All issues · Technology · No. 30

Artificial Intelligence Regulation

Can a fast-moving technology be governed without either freezing it or letting harms compound?

AI systems are being deployed across hiring, lending, medicine, media, and defense faster than legal frameworks have adapted. Debate covers what to regulate, at what layer, and who bears responsibility when systems cause harm.

Regulatory proposals target different layers: the underlying models, the applications built on them, the data used for training, or the outcomes in specific regulated sectors. These imply very different rules and enforcement.

Existing law already applies in many contexts. Discrimination, product liability, consumer protection, and sector-specific rules cover AI-mediated decisions, though how they apply is often unsettled.

Concerns range widely in kind and timescale: near-term issues including bias, fraud, synthetic media, and labor displacement; and longer-term arguments about advanced systems, which are taken seriously by some researchers and regarded as speculative by others.

The positionsFind your officials →

POSITION 1 / 3

Regulate proactively

Waiting for harm to accumulate before acting has failed with previous technologies.

Automated decisions in hiring, lending, and housing can reproduce discrimination at scale.
Synthetic media undermines the evidentiary value of recordings.
Safety testing before deployment is standard in other high-stakes industries.
Concentration among a few firms raises competition concerns independent of safety.

POSITION 2 / 3

Light-touch approach

Premature rules written for today's systems will entrench incumbents and slow beneficial applications.

Compliance costs favor large firms with legal departments over new entrants.
Existing law already reaches most identified harms.
Restrictive domestic rules shift development to other jurisdictions.
Benefits in medicine, science, and productivity could be substantial.

POSITION 3 / 3

Sector-specific rules

There is no single technology to regulate; the risk depends entirely on the application.

Medical, financial, and safety regulators have relevant expertise already.
A recommendation engine and a diagnostic system warrant different scrutiny.
Disclosure of automated decision-making enables recourse.
Liability rules can assign responsibility without prescribing technology.

Terms you will hearFind your officials →

Foundation model: A large general-purpose model adapted to many downstream applications.
Algorithmic accountability: Requirements to assess and disclose automated decision systems.
Synthetic media: Generated audio, image, or video depicting events that did not occur.
Compute threshold: Regulating models above a set training compute level, used in some proposals as a proxy for capability.

What people actually disagree aboutFind your officials →

Should regulation target the model, the application, or the outcome?
Who is liable when an AI system causes harm: developer, deployer, or user?
How should genuinely uncertain long-term risks factor into present policy?

Do something about itFind your officials →

Know where you stand? The people who actually vote on this are reachable, and a message in your own words carries more weight than a form letter.

Find your officials →

Was this page balanced?

We are not asking whether you agree with any position — only whether each one was stated fairly. That is the standard this site is built to meet.

Think a position here is stated unfairly? Tell us what an advocate for that view would say instead.