Expert opinions on AI Safety

For a contract analysis agent, harm is a missed liability clause that costs the company seven figures. [...] Human-in-the-loop review means defining, in advance, which decisions require human confirmation before execution. High-stakes, low-volume dec...

Existing responsible AI governance frameworks mandate human oversight. Yet they share a critical assumption: that competent human overseers remain present and capable of exercising meaningful oversight and control over AI decisions. Accountability mu...

A lot of tools can tell you an AI agent is behaving strangely. Very few can act on that at the network layer, without waiting on a human to approve a ticket. That’s the piece we’re closing.

IMDA and CSA are therefore studying how to strengthen the reporting of serious AI-related incidents, including whether additional reporting or information-sharing arrangements are needed. We will draw on international approaches and build on existing...

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...

Yea this is the most important result of the year, at the level of significance of “Attention Is All You Need”. It brings us from “is alignment even solvable” to “we may solve alignment in the next year.”

I used to always think that, you know, alignment will just fall out. I no longer believe that. So I think this alignment thing may be a harder, different problem.

As I see it, solving alignment for LLMs isn’t just impossible, it’s incoherent: we can’t even articulate what it would look like to succeed.

If Anthropic was just one voice among many, this would not matter. But given its market-leading position and the risk that the industry develops as an oligopoly, society as a whole may wish to have some say over what is good for it. There are ways to...

If consciousness started to emerge in training, I would call for everyone to stop AI development until we could design out those properties.

Governments must assume their responsibilities. It is the first responsibility of any government to protect its citizens – including from the threats of artificial intelligence. National action is essential. But global coordination is also indispensa...

It's actually impossible to control it unless we think the entire world is going to sing kumbaya and say let's slow down, which as you said is unlikely. So if that's the case, fine. You put slowdown here maybe. Well, okay. Does it mean Japan slows do...

Turns out, we were being unnecessarily pessimistic. Even though the planet is on fire, extreme temperatures will not be the ones to knock us out of the evolutionary game. (So much for our soggy paper straws.) Instead, we’ve decided to build something...

To get from today’s AI to human extinction requires a long chain of assumptions. [...] Luckily, none of the steps on this list have come to pass yet. [...] However, something being difficult to disprove does not make it likely. [...] Possibility is b...

We do need to have some safeguards in place to make sure there are no unintended consequences, but, no, I do not think AI is going to imperil human existence.

Some people worry that giving AIs separate legal personhood would let AI companies, or bad actors using AIs, shield themselves from responsibility. But it doesn’t have to work that way… when a Walmart delivery driver runs over a pedestrian, we want t...

There is a reasonable chance, however, that if the vision for AI agents comes to pass, then for our markets and legal accountability regimes to function effectively, agents will have to have the capacity to sue and be sued directly. This is the defin...

Machines are tools, not persons. Today, some propose granting these systems rights, or even legal personhood. Such an approach risks obscuring human responsibility and leaving those who suffer harm to confront a faceless entity.

When an OpenAI engineer sounded the alarm on the rapid pace of AI development, warning that this technology has the potential to have cataclysmic consequences, it wasn’t just deeply alarming—but a call to action. Critically, the 2026 International AI...

The FRONTIER Act is a targeted bill with clear, uniform transparency, and safety standards, instead of a fifty-state patchwork. I'm glad to support this bipartisan effort so that Congress acts before this technology outruns us.

The United States has a clear interest in helping its partners and allies navigate this choice strategically, and there are several steps it should take. Most importantly, Washington should develop frameworks to assure more reliable access to the U.S...

I wish to speak in support of Amendment 41, in the name of the noble Baroness, Lady Harding, to which I have added my name. We heard during Committee not only about the out-of-control behaviour from agents running AI models from OpenAI, Google and Me...

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...

To maximize the effectiveness of such provisions, policymakers should define a sensible range of time frames to report incidents, possibly with a graduated approach according to the gravity of the harm.

Open models have a tendency to be less safe a lot of the time than closed ones. Maybe it could be misused by someone who downloads it and modifies it to remove safeguards or to sharpen its harmful capabilities. This type of fine-tuning could cost hun...

Restricting access to a dual-use AI model is precautionary only if it delays harmful actors more than defenders. That condition varies across actors: a state agency or organized criminal group may obtain a substitute through theft, distillation, inte...

Open-weight models have been a boon for defenders. When Hugging Face investigated how OpenAI-powered agents penetrated its systems, proprietary models refused some of the forensic work, but the Chinese open-weight model GLM-5.2 helped reconstruct the...

External evaluators should be embedded at AI companies, but they should not be restricted to AI safety researchers, and they should not be selected by the company being evaluated. Rather, they should represent communities across society, address a mu...

the bad news: Dario estimates that in 6 months AI will have the ability to take over the internet the worse news: Anthropic's current commitment to stopping this is... putting independent evaluators in desks in their office have a great day

They really become part of the company, which makes it difficult to criticize it because the staffers become almost your friends.

A Global Call for AI Red Lines, launched at the UN General Assembly in September 2025, calls for categorical prohibitions on what an AI system may be built or used to do and presses governments to conclude a binding international agreement.

AI isn’t going away. Barring unprecedented global government coordination, it isn’t even slowing down. The only real decision is whether to centralize power, or disperse it. Do you fear the technology? Or a small number of people controlling the te...

The United States totally rejects any attempt to construct a globalist scheme of control for the artificial intelligence being spoken of so much now – hereinafter officially called super intelligence.

Which is why US lawmakers put forward legislation, dubbed the ‘Kill Switch Act’, which would require companies to have a way to shut down problematic AI tools. Tragically, the proposal was quickly shot down.

If you’re relying on a ‘kill switch’ to stop an AI agent, you’ve already made a fundamental mistake. A properly governed agent is simply never handed the task, so there is nothing to kill.

Embedded evaluators are independent external experts placed within frontier AI companies to provide ongoing assessments and validation of a company’s safety, security, and operational systems, practices, and claims. [...] Contracts should contain spe...

Connecticut already has strong whistleblower laws that protect employees from retaliation. [...] creating a separate, AI-specific whistleblower regime could lead to confusion, duplication, and unintended consequences for employers operating in highly...