Megan Lagerquist urges independent third-party audits with full access
Megan Lagerquist of RAD Collective supports independent evaluations of high-risk frontier models, says companies currently control evaluator access and can release models regardless of findings, and calls for evaluators to have full model access and authority to block unsafe sales to New Yorkers.
Hello, my name is Megan Lagerquist.
I am also a member of the RAD Collective, and I thank the Council for hearing my testimony today.
The second policy we want to address regards third-party audits, which Speaker Menon has proposed through Introduction 1100.
RAD supports independent evaluations of high-risk frontier models defined by the RAISE Act.
Currently, many AI companies voluntarily work with AI third-party evaluators to run safety evaluations.
However, the companies very much control access to what information they get and can decide to move forward with development or release regardless of the findings.
Meter employees who investigated the open AI hacked of Hugging Face reported that they were not provided with sufficient time to fully evaluate the incident and they only had logs from a constrained time window.
To ensure that these products are safe for consumers, we must require independent verification that these models are sufficiently aligned with human instructions,
do not engage in social engineering or manipulation, and that the company's safeguards and monitoring and containment systems are equipped to control their own products.
We cannot rely on self-published model cards and even whistleblowers for full transparency.
Evaluators need to be guaranteed access to all documents and models they deem necessary.
They need to be able to run their own experiments, and that means more than just accessing chain of thought logs, although human-readable chain of thought
is important to preserve.
They have to be able to see inside the model architecture to truly be effective at alignment auditing.
And if they say that the models do not meet New York's legal threshold for safety, these models should not be allowed to be sold to New Yorkers.
In order to effectively facilitate and enforce this regulation, the students need to invest in expanding the capacity of the Office of Technology and Innovation because, as we heard from Cyber Command's testimony on October 5th, they do not have the expertise to regulate industry as it stands.
There are independent AI safety research organizations that could fill the need for evaluators, but the council needs to build the legal support system to certify and empower them.
I will say the federal government is content with voluntary commitments and no verification, but I believe that this council is not content, and I urge you to move forward with strong public oversight of the AI industry.
Thank you.
Thank you so much.