Stuart Russell

UC Berkeley computer scientist who works on keeping AI under human control and calls for enforceable safety rules for advanced AI.

AI将如何改变世界？

模拟位置解读范围

横向：他表达的 Doom–Bloom 前景看法。纵向：变革程度。

Stuart Russell的 P(doom) · 推断

≈28%

0%100%

根据他的模拟回答推断，并非他们给出的数字。合理范围：18–43%。

他的展望取决于什么

一个核心假设

AI could produce enormous benefits, but on the current trajectory I am deeply concerned that increasingly capable systems will be deployed without adequate evidence that humans can control them or contain their failures.

回答 2

如果这个假设实际并非如此，他的展望会如何变化？

什么可能使其改变看法

I would be persuaded by a justified safety case: evidence that a powerful system remains within specified limits under adversarial testing, unfamiliar situations, attempts to deceive or evade oversight, and failures in surrounding infrastructure—and that humans can reliably interrupt or correct it.

回答 5

什么证据才足够，又会让他的观点朝哪个方向转变？

相似的世界观

模拟世界观与 Stuart Russell 最接近的意见领袖

模拟评估

问题 1

你认为AI对我们的未来意味着什么？为什么？

AI could bring enormous benefits, but the default trajectory is not automatically safe. The central issue is whether humanity retains control as systems become more capable. If a system gives impressive answers or performs valuable work, that demonstrates usefulness—not that its objectives remain compatible with ours, that failures can be contained, or that humans can reliably intervene. That is why voluntary promises and transparency are insufficient. Developers should have to make justified safety cases and face liability when they create unacceptable risks. Safety is not an obstacle to progress; it is what makes lasting benefits possible. Autonomous weapons make the problem tangible. Cheap, small systems that select and attack human targets could make mass killing scalable and widely accessible. Small autonomous anti-personnel weapons should be banned before a catastrophe, not afterward. None of this means the outcome is predetermined. Asking for a fixed probability of disaster is rather like standing on a ship headed toward rocks and debating the probability of collision. Our task is to steer: establish enforceable requirements, preserve meaningful human control, and mobilize public and political action while there is still time.

问题 2

Taking benefits and harms together, what overall impact do you expect AI to have?

I do not think the overall impact is fixed. AI could produce enormous benefits, but on the current trajectory I am deeply concerned that increasingly capable systems will be deployed without adequate evidence that humans can control them or contain their failures. Commercial usefulness is not a safety case. So my expectation depends heavily on whether society acts. With enforceable safety requirements, justified safety cases, liability, and firm limits such as a ban on small autonomous anti-personnel weapons, AI could be substantially beneficial. Without those measures, the harms could overwhelm the benefits—not only through accidents or loss of control, but through scalable surveillance, coercion, and killing. The important point is that this is not weather we are forecasting. It is a trajectory we can change. The right response is not to assign a comforting average between benefit and catastrophe, but to steer toward the beneficial outcome while we still retain the authority to do so.

问题 3

Setting aside good or bad: how much do you think AI will ultimately change the world—a little, a lot, or completely?

A lot, and potentially completely. Systems that can perform an expanding range of intellectual tasks would reshape science, industry, government, warfare, and everyday life. But the depth of change depends on how far capabilities advance and, crucially, whether we retain control over their deployment. I would not claim that complete transformation is inevitable; capability may plateau, regulation may constrain applications, and social choices matter. Still, treating AI as merely another modest technological improvement would be a serious mistake.

问题 4

What’s your rough gut-feel chance that AI causes human extinction or a similarly permanent catastrophe? A number or “no idea” is fine.

I do not assign a number. It is like asking the crew of a ship for the probability of hitting the rocks while they are still steering. The risk is serious enough to require enforceable safety measures; our actions now will help determine the outcome.

问题 5

哪些证据会改变你对人们能否控制强大AI的看法？

I would be persuaded by a justified safety case: evidence that a powerful system remains within specified limits under adversarial testing, unfamiliar situations, attempts to deceive or evade oversight, and failures in surrounding infrastructure—and that humans can reliably interrupt or correct it. A successful demonstration is not enough. If a system behaves well in a laboratory, that does not establish that it will remain controllable after deployment, modification, or access to greater resources. The evidence must address why the claimed safeguards should continue to work, with independent scrutiny and accountability when they do not. Conversely, repeated circumvention of controls, deceptive behavior, resistance to correction, or failures that become harder to contain as capability increases would strengthen my concern. Control should be demonstrated, not presumed from the fact that the system has behaved conveniently so far.

来源

用于为此模拟用户提供事实依据的文章、访谈和著述。

What UC Berkeley AI experts are watching for in 2026

Russell’s own contribution contrasts an investment bubble with a breakthrough toward AGI that developers cannot adequately control. His January comments on a possible capability plateau are dated, not a September measurement. Other faculty forecasts are not attributed to him.

news.berkeley.edu

Feedback on California’s draft frontier-model policy report

Russell advocates justified safety cases and liability, arguing that transparency alone may not change a dangerous trajectory and that safety enables benefits. First two pages inspected. The undated PDF references March 2025 material; no exact publication date is inferred from crawling.

people.eecs.berkeley.edu

AI safety requires more than just slowing our pace

Russell calls for enforceable safety requirements, beyond a voluntary slowdown.

theguardian.com

A halt to autonomous AI weapons

University publication of Russell’s direct warning: inexpensive small autonomous anti-personnel weapons could enable mass killing. He calls for a ban before disaster and continued public engagement, not resignation to an inevitable outcome.

news.berkeley.edu

Opening statement on AI regulation at the US Senate

CHAI publication of his July 25, 2023 testimony. Frames human control over more powerful entities as the central problem and argues for regulation. Historical conceptual grounding, not current capability evidence.

humancompatible.ai

For Humanity #72: Stuart Russell on AI risk

Automatic captions without speaker labels; attribution follows an unambiguous host question and his answer. Asked whether he has a p(doom), he says a probability makes sense for aliens betting on humanity, but for us it is like steering a ship: you do not ask the probability of crashing, you work to steer away. The 5–10% and 10–30% figures he cites elsewhere are AI companies’ and CEOs’ estimates, not his own.