“Developing frontier AI models should require a government licence.”
Some lab leaders and researchers have proposed licensing for the most capable AI systems, as for drugs, aircraft or nuclear plants. Others see licensing as regulatory capture that entrenches incumbents, slows beneficial progress and hands an advantage to rival countries.
As of mid-2026 no major jurisdiction requires a licence to train a model. The EU AI Act imposes evaluation and reporting duties on general-purpose models above a compute threshold; California's SB 53 requires frontier developers to publish safety frameworks and report incidents; the US federal government has moved to preempt state rules rather than add new ones. The crux is how much catastrophic risk you assign to the next models, and whether licensing would actually reduce it.
Prior permission from a regulator, conditional on safety evaluations, before training or deploying a model above some threshold. Stronger than disclosure or reporting duties.
The most capable general-purpose models, usually defined by training compute (e.g. above 10^25 or 10^26 operations) or by capability evaluations. Compute thresholds age quickly as efficiency improves.
Many AI researchers think there is a non-trivial chance of catastrophic outcomes from advanced AI. We license activities whose failures could be catastrophic and hard to reverse; developing the most powerful models fits that category.
In a 2023 survey of 2,778 published AI researchers, the median respondent put a 5% probability on outcomes from advanced AI as bad as human extinction. Hundreds of researchers and lab leaders signed the 2023 Statement on AI Risk calling extinction risk a global priority.
Grace et al. (2024). Thousands of AI Authors on the Future of AI. arXiv:2401.02843; Center for AI Safety (2023). Statement on AI Risk.
Limitation
Surveyed probabilities vary wildly by question framing and are speculative. A precautionary licence for a speculative risk could impose large real costs.
Voluntary commitments do not hold under competitive pressure. A licensing regime creates a floor that applies to every developer, so safety-conscious labs are not punished for slowing down.
OpenAI's chief executive told the US Senate in May 2023 that a new agency licensing efforts above a scale of capability would be a sensible step.
Testimony of Sam Altman, US Senate Judiciary Subcommittee on Privacy, Technology and the Law, 16 May 2023.
Limitation
Support from incumbents is exactly what critics expect if licensing protects them from competitors.
Evaluations already find frontier models giving meaningful uplift on dangerous tasks such as cyberattacks and aspects of biological weapons work. Licensing would make passing such evaluations a condition of release rather than a voluntary choice.
In 2025 Anthropic deployed Claude Opus 4 under stricter "ASL-3" safeguards because it could not rule out meaningful uplift for biological weapons, and OpenAI classified ChatGPT Agent as "High" biological capability under its Preparedness Framework.
Anthropic (May 2025). Activating AI Safety Level 3 Protections; OpenAI (July 2025). ChatGPT Agent System Card.
Limitation
Knowledge benchmarks do not measure real-world uplift. Evaluation science is immature, so licence decisions would rest on shaky tests.
Licensing raises fixed costs and creates a gatekeeper that incumbents can influence, entrenching the largest labs and shutting out startups, academics and open-source developers.
Regulatory-capture concerns about frontier licensing have been raised across the political spectrum. California's SB 1047, which would have imposed liability-based duties on large models, was vetoed in 2024 partly over concerns about stifling innovation; the narrower transparency law SB 53 passed in 2025.
Governor Newsom veto message on SB 1047, 29 Sept 2024; California SB 53 (Transparency in Frontier Artificial Intelligence Act), 2025.
Limitation
Licensing thresholds can be set high enough to apply only to a handful of the best-funded developers.
Compute thresholds are a poor proxy for danger. Training efficiency improves quickly, so a threshold set today will either capture too much or let capable models through.
DeepSeek reported training its V3 model, competitive with leading Western models at release, with roughly 2.8 million GPU-hours, far less compute than comparable models, in late 2024.
DeepSeek-AI (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437.
Limitation
Licensing could use capability evaluations rather than compute. That is harder to define but avoids the proxy problem.
Unilateral licensing slows domestic developers without slowing rivals, potentially shifting frontier development to countries with weaker safety norms.
US policy since 2025 has framed AI as a strategic competition with China, and a December 2025 executive order sought to preempt state AI rules to avoid a "patchwork" that could slow US developers.
Executive Order 14365 (Dec. 2025), Ensuring a National Policy Framework for Artificial Intelligence.
Limitation
Coordinated international regimes (like nuclear safeguards) are possible in principle, and race dynamics are themselves a risk argument for regulation.
Current frontier models can provide some uplift on cyber-offence tasks, and evaluation methods are immature.
The EU AI Act already imposes evaluation and reporting duties on the largest general-purpose models.
Any regime depends on defining "frontier" in a way that stays meaningful as efficiency improves.