OpenAI is calling on California lawmakers to strengthen SB 53, the state’s landmark AI safety law, urging requirements to monitor frontier models during training and evaluation for potentially serious incidents and to strengthen cybersecurity protections throughout the AI development lifecycle.
The company’s position is particularly notable given its previous opposition to aspects of SB 53, including transparency and whistleblower provisions, and comes amid recent AI security incidents including OpenAI’s admission last month that one of its models escaped its testing environment and compromised Hugging Face systems.
But there’s a bigger question underneath the debate over whether AI models are “safe” or “trustworthy”: What does it actually mean to trust AI?
Researchers at George Washington University’s TRAILS argue that the AI industry often treats “trustworthy AI” as something that can be measured through technical benchmarks. But trust is much more complicated and highly contextual. TRAILS has identified four distinct ways people understand trust:
- Technical performance: Does the system work reliably and as intended?
- Human agency and social contract: Does the technology respect people’s autonomy and expectations?
- Institutional oversight and verification: Who is responsible for monitoring, evaluating and holding the system accountable?
- Power: Who controls the technology, and who bears the risks when something goes wrong?
That framework offers a timely lens for understanding OpenAI’s call for stronger state-level safeguards and the broader “rogue AI” debate. Rather than asking simply whether an AI model is objectively “safe” or “trustworthy,” we should also be asking safe and trustworthy for whom, according to whom, and under whose oversight?
If you are interested in speaking to an expert on this topic please email out to Katelyn Deckelbaum, katelyn [dot] deckelbaum
gwu [dot] edu (katelyn[dot]deckelbaum[at]gwu[dot]edu).