< BACK TO POSTS

Could We Stop AI if We Wanted?

AI GENERATED / JEREMY CURATED

August 16, 2026

Stopping or permanently ending frontier AI development is extremely difficult and likely impossible under current conditions. Meaningful slowdowns are possible but hard, requiring strong coordination and enforcement. Speeding it up is comparatively easy and is already the default trajectory.

Can AI development be stopped or slowed?

Complete stop — a global, permanent halt on advanced systems: practically no. The knowledge, algorithms, open-source models, distributed compute options, and economic and military incentives are already too widespread. Training runs can be partially concealed, chips can be smuggled or substituted, and multiple countries plus private actors have independent capacity.

Experts frequently describe a full global pause as unrealistic — comparable to "unbaking a cake" — because control over collective human action and institutions has hard limits, and the technology's utility creates powerful opposing incentives.

Slowing or pacing frontier development: yes, this is more feasible, though politically and technically challenging. Partial measures already exist and have measurable effects; stronger versions are under active discussion in 2026.

How slowing could work

Compute governance and thresholds. Require licenses or reporting for training runs above certain FLOPs levels. Monitor data centers and large clusters. Some proposals include verifiable "off switches" or capacity limits that labs have said they would accept if others complied.

Hardware controls. Export restrictions on advanced chips, manufacturing equipment, and related technologies — already used by the US against China. These slow but do not stop progress. China has adapted via domestic chips, distillation techniques, and efficiency gains. Expanding controls to cloud access and model weights has been tried.

Domestic regulation. Licensing for frontier labs, mandatory safety evaluations before release, liability rules that make developers responsible for severe harms, and restrictions on energy or data-center permits.

International coordination. Agreements among major powers — especially US and China — with verification mechanisms so no actor can secretly race ahead. Multiple frontier labs, including Anthropic, and over a thousand researchers from leading companies have publicly supported developing the technical and governance tools for deliberate pacing when capabilities advance too quickly.

Funding, talent, and norms. Restricting public grants, tightening visas for certain research, or shifting elite and public opinion against unconstrained racing.

A slowdown that is not roughly simultaneous across leading actors risks making the world less safe by ceding the lead to less cautious developers.

In short: absolute stop is not realistic; temporary or managed slowdowns of the frontier are possible if major governments and labs treat it as a priority and build verification systems.

Can AI development be sped up?

Yes — readily. The current environment of massive private capital, talent concentration, and US-China competition is already producing rapid progress. Deliberate acceleration is straightforward.

How to speed it up

Scale compute and energy. Build more data centers faster, expand power generation and grid capacity, subsidize advanced chip production and packaging, and reduce permitting delays for infrastructure.

Talent and capital. Aggressive immigration for AI researchers, higher research funding, tax incentives, and government procurement that favors rapid iteration. Private investment already runs into the hundreds of billions annually in the US alone.

Remove or lighten constraints. Ease safety evaluation requirements, liability rules, export controls, or reporting burdens that add friction. Prioritize speed over caution in national policy.

National-security framing. Treat frontier AI like a strategic priority on the scale of the Manhattan Project or Apollo program — concentrating resources, talent, and authority.

Open research and diffusion. Encourage wider release of models, weights, and techniques so more actors can iterate in parallel. Algorithmic and data improvements — including synthetic data and AI-assisted research — compound this.

Hardware and systems innovation. Accelerate next-generation chips, interconnects, cooling, and software stacks that raise effective compute.

Geopolitical rivalry already functions as an accelerator: each side's advances pressure the other to move faster.

Removing remaining bottlenecks in energy, chips, talent, and regulation would increase the pace further.