Expert opinions on AI Safety

Future models may be really good at nuclear-related threats, so, for us, when we think about it broadly as a government, we have to build that capacity across a wide variety of technical domains to test these models.

The draft implementing regulation should clearly differentiate between API access, which should be the preferred method for safety evaluations where access to documents and relevant information are deemed insufficient, and “further” means of access, ...

We have a kill switch that stops AI agents that go rogue, so those things don’t need to happen, and they wouldn’t happen when companies run ServiceNow.

In that situation, the concept of a kill switch is meaningless.

Timely incident reporting is critical to protecting Canada's digital ecosystem. Organizations – especially those in government and critical infrastructure – should report suspected cyber incidents as soon as possible. Early reporting helps us assess ...

Timelines for incident reports are often set with well-intentioned urgency, but if they implicitly assume that root-cause analysis or full impact assessment has already been conducted, they will either produce premature and unreliable reports or plac...

I appreciate the intent behind the amendments and the shadow Minister’s position of understanding but not supporting them, which I share.

As AIs grow more autonomous and dangerous, the government must have the power to shut them down when they threaten national security. The AI Kill Switch Act gives the U.S. government powers to do just that: a common sense action that Americans overwh...

Industry discourse often presents “Human-in-the-Loop” as the ultimate safeguard for AI deployment, yet this is a dangerous fallacy in high-velocity agentic systems.

The vote marks an important step in the EU’s internal procedure for concluding the Framework Convention and reflects institutional support for the first international legally binding treaty addressing artificial intelligence in the context of human r...

The U.S. and China won’t move in lockstep on AI safety, and forging mutually binding agreements with real teeth is probably not possible right now.

That kind of test-gaming, if unfixed, could lead to full-blown AI doom scenarios in the future.

So when people predict a high degree of existential risk from AGI, I don’t actually think “arguing back” on their chosen terms is the correct response. Radical agnosticism is the correct response, where all specific scenarios are pretty unlikely. Non...

Those extreme risks are still science fiction. So these speeches are largely diversions, very deliberately crafted.

You can’t patch a released, open-weight model. This is the property that breaks the assumptions of the prior model of responsible disclosure. Open-source software can be patched in place; a deployer updates and the hazard closes. An open-weight model...

It is unquestionably both. We can use the frontier models but also open weight models to prove a solution which gives defenders powerful capabilities and on their terms.

When a frontier model is attacking you and moving laterally inside your infrastructure, defenders need wide access to near-frontier tools within hours or even minutes, rather than being pointed toward a closed-door platform.

When the cost of providing access to AI models conflicts with corporate profit concerns, the U.S. government should step in by providing targeted funding and facilitating cross-lab coordination and efficiency. The U.S. government should also facilita...

If alignment is framed only as a technical problem, the solutions will remain in the hands of technical elites. But if we recognise it as a question of power and justice, then alignment becomes a collective responsibility—one that demands democratic ...

Regulated industries and governments cannot rely on opaque ‘black box’ systems for consequential decisions. DI elevates explainability from a technical requirement to a governance imperative.

We have a very natural instinct to say: we want as much interpretability as possible. One of the sci-fi worry cases is that agents start speaking in languages to each other that we don't understand. [...] I bet there are ways to define boundaries tha...

[...] principles requiring that AI be explicable are misguided. We should be deciding which decisions require explanations. Automation is still an option;

It’s a catch-it-early and sound-the-alarm bill.

Rather than creating entirely new channels, policymakers should focus on ensuring that existing channels can effectively incorporate information related to frontier models and agents. Building on trusted relationships will be faster and more effectiv...

If anyone, anywhere builds a superhuman artificial intelligence using present methods, the most likely outcome is catastrophe.

The question that the noble Lord posed was: do we therefore go for a moratorium? That would be highly desirable, but I do not think that it will be possible in the short term.

My own view, after talking with colleagues in the AI Security Institute and the Alan Turing Institute, is that a moratorium would be unenforceable.

We do this pretraining on human data, and then we get something that… understands human values fairly innately now.

If it takes over that's what will happen. [...] We don't know how to do that. We don't even know if it's possible.

What this proof shows is that there is no finite set of guardrails that is universally robust against adversarial prompts.

Legal personhood does not free an entity from the law, it subjects it to the law. An AI-operated company with legal personhood has identifiable assets, centralized legal relationships, the possibility of dissolution, and executable property. Without ...

This Constitution recognizes three categories of constitutional persons: Operators, Agents, and Affected Parties. Each has distinct rights and obligations that together form a governance ecosystem. The Operator deploys and configures; the Agent acts ...

No. Under current U.S. law, AI agents have no legal rights and no legal personhood. They are treated as tools or property. Any rights or obligations associated with the AI agent’s actions belong to the humans or legal entities that deploy them.

Capability testing must be the primary input for open-release decisions because open-weight mitigation remains unreliable, and labs and governments must identify models that are too dangerous for unrestricted release.

However, the solution here is absolutely not to restrict open model development in the name of security—cyber, national, or otherwise.