Don’t let California turn universities into AI gatekeepers
The state can ensure responsible GenAI use in higher education without undermining its educational and research potential.
Universities exist to advance knowledge through cutting-edge research and by training the next generation of thought leaders. Artificial intelligence is a growing part of that work. Yet a bill moving quickly through the California Assembly threatens to place political constraints on its reach as a tool for research and exploration.
Assembly Bill 2392 would require the California State University and the Chancellor of the California Community Colleges to convene a working group and establish procurement standards for generative AI — or GenAI — in line with certain state requirements. These requirements include ensuring that GenAI tools are evaluated for “potential harms, misuses, abuses, and bias,” and ensuring they will not produce “harmful” or “illegal” content. Although the bill provides a few examples of covered content, such as “disordered eating,” they don’t cure the underlying issue: these terms have no settled meaning and often depend on contested political and social judgments, making this mandate as broad as it is subjective.
And while universities have a legitimate interest in prohibiting and responding to unlawful uses of their resources, vague anti-bias and anti-harm restrictions could prompt institutions to procure watered-down GenAI that avoids controversial ideas.
Imagine a professor using GenAI to identify statistical correlations between crime and mental illness — research that might raise uncomfortable conclusions about institutionalization, homelessness, and public safety. She won’t get very far with GenAI that finds support for institutionalization stigmatizing and avoids “harmful” outputs. This tool might avoid such sensitive correlations or qualify answers so heavily that analysis becomes difficult. These system reluctances could permeate other delicate subjects like extremism, poverty, sex and gender, and race.
This is particularly unacceptable at universities, which are supposed to encourage students and faculty to confront difficult truths about our society and ourselves. In this way, members of the campus community educate themselves and push the boundaries of human knowledge. But that process short circuits if the tools provided by universities prevent unvarnished truth seeking.
These concerns extend beyond research. A philosophy student can explore an unpopular worldview by engaging GenAI in a back-and-forth debate. Such exercises are not only common in higher education, but indispensable to rigorous inquiry. It’s one way to avoid what John Stuart Mill called the “dead dogma” problem, where people continue to profess a belief without understanding or being able to defend the reasons for it. GenAI systems are ready-made devil’s advocates — available on demand to every student. But they can’t fill this role if they are trained to avoid the arguments students need to confront.
To be clear, nothing in this bill directly prevents people from using non-procured GenAI. However, university policies may only allow school-approved GenAI systems in various contexts, including when working with student records, unpublished research, sensitive data, or information subject to confidentiality obligations. If AB 2392’s standards cause schools to procure heavily limited tools, many people on campus will be left with AI systems incapable of facilitating intellectual pursuits.
California can ensure responsible GenAI use in higher education without undermining its educational and research potential through restrictions that cause it to deny, obscure, and avoid sensitive issues or “biased” answers. Lawmakers, for example, can order public institutions to provide training on the lawful use of GenAI or how to use AI effectively. These steps do not require restrictions that may limit a GenAI system’s ability to explore controversial or sensitive subjects.
Higher education should be empowering today’s experts and tomorrow’s leaders to break new ground across a wide array of subjects. That’s hard to sustain when state rules push campus tools to dodge difficult, unpopular, or outright offensive topics.
 | A guest post by
|
 | A guest post by
|
As a professor at a UC, this bill horrifies me. The bill will create many potential filters upon AI models based on the personal and subjective criteria held by members of the working group, rather than allowing faculty researchers and students the ability to make informed choices on their own. The bill would create yet another bureaucratic obstacle that would intrude into the inner workings of research labs. The bill dictates that AI models should be assessed and evaluated for their "bias", but does not define what is meant by bias, nor how it should be measured. This is a trojan horse. The bill does not explain what kinds of bias will be evaluated (political, ideological, demographic, statistical, etc.), how bias should be measured, what level of bias would be considered unacceptable, who determines whether a model is biased, or what mitigation techniques would be appropriate, if any. This "working group" will have its own extraordinary power to define these things on the fly, suited to their own personal definitions of bias; this is an extraordinary power that in a rational democratic society, they should not have. There is no consensus definition of bias, in any case, that will align with the diverse uses cases to which these models are put in research settings. After all, the many different kinds of biases that exist in these models may itself be a subject of scientific research. Faculty and students need to be able to engage with a diversity of AI models. It is the responsibility of faculty and students to not use these models for illegal ends. That is where the moral and legal responsibility should remain-- with the adult users.