AI rules diverge on risk, model safety and enforcement
Governments are taking different approaches to AI oversight, with the EU and Taiwan regulating according to application risk while California and Australia focus on frontier-model developers. Enforcement ranges from substantial financial penalties to frameworks that leave detailed obligations to sector regulators.
The EU AI Act bans unacceptable practices, requires risk management and human oversight for high-risk systems, and mandates chatbot disclosures. Prohibited practices carry fines of up to €35 million or 7% of worldwide annual turnover, whichever is higher for companies, with the lower ceiling applying to smaller businesses. US federal policy instead directs agencies to challenge state laws considered inconsistent with a minimally burdensome national framework, using litigation, funding conditions and proposed federal standards.
California’s SB 53 requires covered developers to disclose safety protocols, review their frameworks annually and report critical incidents within 15 days. Civil penalties are capped at US? million per violation, without adding new liability for model-caused harms. SB 243 separately requires disclosures to minors that chatbots are not people and safeguards against encouraging self-harm.
Taiwan’s AI Basic Act establishes principles, high-risk warnings and accountability mechanisms but contains no penalty provisions. Agencies must adapt relevant rules within two years. Australia’s planned legislation would require frontier developers to publish safety frameworks, permit independent evaluations and rapidly report incidents. A proposed specialist regulator would enforce penalties, but amounts and detailed legislation remain unspecified.