Managing AI Risks: An Economic View

How economists think about the risks of AI, from misuse and accidents to concentration of power, using tools like externalities and public goods.

AI brings large potential benefits and also real risks. Economists analyse these risks using familiar concepts.

Types of risk

Misuse: AI can be used for fraud, scams, deepfakes, cyberattacks or disinformation.
Accidents and errors: AI systems can make mistakes in high-stakes settings, such as medicine or finance.
Bias: AI trained on biased data can discriminate.
Economic disruption: rapid job changes in some occupations.
Concentration of power in a few companies or countries.
Longer-term risks: some researchers worry about highly capable future AI systems acting in ways their developers did not intend.

Externalities

Many AI risks are externalities: costs imposed on people outside the transaction between an AI developer and its customers. A company selling an AI tool may not bear the full cost if the tool is misused for fraud against third parties. Like pollution, such externalities can lead to underinvestment in safety without rules or incentives.

Safety as a public good

Research into making AI safe and reliable, such as testing methods and ways to understand how models work, benefits everyone, including competitors. Like other public goods, it may be underprovided by markets. Governments have responded by creating AI safety institutes, including in the United Kingdom and United States from 2023, to test models and fund research.

Racing dynamics

When companies or countries compete intensely to be first, they may cut corners on safety, a race to the bottom. Coordination, standards and agreements can help, similar to other areas where competition creates collective risks.

Liability

Making developers or deployers liable for certain harms gives them incentives to prevent them. Debates continue over who should be responsible when AI causes damage: the developer, the company using it, or the user.

The deepfake scam

Fraudsters use AI to clone the voice of a company executive and call an employee, instructing an urgent money transfer. The employee complies, and the company loses money. The AI developer earned revenue from the voice tool, but the loss fell on the company. Economists see this as a misuse externality, which rules on verification, liability and detection can help address.

Thinking AI risk is either nothing or catastrophe

AI risks range from everyday problems like scams and bias to more speculative long-term concerns. Economic tools help weigh different risks and design proportionate responses.

Key takeaways

AI risks include misuse, errors, bias, disruption, concentration and long-term concerns.
Many AI risks are externalities that fall on third parties.
AI safety research is a public good, which governments support through safety institutes.
Racing dynamics and liability rules shape incentives for safety.