Trust, innovation central to Colorado’s new AI guidelines, says state data chief

Two weeks ago, Colorado released its Guide to Artificial Intelligence for state agencies. But Colorado’s chief data officer said the success of agencies using generative AI will hinge on innovation, trust and the quality of the state’s data governance program.

Colorado has been forward-leaning with AI regulations — in May, Gov. Jared Polis signed into law SB 24-205, a law offering consumer protections for AI systems, becoming one of the the first states to do so. But Colorado Chief Data Officer Amy Bhikha said in-house use of AI has been more measured.

The new AI guide includes Colorado’s statewide generative AI policy, which Bhikha said was written over the last six months. It features more-formal rules around development, deployment and use of AI systems, products, services, tools and content within state agencies.

Central to these guidelines and policies is balancing innovation with trust. Bhikha said the key way Colorado is carrying out that balancing act is through the state’s data governance program. Before building out the AI policy and use framework, she said the state had to re-imagine its relationship with data.

“In the past, you were looking at data governance truly from that data privacy lens, you know, making sure that you’re keeping your HIPAA data or your [personally identifiable information] data governed in a way. But now, it is stretching out to looking at how could it potentially be used in the bias components, pieces of that nature,” she said.

Colorado began building a data inventory so officials will know which data the state has and where. Bhikha said her office had begun that task before she was formally charged with overseeing the state AI program earlier this year. Bhikha said the inventory will give officials confidence that their data is trustworthy, giving them a starting point to look for possible bias.

“Anytime you’re doing anything with a gen AI model, you want to make sure you know exactly where this [data] came from, what its quality was, how it’s going to be able to be repeatable, and then, not only the quality of that data, is it complete or accurate, but also, are we making sure there’s no unintentional bias in that data,” Bhikha said.

Another way the state’s AI policy reduces the chances for bias is a requirement that all generative AI uses — including via vendors — pass a risk assessment conducted by the Colorado Office of Information Technology.

Vetting AI uses has enabled the state to provide safe spaces for innovation, she said.

“We have a pilot going on right now where we have 150 people working in like a walled space to say, ‘Okay, how are you going to use it? What are the parameters we need to stay within for those pieces?’ And then working on specific use-cases.”

Bhikha added that the pilot also allows the technology office to evaluate the needs for AI literacy within state agencies and plan education programs.

“We want to be on that balance of being really innovative, but then also making sure that we have everything in place, and that it’s coming from a place of trust, and so really building in all the pieces to mature that from a state perspective,” Bhikha said.

She said collaboration with other states and their chief data officers has been essential for developing new guidelines and policies. Bhikha said the work has also uncovered some unexpected findings.

“I know, from my perspective, I’ve been surprised how much procurement has been a factor in this. Are the terms and conditions of these packages and software that we’re looking at, even something that we can commit to?” she said.

Bhikha said Colorado may not have a chief AI officer or a dedicated AI team, but “we’re finding out right away, it’s all the same.”

“When it comes down to it, it is really almost the old school data governance things that become the most fundamental,” she said. “I think people keep thinking this forward tech, but I find talking with other states and other programs, ones that who have led farther in that space, getting out there, have succeeded or not succeeded because of their data governance and quality programs.”