorithmic efficiency improves. Over time, “frontier” training runs that once required national-scale clusters become accessible to medium-sized firms or consortia.
– Culture and momentum favor scaling. A generation of researchers and entrepreneurs is trained to equate progress with larger models and faster deployment. Capital markets have poured tens of billions into AI infrastructure, creating sunk costs and quarterly pressures to ship.
– Fairness arguments cut both ways. Slowing can look like entrenching incumbents and high-income countries at the expense of smaller labs and the Global South. That makes a universal pause politically contentious and morally contested.
What has worked elsewhere—and why AI is different
The world has coordinated before. Nuclear proliferation slowed because the core materials are scarce, bulky, and detectable; the technology is capital-intensive; and international treaties can verify compliance. Biotech, despite dual-use risks, benefits from physical containment and licensing bottlenecks around pathogens and equipment. Climate action is harder: emissions and benefits are diffuse, free-riding is tempting, and the atmosphere is a shared sink. The internet spread fast because it was mostly software and incentives to connect were huge.
AI is closer to the internet and climate than to nuclear. The raw material is compute and data, both widely available and getting cheaper. Knowledge is codified and portable. The payoff from acceleration is immediate and local; many of the risks are delayed and global. That is a classic recipe for coordination failure.
What “slowing” could even mean
Calls to “pause AI” often blur distinct levers:
– Training slowdowns: limits on the size or frequency of frontier training runs.
– Deployment throttles: staged rollouts, access restrictions to powerful features, and mandatory evaluations before scaling to millions of users.
– Diffusion controls: limits on releasing model weights, code, or dangerous datasets.
– Compute governance: caps or permits for large training clusters, export controls on advanced chips, and telemetry on compute usage.
– Market pacing: tying deployment to safety benchmarks, third-party audits, or insurance requirements.
Each lever faces leakage. Compute can be rented through cloud providers in friendly jurisdictions. Weights can leak or be approximated. Firms can relabel or fragment training runs to avoid thresholds. Meanwhile, blunt tools risk stifling beneficial applications and smaller innovators while leaving well-funded actors relatively unaffected.
Levers that can still matter
The likely failure of a blanket slowdown does not make governance futile. It shifts the emphasis from “stop” to “shape.”
– Compute and capability thresholds. Require registration and safety plans for training runs above defined compute or capability levels. Embed hardware-based attestation in advanced chips and datacenters to support audits without revealing trade secrets.
– Independent evaluations. Fund third-party “model evaluation institutes” to test for dangerous capabilities (e.g., autonomous replication, high-skill cyber exploitation, biohazard assistance), systemic vulnerabilities, and reliability under adversarial prompts. Make evaluations a precondition for high-scale deployment.
– Liability and insurance. Clarify product liability for AI-enabled harms, require risk-based insurance, and let premiums reflect safety posture. Markets can create continuous pressure to harden systems.
– Secure-by-design standards. Treat AI infrastructure like critical software: secure training and inference environments, robust red-teaming, incident reporting, postmortems, and patch cycles. Tie compliance to procurement eligibility for public contracts.
– Provenance and authenticity. Deploy content provenance standards and watermarking where feasible; pair them with verifiable digital identity, bot labeling, and strong penalties for undisclosed synthetic manipulation in elections and financial markets.
– High-risk domain gating. In areas like bio, critical infrastructure, and autonomous cyber, require restricted-access models, strict user verification, and usage logging with privacy protection.
– International norms. Even imperfect agreements—on incident reporting, model evaluation interoperability, and red lines for weaponization—can reduce miscalculation and create shared expectations.
– Workforce and social policy. Invest in transition supports: portable benefits, wage insurance, rapid skilling, and regional development. Cushioning shocks makes societies more willing to regulate without fear of losing competitiveness.
Why these won’t add up to a true slowdown
Even if implemented, these measures mostly add friction where it matters most; they do not reverse the overall slope of progress. Compliance costs are manageable for major players. Open-source ecosystems will continue to backfill capabilities. Nations will prioritize strategic uses. And each safety breakthrough often enables more ambitious deployments, not fewer.
The “perimeter problem” persists: it is hard to cordon off know-how once published, and very hard to un-ring the bell once weights or recipes spread. Over time, better tools for sandboxing, cryptographic attestation, and monitoring can help, but they rarely outpace motivated actors everywhere.
The pragmatic stance
If we accept that a wholesale slowdown is unlikely, two priorities follow.
First, bend the trajectory. Push the ecosystem toward safer defaults, stronger guardrails, and better governance without presuming that we can halt frontier development. Aim to slow where risk is concentrated—frontier training runs, high-risk domains, and mass deployment channels—while letting low-risk, high-benefit applications flourish.
Second, build resilience. Assume rapid capability growth and plan for shocks:
– For governments: modernize agencies with technical talent; establish rapid-response units for AI incidents; stress-test elections and public communications for synthetic media floods; update education and civil service hiring to leverage AI safely.
– For companies: adopt internal risk taxonomies; gate access to powerful tools; log usage; practice incident response; and test supply chains and customer support for synthetic attacks.
– For institutions: invest in high-integrity media, open science on safety and evaluations, and civic capacity to handle more information, faster.
– For workers: provide lifelong learning infrastructure and portable credentials; encourage firms to pair automation with augmentation.
A note on optimism
Skepticism about a broad slowdown is not the same as fatalism. The story of other technologies shows that norms and institutions do catch up—often after near misses, sometimes after real harm. Early aviation was dangerous; now commercial flight is remarkably safe. Software security was an afterthought; now secure development lifecycles, bug bounties, and coordinated disclosure are industry staples. None of this happened overnight or by wishing for a pause; it happened through standards, liability, market discipline, and professionalization.
AI will be harder because it is general-purpose, cheap to copy, and culturally celebrated. But that only raises the stakes for channeling its momentum. Where restraint is feasible—on frontier training, on dual-use capability access, on election-season rollouts—it should be pursued vigorously. Where restraint will fail, we should prepare defenses, build buffers, and cultivate the muscle memory of responsible deployment.
There’s every reason to slow. There’s every reason to doubt we will. The work, then, is to make acceleration survivable and as broadly beneficial as possible, while carving out the specific cliffs where slowing is both necessary and achievable.
