Precautionary Containment Gate
Slow high-impact capability until its harms can be credibly controlled
- Difficulty
- Expert
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 97%
Suleyman argues that unusually capable dual-use technologies should pass a precautionary gate before broad release. Define the dangerous capability and the scale of possible harm, then identify the inputs that make it possible: compute, model software, cloud access, chips, APIs, or biological materials. Restrict access to the highest-risk combinations while safety measures and audits are developed. Use physical and economic choke points where they exist, coordinate internationally to reduce regulatory flight, and reassess as capability, cost, and access change. The mechanism is containment through friction: fewer actors receive the means to create an existential threat, and deployment slows until control is more credible. Suleyman presents this as a policy imperative, not proof that containment will succeed.
Origin
Suleyman develops containment as the central challenge of The Coming Wave. In the interview he applies the precautionary principle to advanced AI and synthetic biology, arguing that society may need to forgo some benefits until control is more credible.
Core principles
- 01High upside does not cancel catastrophic downside
- 02Access controls can target dangerous capability rather than all innovation
- 03Forgone upside can be justified while containment remains uncertain
- 04Regulation is necessary but insufficient without technical and cultural participation
- 05The burden rises with autonomy, scale, and irreversibility
How to run it
- 1
Define the dangerous capability
Specify the behavior, access pattern, and impact that would cross the containment threshold.
Pro tip Distinguish a demonstrated capability from a forecast about future systems.
Watch out A vague label such as powerful AI cannot support a proportionate control.
- 2
Map access and proliferation
Identify who can obtain the tools now, how costs are changing, and how open release could expand the actor set.
Pro tip Revisit the map as yesterday's frontier becomes cheaper and smaller.
Watch out Restricting one provider does not contain a capability available elsewhere.
- 3
Find the choke points
Locate controllable dependencies such as advanced chips, large compute clusters, model weights, APIs, cloud environments, or hazardous substances.
Pro tip Choose controls closest to the dangerous capability rather than burdening every user.
Watch out Choke points can create concentrated power and require oversight.
- 4
Install friction and audits
Limit access, monitor use, test system behavior, and require evidence before capability or autonomy expands.
Pro tip Pair policy controls with technical safety work and independent scrutiny.
Watch out Regulation written without technical understanding may miss the actual risk.
- 5
Coordinate beyond one jurisdiction
Align the main states and organizations so a local restriction does not simply move development elsewhere.
Pro tip Start from shared exposure to harm rather than adversarial rhetoric.
Watch out Coordination is an objective in the interview, not an established outcome.
- 6
Reopen only with evidence
Expand access when safeguards, monitoring, and containment evidence justify the additional risk.
Pro tip Set review triggers tied to capability and cost, not only calendar dates.
Watch out Commercial pressure is not evidence that a system is contained.
In the wild
A cloud provider finds that a frontier model paired with specialist biological tools could enable dangerous experimentation. It limits that workflow to vetted researchers, logs use, commissions audits, and expands access only after safeguards demonstrate reliable controls.
→ Access remains available for legitimate work while the highest-risk combination receives additional friction and oversight.
Common mistakes
Treating all access as equally dangerous
The framework targets capability and combinations of inputs; indiscriminate restriction can suppress low-risk benefit without improving containment.
Assuming national regulation is enough
Suleyman argues that technical development, culture, and international coordination must accompany regulation.
Is it for you?
Best for
It is best for powerful dual-use systems where access, compute, software, infrastructure, or hazardous materials can be controlled.
Not ideal for
It is not ideal as a blanket justification for restricting low-risk tools or concentrating power without oversight and evidence.
From the transcript
“We have to approach this with the precautionary principle.”
“Give up some of those potential upsides or benefits.”
“Choke point refers to points in the supply chain where you can throttle who has access to what.”
From the episode
Google DeepMind Co-founder: AI Could Release A Deadly Virus - It’s Getting More Threatening! Mustafa Suleyman