Containing AI Is Possible
History has contained transformative and dangerous technologies before, nuclear energy, genetic engineering, autonomous trading. AI requires the same treatment: not prohibition, but verifiable, phased, international containment.
What the Experts Say
Researchers and policy scholars have studied arms control, verification and here's what they think about containing AI.
“It's not risks vs. benefits, it's you can't have the benefits unless you address the risks... Is nuclear technology a force for good? If we choose to use it for making cheaper electricity without pollution then, yes, but if we choose to use it to make weapons to kill each other then, no.”
— Prof. Stuart Russell, UC Berkeley, AI for Good Global Summit
Watch the Interview“The type of regulation you need is regulation such that, if you comply with it, the risk is reduced to an acceptable level. If frontier AI companies can't comply, they shouldn't be permitted to release their systems.”
— Paul Scharre, Center for a New American Security
Read the Article“Governing AI will require widely sharing its benefits while keeping the most powerful AI out of the hands of bad actors. The good news is that there is already a template on how to do just that. In the 20th century, nations built international institutions to allow the spread of peaceful nuclear energy but slow nuclear weapons proliferation by controlling access to the raw materials, namely weapons-grade uranium and plutonium, that underpins them. Governments can control access to the highly specialized chips that are needed to train the world's most advanced AI models.”
— Paul Scharre, Center for a New American Security
Read the ArticleWe Have Contained Dangerous Technologies Before
From nuclear energy to autonomous trading, societies have repeatedly chosen containment over prohibition. The same pattern applies to AI.
Nuclear Energy
- •Too powerful to abolish after the Manhattan Project
- •Governed by the NPT, IAEA inspections, and reactor safety standards
- •Still in use today - under constant oversight
Genetic Engineering
- •Scientists halted work at Asilomar (1975) to assess the risks
- •Created BSL labs and institutional review boards
- •Research resumed - under strict safety rules
Autonomous Trading
- •The 2010 Flash Crash proved automated trading needed limits
- •Introduced circuit breakers, kill switches, and latency floors
- •Trading kept running - with built-in speed limits
Artificial Intelligence
- •Too economically valuable to abolish — too dangerous to leave unchecked
- •Model audits, pre-deployment safety evaluations, and output filters
- •Regulatory frameworks emerging: EU AI Act, US executive orders
- •The same pattern again — containment, not prohibition
The pattern is unmistakable. When a technology is too transformative to stop and too dangerous to leave unchecked, societies choose containment. AI requires the same treatment.
Most Promising Near-Term Paths
Most researchers converge on a phased approach. Each phase builds the foundation for the next. Skipping phases is how treaties fail.
Scientific Consensus
IPCC ModelBuild shared scientific consensus on risks through an independent international panel. Independent risk assessment must precede any binding action. No treaty can succeed if the parties do not agree on what they are trying to prevent.
Reporting Norms
Voluntary but InstitutionalizedEstablish voluntary but institutionalized reporting norms so labs disclose training scale and safety evaluations. This builds transparency, creates baselines, and normalizes accountability before mandates are politically feasible.
Standing International Body
With Technical StaffCreate a standing international body with technical staff to track progress, verify reports, and build trust over time. This body must have real expertise, not diplomats alone, and must be funded sufficiently to monitor frontier development.
Binding Commitments
Once Trust ExistsOnly later, once trust and verification tools exist and are proven, move toward binding commitments. Attempting binding rules before verification is possible, as the BWC demonstrated, produces empty paper.
The Compute-Monitoring Angle
Most Tractable Near-Term VerificationTracking the physical hardware required for frontier training, Nvidia H100s and successors, is widely seen as the most tractable near-term verification mechanism. Unlike inspecting code or model weights, compute is physical, traceable, and countable. It does not require understanding what a model does; it only requires knowing that a training run of a dangerous scale occurred. This is the foundation on which every other verification layer can be built.
- Advanced AI chips are manufactured at a handful of facilities worldwide
- Export controls on cutting-edge semiconductors have already been implemented (US restrictions on China, 2022)
- Cloud providers can be required to report large-scale compute usage and verify customer identity
- Hardware-level tracking does not require inspecting proprietary model weights or source code
Existing Analogous Frameworks
We have contained dangerous technologies before. Each precedent offers lessons for how to approach AI governance.
Nuclear Non-Proliferation Treaty
1968The most-cited analogy for AI containment. The NPT established three pillars: non-proliferation, disarmament, and peaceful use.
AI Mapping
- Non-proliferation → Preventing dangerous AI capability development
- Disarmament → Reducing existing risks from deployed frontier models
- Peaceful use → Enabling beneficial AI applications under oversight
Chemical Weapons Convention
1993Often considered a stronger model because it includes a dedicated verification body (the OPCW) with actual inspection powers.
AI Mapping
- An 'AI Weapons Convention' could establish an international body empowered to audit training runs
- Compute usage and deployment systems could be subject to inspection, analogous to chemical stockpile checks
- Unlike chemical precursors, AI capabilities are largely software, invisible, copyable, and dual-use by nature
Biological Weapons Convention
1972Instructive as a negative example. It banned bioweapons but has no verification mechanism whatsoever, making compliance essentially voluntary.
AI Mapping
- Banned bioweapons by treaty but provided no means to verify labs or stockpiles
- Relies entirely on national self-reporting and goodwill
- Most AI policy experts argue strongly that any AI treaty must avoid this failure mode
IAEA Safeguards Model
1957The International Atomic Energy Agency monitors nuclear material through on-site inspections, material accounting, and satellite surveillance.
AI Mapping
- Monitoring large compute clusters, Nvidia H100s and successors are trackable physical goods
- Auditing frontier labs for undeclared training runs
- Requiring 'know your customer' obligations on cloud providers for large-scale compute
Current Institutional Landscape
Several nascent structures exist. None yet binding. Understanding what is already in place is the first step toward making it work.
The Bletchley Declaration
2023Established a loose multilateral dialogue among ~30 nations, producing shared risk assessments but no enforcement mechanism.
- Signed by 28 countries plus the EU
- Acknowledged frontier AI risks but created no binding commitments
- Served as the foundation for subsequent summits in Seoul and Paris
AI Safety Summits
2023–PresentSeoul and Paris summits continued the Bletchley process, expanding participation but remaining dialogue-based with no verification or enforcement.
- Expanded beyond the original 28 to include additional nations and industry observers
- Produced voluntary safety commitments from some frontier labs
- No standing secretariat or institutional memory, each summit resets the conversation
UN High-Level Advisory Body on AI
2024Released recommendations for an international panel modeled on the IPCC, scientific consensus-building rather than regulation.
- Proposed a global scientific panel to assess AI capabilities and risks
- Explicitly modeled on the Intergovernmental Panel on Climate Change
- Focus is on building shared understanding before attempting binding rules
EU AI Act
2024The most ambitious domestic framework, creating risk tiers and mandatory requirements, but it is jurisdictional, not global.
- Bans social scoring and manipulation by government
- Creates mandatory pre-deployment testing for high-risk AI
- Limited to EU jurisdiction, does not cover training runs in the US, China, or elsewhere
OECD AI Principles
2019Soft norms adopted by ~50 countries, setting expectations for transparency, safety, and accountability without enforcement.
- Adopted by most OECD member states and several non-members
- Non-binding, compliance is voluntary and unmonitored
- Provides normative foundation that more binding frameworks could reference
Core Design Challenges
Any serious global framework runs into several hard problems. Pretending they do not exist is how treaties become empty paper.
These objectives are ambitious, and they require a movement behind them. Whether you sign a petition, volunteer, or donate, your action brings these goals closer to reality.