Significant developments in government’s approach to artificial intelligence are the latest topic at Hyperdimensional, the online newsletter about emerging technology published by Dean Ball, a board member of The Alexander Hamilton Institute for the Study of Western Civilization (AHI). His latest post, “What Good AI Policy Looks Like,” highlights governmental actions that “move the ball forward in a positive manner.”

California’s extensive proposed AI law, SB 1047, is encountering substantial “public pushback”—a good thing in his view, since the “simplistic” bill would generate “needless complexity.” In effect outlawing any AI model “that could possibly be used to do very bad things,” it would be comparable to outlawing the manufacture of a computer that could be used for “very bad things.”

“Criticizing bad legislation is an important part of policy scholarship,” notes Mr. Ball, who recently joined George Mason University’s Mercatus Center as a research fellow focused on AI, but “I prefer building to tearing down.”

Mr. Ball praises Connecticut’s SB 2 as “some of the best AI legislation” in the world. A few key provisions: Developers would have to publicize such basic information as their models’ intended uses and their architecture. AI models would have to be labeled as AI, not “pretend to be humans or let their identity remain ambiguous.” If an AI developer complies with certain international standards, the state would presume its own concerns are sufficiently addressed.

The bill’s definition of “high-risk” artificial intelligence may be seriously problematic, however. Any AI model can be used for a high-risk purpose, and its developers would have to “perform algorithmic impact assessments, develop risk management plans, and take several other steps.” An example of potential hassles for small business:

“Imagine you’re [a heating and air conditioning, or HVAC] contractor who would like to make an AI language model to help draft bids to send to clients. You might fine-tune this model with examples of previous bids … might also ‘train’ the model by feeding it additional data beyond what its developer used … If fine-tuning is considered a ‘substantial modification’ … our HVAC contractor may well have to comply with technical legal requirements simply for using a product as intended.”

But although the Connecticut bill “could create a regulatory burden on a far greater range of corporations than its authors likely had in mind,” sufficient clarification in its language could prevent that. Overall, it shows “a thoughtful approach to AI regulation, balancing the need for safety, transparency … innovation and adoption.”

Mr. Ball also praises a noteworthy initiative in Ohio, which in 2020 began reviewing its lengthy administrative code to identify what is redundant or outdated in the text. Using an AI system, the state has “cut 2.2 million words of unnecessary regulation.”