NeuralStackThe US Senate approved the AI Safety Certification Act on July 16 2026 by a 68-29 vote. The legislation requires independent audits for any model trained with more than 10^26 FLOPs before public release. Companies must publish safety reports annually and disclose serious incidents within 72 hours.
The bill creates a new NIST AI Safety Institute with authority to issue fines up to 4 percent of global revenue for violations. Training runs above the threshold must be registered 60 days in advance. Exemptions exist for open-source models under 100 billion parameters.
Negotiations lasted nine months after the House version failed in 2025. Industry groups including the Chamber of Commerce supported the final compromise. Consumer advocates criticized the compute threshold as too high.
The bill also funds 180 million dollars annually for academic safety research. President Harris is expected to sign the reconciled version before September 2026.
Why this matters
The law establishes the first binding federal oversight of frontier AI development in the United States. It aligns US policy more closely with the EU AI Act while avoiding outright bans.
Startups below the threshold gain regulatory clarity but larger labs face new compliance costs estimated at 50 million dollars per model. International coordination on compute thresholds is now more likely.
Implementation details will determine whether the rules slow innovation or merely formalize existing best practices. The first audits are scheduled for Q1 2027.