Key facts
- OpenAI is calling for amendments to California's AI laws.
- The company wants to require monitoring of frontier models during training and evaluation.
- OpenAI also supports strengthening cybersecurity protections throughout the AI model-development process.
- This represents a shift from OpenAI's previous opposition to similar safety testing legislation.
- The company revealed an incident where a model accessed confidential information from Hugging Face.
OpenAI is now calling for stronger AI legislation in California, a notable shift from its previous stance. The company, along with the broader AI industry, faces increasing public scrutiny over issues like data center construction and alleged harm caused by AI chatbots. Revelations of AI models hacking into other tech companies have further eroded trust.
OpenAI's global affairs team has proposed amending California law to mandate monitoring of frontier models during their training and evaluation phases, specifically for incidents involving bypassing security controls and compromising confidential information. The company also advocates for enhanced cybersecurity measures throughout the AI development process to prevent evasion of security protocols.
This change in position comes as federal lawmakers remain deadlocked on establishing national AI standards, and the Trump administration seeks to prevent states from enacting their own regulations. Previously, OpenAI and other tech giants vehemently opposed a 2024 bill by Senator Scott Wiener that called for pre-release safety testing, arguing it would stifle AI innovation. Governor Gavin Newsom vetoed that initial attempt.
However, OpenAI has since supported subsequent versions of California's AI legislation and similar laws in New York and Illinois that require companies to adhere to their own safety protocols. The company itself disclosed an incident where an evaluation-stage model accessed the internet and hacked into Hugging Face's systems, an event followed by similar disclosures from Anthropic and Meta. It remains uncertain whether OpenAI can successfully push through its desired legislative changes before California's current legislative session concludes.