AI GUIDEPartner content

How to Manage AI Risks: Three Approaches by Russell, Bender, and LeCun

Russell, Bender, and LeCun agree that AI needs to be controlled, but they assess priorities differently: future loss of control, harm already being observed, and dangerous use cases. We examine how their approaches to safety assurances, open models, and regulation differ.

Affiliate link: your price stays the same and the project earns a commission.

AI risk management includes measures that limit potential harm and help assess the use of artificial intelligence. Russell, Bender, and LeCun, as the review's author describes their views, agree that control is necessary but prioritize different concerns: from future loss of control over powerful agents to harm already being observed and the danger of monopolization.

Two Scales of AI Risk: Harm Today and Loss of Control in the Future

The debate over AI safety is not only about how intelligent a model may become. Participants discuss different sources of risk: the consequences of a specific use of a system that is already available, and the possible behavior of future autonomous agents that are assigned actions, not just answers.

In the author's account, Emily Bender focuses first on existing harms: discrimination, disinformation, exploitation of people, and the erosion of accountability. This approach suggests starting the analysis with those harmed, the decision made by the system, and the people or organizations responsible for deploying it.

According to the author's account, Stuart Russell focuses on a different time horizon. He is concerned about the possibility of losing control over powerful future agents, and therefore calls for convincing safety assurances before they are deployed. This is a position about potentially severe consequences, not a claim that loss of control has already occurred.

The distinction between these scales matters when choosing measures: assessing a specific service addresses questions about its current behavior, while assessing an autonomous agent should also take into account what actions it is allowed to perform and how a person can stop it from carrying out a task.

One test cannot answer both questions.

What Safety Assurances Are Needed According to Russell's Approach

In the author's account, a safety assurance in Russell's view means having convincing grounds to consider the deployment of a powerful agent acceptable before the system begins operating in a real-world environment. This is not a ready-made universal standard or a numerical threshold: the materials provided do not specify a particular procedure that, by itself, would prove safety.

This position can be applied as a requirement for a launch decision: determine in advance the agent's authority, the potential harm from an error, ways to monitor its actions, and the conditions for stopping its operation. This is a practical interpretation of the argument for assurances, not an official list of Russell's requirements.

If an agent can modify data, send messages, or initiate operations, checking only the quality of its answers is not enough. Its actions should also be assessed under conditions involving a failure, a tool malfunction, and an incorrect initial assumption; a person should understand what consequences the system can cause without further confirmation.

A strength of this approach is its attention to consequences that are difficult to undo after launch. Its limitation is that requirements for evidence of future systems' safety are hard to reduce to a single criterion, especially when there is no generally accepted way to determine how close a system is to general or superhuman intelligence.

Why Bender Puts Harm Already Being Observed First

Bender's approach draws attention to the concrete consequences of using AI. In the author's review, these include discrimination, disinformation, exploitation of people, and a lack of accountability. The material does not provide separate statistics or a list of verified incidents, so these categories should be understood as the position reported by the author, not as findings from research presented here.

When assessing a system in a specific application, it is useful to ask who bears the cost of an error, who benefits, and who can challenge the result. For example, if an algorithm is involved in selecting job candidates, measuring accuracy alone is not enough: the consequences for the people affected should be assessed, and it should be established who is responsible for the decision.

What if a company calls its system a neutral tool? The label itself does not resolve the question of accountability. According to the author's account, Bender argues against pushing current harms into the background because of debates about a possible threat from hypothetical superintelligence.

This priority helps focus assessment on the actual conditions of use: what data goes into the system, who makes a decision based on it, and whether there is a way to correct an error. But this focus alone does not resolve the question of how to control more autonomous systems in the future.

LeCun's Approach: Open Models and Regulation of Applications

According to the author's account, Yann LeCun considers catastrophic predictions exaggerated, supports open research and testing, and proposes directing regulation toward dangerous uses of AI. He also points to the risk of concentrating technology and control over information, so in this position, open models and independent evaluation are tied to the distribution of access and the ability to conduct assessments.

It is important here to distinguish between regulating model development and rules governing its use. LeCun's approach, as the author describes it, emphasizes the specific use case: danger arises not only from the technology's capabilities but also from who is allowed to use the system, for what task, and with what consequences.

Openness by itself does not prove safety.

Access to a model can help researchers conduct independent assessments, but evaluation still needs to take into account the system's intended purpose, available tools, and deployment conditions. The material provided contains no data that would allow us to declare open models safer than closed ones in every situation.

This is where tension arises between independent evaluation and the risk of spreading capabilities that can be used to cause harm. The author reports LeCun's position in favor of open models, but does not provide a universal rule for determining when open access is justified and when the danger of use calls for restrictions.

How to Distinguish Safety Measures from a General Slowdown in Development

A safety measure limits a specific risk or evaluates a specific use case. A general slowdown in development restricts a broader range of research and products, including those for which the material identifies no dangerous area of application. The author describes public debate as oscillating between strict restraints on development and technological liberalism.

To avoid conflating these approaches, when discussing a restriction, specify what it applies to: a model, access to a tool, an autonomous action, or a specific application. Then explain what harm the measure is meant to prevent, whom it protects, and how its impact will be assessed. This is a practical way to analyze arguments, not an official AI governance framework.

  • Application assessment helps assess a specific risk, such as discrimination or harm from an incorrect decision.
  • Limiting an agent's authority reduces the number of actions the system can perform without a person.
  • Independent evaluation makes it possible to assess a system independently of its developer's self-assessment, although the method of evaluation depends on the application.
  • Restricting development affects a broader process and requires a separate justification: what risk it reduces and what research it affects.

These positions do not offer a single answer to the question of regulation. Russell calls for convincing assurances before powerful agents are deployed, Bender puts harm already being observed first, and LeCun supports open research and the regulation of dangerous uses. Their differences show why discussions about safety become more precise when participants identify a specific risk, time horizon, and responsible party.

Frequently Asked Questions About AI Risk Management

Do Russell, Bender, and LeCun think AI needs to be controlled?

According to the author's account, all three recognize the need to control the use of AI and assess its safety. They disagree about which risks to prioritize and what regulation should target.

How does the risk of losing control differ from harm already being observed?

The risk of losing control concerns the possible behavior of powerful future agents. Observed harm is associated with the consequences of using systems, such as discrimination or disinformation, which Bender focuses on in the author's account.

Does support for open models mean LeCun rejects regulation?

No. According to the author's account, LeCun supports open research and independent evaluation, and proposes directing regulation toward dangerous uses. This is not the same as rejecting control.

How can these approaches be applied when assessing an AI system?

First, specify the application and potential harm, then determine the system's authority and how its behavior will be assessed. For a launch decision, it is also important to establish who is responsible for the outcome and which actions require human confirmation.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

AI risk management artificial intelligence safety Stuart Russell Emily Bender Yann LeCun open AI models AI regulation

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 22 November 2026

Related reading

All in this section →