SmartTranscript of House Energy and Digital Infrastructure 2025-05-06 2:30pm

[Mark Holmes]: Alright. We are live on YouTube. [Chair Kathleen James]: Alright. Welcome back, everybody, to, House Energy and Digital Infrastructure. It's May sixth. And we are here with, some folks from ADS and the Vermont AI Council. We will go around the room and introduce ourselves, including our folks on the screen. And, then we always ask our guests to introduce themselves and then we'll turn it over to you guys for the record. So, I'm representative Kathleen James from the Bennington four district. [Mark Holmes]: Richard Bailey from Memorial two. Chris Morrow, Windham, Windsor Bennington. Michael Southworth, Caledonia two. Christopher Allen, Rotman four. Garrett, Torrey, Washington two. [Bram Kleppner]: Brent Klubner, Chittenden thirteen, Burlington. [Chair Kathleen James]: Laura Smith, one one two. [Vice Chair R. Scott Campbell]: Jack Campbell, representative from St. Johnsbury. [Mark Holmes]: Jack Laucher, community assistant. John Kelly, chief of staff aide. Yes. I'm Justice Collin from law supreme court and on the, from law supreme court's commission on AI use in the courts and in the legal profession. Super. [Chair Kathleen James]: Alright. For the record [Mark Holmes]: For the record, I'm Mark Holmes. I'm the chief technology officer of the agency of digital services as well as the cochair for the Vermont AI Council. Great. [Desiree Shanks]: And I'm Desiree Shanks, chief data and AI officer. [Chair Kathleen James]: Thank you for coming back. [Mark Holmes]: Thank you. [Chair Kathleen James]: Alright. Over to you guys, do you have, I know you shared some testimony. Do you wanna, share your screen, or do you wanna just walk us through it? [Mark Holmes]: I we'll just walk through it because I I think it's really more to provide, you know, opportunities for you guys to ask questions and, you know, keep us in the hot seat preferably. So as I said, I'm the cochair of the Vermont AI Council. We were formed in twenty twenty two coming out of the legislative session based upon some work that I'd also been a member of the group for, which was a a group which was studying the needs, you know, needs, opportunities, risks, around AI and and what the state could do to either a combat or whatever and how we would respond. So one of the recommendations of that task force was to this creation of AI council, which provides sort of a more constant set of oversight for AI activities within the state as well as become sort of a point for education and outreach within the state as well. So we have membership as defined in the statute from within the state that from various folks such as public service, serial services, the attorney general's office, then we have some outside council members as well who represent education and ACLU and other public interest groups. So that's that is the history of the council. Anything there you want me to touch on further? We can go on to the next one. I don't I don't [Chair Kathleen James]: think so. We just had a review of some of this background with ledge councils. Oh, perfect. Okay. [Mark Holmes]: Great. Then wherever you see need for me to kinda speed ahead, just give me a wink or something. Okay. Yeah. I'll be just So as I said, we were part of we were we're we're enacted through act one thirty two in twenty twenty two. We oversee the state's use of AI to ensure ethical use as well as provide public engagement on our use. Also within that was the requirements for us to meet once a month and provide an annual report, which includes recommendations that are taken from the various agendas and testimonies and such that we gather throughout the year as well as working with my compadre here, who is the chief data and AI officer, who is actually more the operational day to day within AI within the state. So we intersect with him to look at the the state's current inventory and report on some of the progress that the state is making internally. So that annual report, I think, is what you are is the first part of this. So it's it's the second of our I think it's our second annual report as a council. It was this year. Third third. The card benefits on the report. Okay. Yeah. Okay. So we had some recommendations within that, which I think are probably, like, the source on most of your, you know, questions with us. Through [Desiree Shanks]: meeting for [Mark Holmes]: the last year, we had a lot of topics that we wanted to dig further on and, you know, really get more meat to it too. As far as where we sit as a state, I'll say upfront that thanks to the the speed of execution and the abilities of Josiah and his team, I think Vermont sits out in front of a lot of other states, I'll be honest. We do we were one of the first to have a code of ethics for AI use within the state. We currently today have published guidelines for the use of AI, and this these were all reviewed and, you know, voted on and as part of the council as well. So they were things that Josiah worked on and brought through the council and add that. Yes. [Chair Kathleen James]: I I do have a question about that. I I didn't wanna interrupt you. No. You can, please. So I'm I'm curious about the code of ethics and the implementation guidelines in particular. How are those rolled out to state employees? That's a that's a lot of folks, and I assume that it's more than just sending everybody an email saying see attached. So how do you how do you roll that out? How do you train people? How do you know that folks are taking it seriously and abiding by it? [Mark Holmes]: Yeah. I'll let him jump in and give you the specifics, but I'll say that it it has required a lot of dedication to internal outreach, lunch and learns, a lot of time spent with people getting, you know, getting the information out there really kinda almost to a one on one basis. [Desiree Shanks]: Yeah. So me and a few other folks on my team have have held a series of trainings, probably eight or nine at this point, on AI related topics. And those have varied from, like, trainings for supervisors and managers, which focus more on the code of ethics and the ethical guidelines and ensuring compliance to, like, stuff for employees, which is more of, like, how do you use a tool like Chavity that I'll talk about in a little bit? How do you use that kind of tool effectively? And as part of that, go talk to your supervisor because that's what we've said in our guidelines is you need to clear things with your supervisor. And, and I'm kind of constantly, usually several times a month, engaging with various units around the state that are like, we wanna do this thing with AI. And then I have conversation with them and make sure that they're they're in alignment and come back to the AI council. If it's like, hey. Here's a use case we haven't really thought about. How do you all think we should approach this? And I'll get their input on those things. But we kind of try to do a continuous set of trainings and then use the ADS oversight of procurements to make sure that we are making sure that tools and physical objects that have AI in them aren't getting procured without eyes on them to make sure they're compliant. Representative Southwick? Do you have modules through the LMS? Not yet. We've done we try to do it as, like, one hour, like, a complete training. We've been talking about doing modules through the LMS, and when I bring on the new AI director, which will be any day, that's one of his first activities. We probably must get it. [Chair Kathleen James]: What's the LML? [Mark Holmes]: Learning management system. Learning management system. So it states training. It's an online training tool. [Chair Kathleen James]: Okay. So [Mark Holmes]: if you've done any online training, you go and click on a class. So we have one of those from the state, and we can upload numerous topics of, you know, coursework and such. Okay. [Chair Kathleen James]: I see you guys done it. Yeah. [Mark Holmes]: Which are not required. You'll get an email saying you have this and, yeah, you don't need too much time to complete it, that sort of thing. [Chair Kathleen James]: Okay. Grup Klapner, was that your question? [Bram Kleppner]: You mean to it. [Chair Kathleen James]: Okay. [Mark Holmes]: So the the the report itself, you know, talked about the successes that we've had, but also really, you know, to where we're looking ahead, right, where we're headed. AI is not a stagnant thing. It's not anything that anybody's got their hands around and said this is what it is. It's it's changing and growing day by day as we've sat here. It's it's morphed to something else. So that's one of the complexities that that the council is facing at the moment. It's just that the group of people that we brought in to originally talk were you know, it's great, but there's so many new technicals to AI that maybe, you know, broadening the group so we can get some different roles and different, you know, perspectives for some of these newer topics in there. There's been questions around, you know, intellectual property rights and and digital, you know, image rights and things like that. And we talk about them, but at at a, I would say, at a very high blush type level. And so we're interested in the potential to create subcommittees, which are actionable, that include folks who are not necessarily committee members on these subcommittees, to investigate some of these, you know, more in-depth topic areas. And then we're also, you know, requesting essentially to to broaden sort of the coop the council's membership to include a few other perspectives. [Desiree Shanks]: I I think one of the items we talked about was, like, when we first constituted the council. Yeah. Nobody really expected image generation in art was going to be, like, a really big deal. And it may not be a really big deal for forever, but for now, that's a topic that's on the minds of the art community in Vermont. And as it touches copyright, will be a a deal for much longer. And so the opportunity to bring in some subject matter expertise in that space would be helpful as an example. Education is another big one. AI is really transforming how education works. And, again, I think that wasn't necessarily something we expected right off the bat. You know, some code writing in that as well? Yeah. Yeah. Yep. [Mark Holmes]: And then as I think the nature of your first question, you know, around how do we get the word out, that public engagement's important too. So having different parties from different, you know, backgrounds or different segments of the Vermont picture really helps with that that public location as well. Vee [Chair Kathleen James]: Sorry. I can't remember. How's how's education represented on the council? I can't remember who the actual appointed representative is. Is it public ed [Mark Holmes]: We we have [Chair Kathleen James]: two twelve? [Mark Holmes]: No. We have two folks who represent universities within the state. Doctor Suessman, who's at Norwich, and he can, you know, speak to issues within the education community as well as within the cyber community. And then I so I'm not so first. Yeah. Joe Joe Neer, who he's in a UVM faculty as well as representing the ACLU at the time at the moment. So so [Chair Kathleen James]: So higher ed is not [Desiree Shanks]: their primary thing. Like, that's nobody's primary Right. Piece. [Chair Kathleen James]: Okay. [Desiree Shanks]: We have we feel like we have some okay connective tissue at the higher ed level, but but the k twelve is that's another gap to be identified. Yeah. [Chair Kathleen James]: What page are you on? [Mark Holmes]: I'm actually so I yeah. My apologies. Like yeah. [Chair Kathleen James]: And I'm still on the first page. [Mark Holmes]: Oh, you are. Okay. So we're we're on the anticipating AI challenges. Five. Page five. So Oh. Yep. [Chair Kathleen James]: K. [Mark Holmes]: So, you know, as we're talking about the the future AI interest, so there are emerging technologies. I don't know if you wanted a definition of agentic AI, but it's it's essentially at a high level. It's the AI that's you're using to perform a it's not like the chat gbt, which is a interactive you know, let me ask you a question. You give you a spot. This is a heads on ag, you know, some degree of independence as an as an agent, as a worker, and doing things. So we're, you know, monitoring developments in those areas because that degree of autonomy is different than what we've talked about in the past. If you're talking about the information you get back from chat GPT, you're you're reading it. You're evaluating it. But now you're talking about with the agenda k I. There's independent actions. Of course, there's supposedly learning and policy and everything else that guides those actions, but they're still independent actions. Actually, what on that one? Yeah. So [Desiree Shanks]: as a as a for example, one of the trends that we're gonna be talking about at the next AI council meeting is kind of consumer grade AI agents. So you may have seen videos of, give it your grocery list, and it goes to a site like Hanford to go, puts it all in your cart, and checks out for you. Mhmm. And, like, using AI agents to do that. So if it's shopping for your groceries, that's maybe low risk. I think one of the early videos, the person ended up with, like, twelve pounds of pasta. But, like, that's not that bad. But, like, if you do that to renew your driver's license Mhmm. All of a sudden, if the AI agent makes up an answer that is not accurate Right. Normally, we call that fraud if a person gives us inaccurate information. If they're kind of one level separated from the AI agent who's entering the information, does it count as fraud still? Like, there's some things to think about as we're moving into this AI agent world and when it's appropriate to use as, say, an assistive technology and when it's not. So we don't have answers on that yet. That's what the next council meeting will be for, is to talk through that as as one example of kind of those emerging tech challenges that we are addressing. [Mark Holmes]: And then another, leads into the area of, cyber risk, which, you know, at at a very high level, it's really got to the point where it's agent or AI against AI out there, meaning that we're successful in protecting the state's network, for example, because we employ AI to counter attacks that are AI driven, and that's becoming more and more the case. And so as, you know, the AI is on one side, the bad side gets smarter. The AI is on the good side, we get smarter. And so we we follow those a lot, and we have, you know, some degree of conversation about them, but we're not in the depth of those that we'd like to be able to monitor. Policy development. So I know doctor Stusman was invited here and wasn't able to attend, but one of the things he wanted me to make sure that we put forward is that he, along with Norwich University, have a four hour training program, which is AI for policymakers, and they would love to reengage with you and provide us an opportunity. But we try to track various AI policies or Josiah specifically spends a lot of time tracking what other states are doing in in policy areas of AI. You you know, working with John Kelly over here, we track a lot of what's going on within our statehouse and try to stay ahead of the curve there so we can come in and provide good testimony that, you know, gives you guys helpful answers. Thanks. [Chair Kathleen James]: We were going back and forth with him. Do you know if the is that an online course, or is that something we got to do in person? I think that's where the conversation bogged down. [Mark Holmes]: I'm not sure. I can find out, though. We we will find out for you. [Chair Kathleen James]: Or we can. It's Okay. I think if if I'm being sitting in my email somewhere, I [Bram Kleppner]: Okay. [Chair Kathleen James]: I might have written or intended to write and see whether that was something we could do in the waiting days of the session here, which should be great because we have time. Yeah. [Mark Holmes]: Apologize. Thank you for coming. No. Thank you for having me. [Chair Kathleen James]: Yeah. Appreciate you. We look forward to seeing the the judiciary report. I'll I'll have [Mark Holmes]: I'll have Terry send to the committee. Super. [Chair Kathleen James]: Yeah. Thanks, Deborah. You. Neat to see you here. That was crowded here. [Mark Holmes]: I had no idea. Yeah. So as a as a crossover went there, on our council is justice Dooley, who's a retired secretary of state supreme court justice. He was part of the original task force as well, and he's on our council as well. I do have to answer your question. It's in it's in person. It's a four hour [Chair Kathleen James]: Is it so do they operate here? [Mark Holmes]: Or They can be wherever you want. [Chair Kathleen James]: I think it's really cool. It's like, okay. Jack, make note. [Mark Holmes]: Cool. I'm gonna move onward to the next slide, which is slide six. Can we go back to four before we move on? Yeah. [Chair Kathleen James]: Sorry. I I didn't realize we were That's okay. Page four. So [Mark Holmes]: my fault. [Chair Kathleen James]: Oh, that's okay. So I I just wanted to go back to some of your key recommendations. I I wanna make sure that we're taking a minute to talk about that because that seems really important. So are these legislative recommendations things that you need us to do or things that you as a council want to do? [Mark Holmes]: So we're recommending back to the legislature that support to allow the council to do some of the things. So one of these would be, you know, clean up language maybe or changing language to extend our AI oversight into more beyond state government areas, into private sector use in some areas. [Desiree Shanks]: And Yeah. We've done a little bit. So we had a lot of discussion on the council about this because there are a few models that that other states are pursuing. Colorado has a model that's currently being litigated, I think. California has stood up, like, a a separate body basically to do AI regulations throughout throughout their state. So there's a few models of what that could look like. And at this point, as we highlighted in the report, because the AI division is technically one position, we don't really have capacity to do, like, a sweeping regulatory body like California has put in place. Right? Like, that's not feasible. And wouldn't even be feasible, like, if you said, go do it, and here's five positions or whatever. Like, that would not be nothing we could do next year because it's like, there's a lot to figure out between here and there. But it would be helpful to know from the committee's perspective, and we can talk about this at length, like, what do you all envision the future as being? We put forward kind of an intermediary in that that education role, our second bullet point there. We put forward kind of an intermediate phase where we could serve as a best practices clearing house for, Vermont businesses as like, hey. This is what we use within state government. And, like, Mark and I are both regularly at events around the state where we're engaging with business owners on on AI usage, or I recently did a panel on AI and mental health, talking with mental health practitioners about how to use AI well. So these are conversations that we're engaging in, but to start to take some of those best practices and and centralize them and and become a clearinghouse for that, we see as a good intermediate step of building our engagement, our ability to engage with Vermont businesses and would position us to develop the relationships that we could then if if the future was some sort of regulatory body in, let's say, last year was h seven ten and seven eleven. I don't remember what numbers it was from the session. But, anyway, there were some some bills establishing the AI Council as a maybe more regulatory entity. If that's the direction we're going, you would like to see that kind of phase over time so we can build those muscles. And it will require a lot of coordination with the attorney general's office because, obviously, they have enforcement of consumer protections and things like that. So, anyway, we're that most of that section was, we see a need potentially for some of this regulatory body in the future. We're not ready to say, like, let's go do that yet. We think the we don't need to be the first state to figure that part out. We'll let California and Colorado sort that out. But there's a step that we can take to kind of move us in a direction that would be of more service to Vermont businesses. [Chair Kathleen James]: Okay. Oh, rep Campbell. [Vice Chair R. Scott Campbell]: Sure. Thanks. Well, so it sounds like bullet point two is actually first. We you'd wanna you you would wanna set up some sort of role as a as a clearinghouse of information that you could provide to the private sector ahead of any sort of consideration of regulation. Is that is that what I'm understanding? [Desiree Shanks]: Yeah. I think the reason we ordered them this way was it would be helpful to know, like, do you see a regulatory body in the future? If so, then this education role is a good next step while we figure out what a regulatory body ought to look like. So they're, like but, yes, the functionally, the work my team would do would be the other order. It'd be great to have guidance from this body kind of in the order we've laid throughout here. [Vice Chair R. Scott Campbell]: Do you have let me guess you're one person, but do you have any legislative language or, you know, any any sort of more specific or concrete proposals for legislation? [Desiree Shanks]: We don't have any specific proposals yet. We we worked on that for quite a while in the council, and we're like, the models are not mature enough for us to be like, here's what we think would work in Vermont yet. We really wanna see where California's kind of consumer protection oriented language shakes out once it finishes through the the litigation process. That may be a good model. We'll see. [Vice Chair R. Scott Campbell]: Do you have a sense of a timeline for that? Is there anything that we should be thinking about for next year? [Desiree Shanks]: I expect we'll have more clarity on it next year. I I don't know how long that litigation will take, though. [Mark Holmes]: Okay. Okay. Thanks. [Chair Kathleen James]: Yeah. [Speaker 5 ]: Slightly different topic. I'm just curious what you see as the the, the value proposition for using AI in state government to either get more efficient, reduce costs, produce better outputs. And we're talking a lot about the risk, which is appropriate. It's scary, but there are, you know, there are opportunities, I would assume, for [Mark Holmes]: And we're taking advantage of some of them today. So in example, like I was saying before, like, in the cyberspace, you know, we're we have gotten ourselves up to a level of efficiency by using AI, you know, without having to grow a workforce to eighty people, that kind of thing. There's other areas within the state. I think if [Desiree Shanks]: If you jump up to slide seven Yeah. We actually I've covered a few of them. Okay. [Mark Holmes]: Yeah. [Desiree Shanks]: Okay. So I can talk through a couple of these, and I I won't spend a lot of time on it. But we've done, a lot of AI pilots. I think one I'll I'll hit a couple of of them that are pretty tangible. One was we had a research university reach out to public safety wanting to have officer narratives about incidents involving certain types of drugs, or a study that they were doing. And we wanted to support this study, but those officer's names are full of, like, people's names and places, and, like, you you can't just, like, give those away. So we we did a pilot where we used AI to preliminarily redact that, and it still had to go through a person. But the like, it went from being a thing that would take weeks to a thing that took, like, about one day for her to go through them because the vanity plates the AI couldn't find the vanity plates. She had to go through and find those many links. But, like, that was one example. Another one is the report that ADS puts out each year includes, like, a one pager for every project. And two years ago, I believe, I asked our MO director to try using the time, chat g p to do that report. Right. Because what the feedback we had gotten was this report is really hard to understand because you have to be, like, a nerd from IT land and also a policy wonk from whatever the policy area is order to understand what this project is actually trying to deliver. And that was a hundred page report that year. And so I was I had high hopes that that was going to save a lot of time. And I asked her afterwards how much time it saved, and she said several hours. And I said several hours. Are you serious? That's not very much time for a hundred page report. And she said, no. But this is the first year that quality has been what we really wanted. It's the first year that somebody could actually read this and under like, a normal a normal human could read and understand what we're trying to accomplish with these IT projects. And that theme has come up many times where we'll we'll go into something looking for efficiency and find that we've been doing more with less for so long that we kind of have like, we're not doing things to the level of quality that we necessarily would like to be able to. And so using AI actually frees up our staff to zoom out from the mechanics of the I'm typing the report to, like, making sure that they're in editing mode and they're explaining things well and it's clear and it can be understood. We run into that there. We run into that. Like, my team uses AI to do code reviews before we deploy things. And, again, we didn't find it that much time savings. What we found is our code is now documented, and we can actually maintain it. [Mark Holmes]: And and that aligns to sort of the AI enterprise maturity model. You know, generally, where most enterprises are, including Vermont, is we're at that AI coworker type where, essentially, AI is being used as sort of the peer peer review to what you're normally doing so that the end product's better or that, you know, it you you can create a perspective that's not normally yours, you know, those sorts of things. We're not at the point where agentic AI is happening so that you got one person doing your work out, you know, eighty people out there by pushing a couple buttons, but it is a proof of things. [Desiree Shanks]: So we are seeing some efficiency gains. We're seeing more quality gains Mhmm. Which we now kind of expect first to get those quality gains. And then we start to see efficiency gains as we build more specific power tools Yeah. For our users. Yeah. Right. [Chair Kathleen James]: Brett Campbell? [Vice Chair R. Scott Campbell]: Yeah. Thanks. So do you do you know where AI is being used? Do you need to press the government? I I guess I'm thinking of using it for what I'm thinking of specifically, you go you go to a doctor's appointment, say, and they talk to you about whatever it is, and it turns out that they're using AI to write the write the notes because they're recording they're recording the, the discussion. Are we are we doing that in I don't know where we would do that in state government, but are we doing that in state government? And and do you know where it's being used? [Desiree Shanks]: So we put together an inventory. I believe there are around thirty, maybe thirty four use cases on it right now, and that's on the website and was included in the report, the ADS annual report. That particular use case, we are piloting. We're going to be launching it kind of more broadly over the next few weeks. I don't have a specific date, but taking kind of the transcripts of the meetings that we have in the state that are okay to have a transcript for their their public record type transcript, being able to summarize those. There are some, specifically within government, some public records implications of using AI in that particular way that we're working with Messara on. And some states have said you just can't, but we think that there's a path to being able to take advantage of that type of tool within state government. So something we're working on. I think one other big one, ChatBT, I mentioned on this on this slide. We will first roll this out a little over a year ago, and we did another expansion of use cases of it yesterday, actually. And that's kind of an internally built version of a tool like ChatGPT. It uses a large language model that's deployed in our environment in the cloud. So all the data in it stays private to us. So state employees can use it with things that we could put into the cloud normally. So we don't put, you know, very sensitive confidential information in it, but people can do things like drafting emails and and work like that that normally we wouldn't do in a ChatGPT or a public tool because ChatBT respects privacy and stays within our our, boundary. And we've got, around three hundred active weekly users before our expansion on on Monday. So we're we're getting pretty good usage of that. That is a a states that have been doing chat g p t enterprise pilots or Microsoft Go pilot pilots, are normally running around two hundred users, like, much bigger states than Vermont. So the fact that we have, like, three hundred folks using it every week, it's it's pretty great. And so we have a lot of state employees who are taking advantage of it kind of for, like, those daily productivity tasks. [Vice Chair R. Scott Campbell]: That's that's all internal to, state government employees working with each other and all of that. How about as far as interactions with the public? [Desiree Shanks]: We have a couple of things. During the floods, we we wanted to allow municipalities to purchase off of the state contracts for things like gravel and emergency construction services and things like that. And, the challenge was that that data was being housed in a very old access database being turned into a twenty page tiny print PDF, and it was being updated incredibly fast by people who are very dedicated, but we're making typos because of how fast they were going. And so we actually put an AI interface on top of that. It was our first public facing production AI. I don't think anyone using it would have realized it was AI, but it would let you search for something and it would find the most relevant contracts, for that, which meant you didn't have to use the right words. You didn't have to use the weird misspellings that might be in the document, and you could still find what you were looking for. So that was one public one. I don't think anyone would have been like, oh, there's AI, but there was AI underneath it. It's making it effective. And we're shortly going to be launching and I don't have a date. I'm waiting for the new AI director to finish his background checks. But we'll certainly be launching public facing chatbots. The approach we're taking is looking at publicly available data and data that could be publicly available, turning it into question and answer pairs using AI, letting state staff review those and confirm that they're correct, and then deploying those in a chatbot. So when you ask a question, it'll find the most relevant question we have an answer for, rather than the other way of doing chatbots is to, like, let the AI make up an answer based on your your state websites. We think there's a lot of risk in our websites line. So we're not taking that approach. Instead, we're kind of pre pre loading it with questions and answers, for, like, FAQs and then letting people ask questions and find the most relevant answer that they may ask in another language. They may ask like, use it. They don't know the magic words to ask, to get the right answer. The AI will help them to find, the right answer starting with those. [Chair Kathleen James]: Would that be on, like, an agency by agency? Okay. [Desiree Shanks]: Yep. We're working with HR on a pilot. We're working with the marketing office on a pilot. And then also talking with human services about doing a pilot there as well. [Vice Chair R. Scott Campbell]: Sounds like sort of a sophisticated search [Desiree Shanks]: mechanism. Yeah. Yeah. And the goal is just to make sure that all the content is human curated so we don't have we don't run into the situation where we have our state websites lying. That's what I'm avoiding. For for [Mark Holmes]: a nineteen ninety seven document out there inadvertently, you know, establishing a policy again. Yeah. Right. Let's see. So that was so that we jumped ahead because there was a question about if we go back to slide six. So Yeah. So we were I don't remember exactly what the question I'd let you jump ahead, but rolling back to sales. The another thing that this council is we see ourselves since we've talked about, you know, how these different areas of specific AI expertise are needed. With that, we're foreseeing, you know, a little bit more of the bifurcation of our more generalized AI guidance and and, code of ethics to to also kind of start to lean in those areas as well. So if we have the resources, we can create those subcommittees. We can, you know, recommend more specific policy, more, you know, realistic life driven policy. And so the the questions that were asked of us at a time are just kind of like, you know, there's the report and the council, and then you wanted to talk more specifically about the division within the ABS and artificial intelligence. [Chair Kathleen James]: Yes. Are we on PhD? [Mark Holmes]: Yeah. So seven is really kind of the the where where that part of it starts. [Chair Kathleen James]: Okay. [Desiree Shanks]: And I've covered seven, I believe. [Mark Holmes]: Yeah. So Yeah. We can jump. Okay. We can jump to eight. [Desiree Shanks]: Alright. So at the beginning of of twenty four, we had the opportunity to and ADS to combine the data and AI, the the data office and the AI division and create a data and AI office. And the reason we did that was we saw that all of our AI pilots were starting to run up against data challenges. Right? Like, the data wasn't ready. The data couldn't be used for the thing we wanted to use it for for like, it was collected for one thing, and then we were trying to use it another way. And we're just, like, running into data related issues. So we saw that opportunity to combine the offices. And, and I moved from being the AI director to being the chief data and AI officer and then have been recruiting a new AI director, to come in. So, but what that has let us do is, have those AI conversations kind of all throughout our our life cycle, of projects. So as soon as we engage in a a data conversation on an IT project or somebody comes to us with a a data related problem, we're able to have a conversation about, is AI part of the solution here? Is this is this part of how we resolve this issue? And also on all the teams within the data and AI office, we've been using them as pilots. So I've got folks who are using AI to stand up, like, data intake forms and things like that more efficiently. I've got folks doing code reviews, things like that. We kind of pilot it within the data and a AI office and then share the learnings with the rest of ADAS. And, like, this works really well. Just go ahead and plug this in, and it prevents us from having to kind of relearn or refigure things out, with AI again and again. So, I've I've included kind of this practice oriented it's it's not really org chart, but kind of the five, components of the data and AI office. And this is all pretty new. We created our data practice back when ADS was formed, but it was really focused on serving human services, basically. So it was a pretty significant expansion last year when we we started providing those services throughout the rest of the state from a more centralized location. We also have our privacy and governance practice, which we've we just created last year, and we have a director of data privacy who's engaged in these conversations to make sure that as we're plugging in AI and using these advanced technologies that we're still preserving Vermonter's privacy and also the privacy of our our state employees. So, anyway, he's really great, and we're we're glad to have him on board. The AI practice is still the the AI division of one position, currently vacant, but, we'll have somebody in there shortly. We have a quality management practice that's really focused on how do we deliver solutions across ADS, effectively and make sure we're not having to go back and redo work because we missed something or, you know, things that tend to play in IT projects. And then finally, our constituent engagement, which is our public tracing websites, both our apps and our, like, vermont dot gov websites, how we engage with Vermonters in a consistent way as we think about all the lovely buildings we have here. Right? Like, it used to be Vermonters came to these buildings to interact with us, and now they they largely come to our websites to interact with us. So we're trying to treat them with the same level of of rigor and thought to things like accessibility and the effectiveness of her monitor's interactions online as we would have in in these companies. So that's how we're gonna ask. Any questions on that? Yeah. [Vice Chair R. Scott Campbell]: I'll ask you. So practice is an unusual word here at least in in terms of government. It it sounds is it sort of synonymous with with service? Is that is that what we're Yeah. Or is there is there [Desiree Shanks]: a broader meaning? Yeah. So we're kind of, yeah, we're we're back of the house type type of an organization. So, yeah, they're the services we provide to the rest of state government. We've organized them into these kind of five areas that we've been calling practices. [Vice Chair R. Scott Campbell]: Okay. But but if I substitute the word service, that would be about the scope of what you're talking about? Yep. I I just try to understand. Thanks. [Desiree Shanks]: And then oh, sorry. Are there other questions on this? [Chair Kathleen James]: I don't think so. Oh, okay. Nope. [Desiree Shanks]: So on the next one, slide nine, I've talked about kind of the the priorities that we're working on for this year. So this is all stuff that's in flight that we're doing now in each of these groups. I won't read them all. I'm I'm really excited about the disaster response support that we're we're trying to build in as a a core capability within our data group. We had a lot of people who were being heroic, during the the flooding events of the last few years and really appreciate that, but would like to rely less on heroism and more on, we just know how to do this. And it's it's a standard part of how we do things. So we're we're calling it blue sky, so that, you know, even on days when it's when it's not raining or whatever the bad thing is that's happening, we have the capabilities and we're having those conversations with safety and human services and others that are preparing and doing readiness activities. Another big one under constituent engagement is accessibility. There's a pretty significant push from Department of Justice on improving accessibility of state services. We have until, I believe, it's, twenty twenty six, to achieve kind of a much higher level of compliance than was previously required. And so that's another, big add that my team is engaged in right now. And then on AI, I mentioned the the AI power tool framework. That's how we're talking about AI within state government. We're not trying to, at this point, build kind of autonomous agents where there's, like, a a human over oversight over, you know, many robot minions. We think that that's not the way to deliver good outcomes for Vermonters. Instead, we're we're working on building power tools for our state employees to be more effective with how they with the work that they do. So I and I like the power tool metaphor because it it implies somebody who's reasonably well trained. Like, most of the power tools my ten year old does not have access to in the house. And and she has a lot of oversight when she is using the cordless drill, which is the one that she's able to use. But so a power tool implies a level of training. It implies specific tasks that we use it for, and the tool is used in particular ways to to do that safely. So when we're building out, I guess, tools like our chat BT and and specific things within that, that's really the paradigm we're we're pushing people toward is how do we give you a power tool that helps you think more about outcomes and less about mechanics and meets these other criteria? And that keeps us from falling into some of the common AI pitfalls where there isn't adequate human oversight or, you know, things like that happen. And I think that gets us to the end of the day. So [Chair Kathleen James]: Rev Morrow. Yeah. And Rev Campbell has his hand up with you. [Speaker 5 ]: I have two questions. One is, like, you were testifying the other day about the lot dot gov RFP and all new vendors and all that. So would you incorporate something like AI in an RFP to stipulate that the the next iteration of Vermont dot gov would utilize AI, or is that not appropriate? I believe we did include that under the auxiliary services. [Desiree Shanks]: We don't see it as or yet, but we wanna start moving in that direction. And that is a conversation we're having with every vendor we interview is, yeah, what's your what's your posture on AI? What capabilities do you have? Can you align with hardcoded ethics in our our way of doing it? [Mark Holmes]: That that seems to be the most common pathway at which AI is entering into the enterprise. It's not, you know, going out and getting a whole new AI product. It's it's the embedded AI and the products that we've been using forever. Mhmm. And so we've had to work, you know, within our procurement processes and our requirements gathering processes to suss those out and say, okay. We want this, but, you know, we wanna be careful that you're not providing it without, you know, human oversight and those kinds of things. Mhmm. [Desiree Shanks]: I I think examples that come up a lot, like, a lot of the the TV news cameras. Right? A lot of them now have AI in them. And that's great. It helps it zoom in on the right person who's talking and things like that. But it also means that the contracts for those things need to have terms that say, you can't use the state data to go train your AI model further. You can't you can't take state data and send it to somebody else or share it with anybody else. Like, we have to plug in more terms because there's AI risk associated with these AI in our devices. Yeah. [Mark Holmes]: A great example is, you know, we're all using the Microsoft three sixty five suite, and they are very actively moving the Copilot, their version of AI into it. And as we've been relicensing, we've been chain you know, addressing the terms in the license that don't comply with our AI direction and all the data use and all that. [Speaker 5 ]: My other question is can you explain in layman's terms why AI hallucinates? [Desiree Shanks]: There's there's a variety of reasons. I think the easiest one to understand is AI is I I tell my staff to treat it like a very eager intern. It wants to impress you. It wants to show you that it's smart and it knows what it's talking about. And so it will come up with things and very happily tell you things confidently that are actually not so because they saw it in a hypothetical somewhere or it's it's just getting confused based on the context of what you've given it. It can take, like, things that you say and come back to you with, like, interpreting those things as facts rather than questions. [Mark Holmes]: Yeah. It'll sometimes take data points and extrapolate the relationship between those datas. And they're they're they're in to be the source of the hallucinations. Right? They may be valid data points, but for very different reasons. [Desiree Shanks]: So I guess one one example, because we tried the, can we put a chatbot on our websites, like, right when chat g b d first came out? I was really excited. So, like, this is gonna be great. And I built one around our company, so all that COVID money we received and fed it all of the information we had publicly available and then asked it, if I wanna put a new roof on my school with ARPA money, who do I contact? And so it gave me back an answer that had some of the ESSER, the the education ARPA money information in it, some natural resources, information about runoff because it assumed a roof would have runoff. And then gave me contact information for somebody over at natural resources and said that's the person to [Mark Holmes]: call if you wanna put it in your [Desiree Shanks]: fund or school. And I was Yeah. And then that was when I was like, well, I guess we're not doing AI this way. Yeah. Yeah. Not gonna work. Context is really hard. And grab snippets from lots of places and try to come up with a coherent answer, and it will look good. [Mark Holmes]: Yeah. [Desiree Shanks]: It always looks high quality, but you can't use what it looks like as a proxy to, like, the actual quality of the IPs. [Mark Holmes]: Yeah. I don't know if [Desiree Shanks]: that was layman's terms. It was not blow. Yeah. Okay. [Vice Chair R. Scott Campbell]: In layman's terms, it might be AI explaining. [Chair Kathleen James]: Yeah. Brett Campbell? [Vice Chair R. Scott Campbell]: Yeah. But it does but what you just mentioned about asking a a question about where do I go for what this service, That sounds like really, that's the the immediate application for AI. I'm thinking about the integrated eligibility and enrollment thing that we've been working on for the last, I don't know, decade or fifteen years or something something amazing. That that that would really make that a possible thing. I'm also thinking about permits for for for building, housing, or whatever. I mean, having having a a way to direct people to where they need to go would be fantastic because, obviously, all of the various things that state government does and provides are, you know, incredibly confusing to most people. [Desiree Shanks]: One of the things that we've been talking about with and and with other states is people spend a lot more time figuring out how to interact with state government than the actual, like, interactions. And on the IT side, historically, we spend a lot of time making sure that our forms are easy to fill out and beautiful, or maybe not quite beautiful, but, like, at least adequate. But there people might spend three or four times as much time finding the form and figuring out what they need before they actually, like, go fill out the form. And then it goes into a black hole somewhere, and eventually, they get an answer. And one of the items that we have on our our AI strategic plan is thinking about service like, how we provide services to the monitors all the way from them recognizing that they need something through service delivery and using AI to improve that throughout that whole life cycle. We're trying to avoid optimizing just our back end processes or just our, like, the form intake process and really focus on how can we make sure that Vermonters can interact with us in a kind of a fluent way or, like, just be confident in their interaction with us from realizing that they need something all the way through. Yeah. Being delivered to them. [Vice Chair R. Scott Campbell]: Yeah. Get get get that get that fixed for you. [Desiree Shanks]: Yeah. Yeah. Be right on [Mark Holmes]: Four times a week. [Speaker 5 ]: Well, that was very helpful. Thank you. Not [Mark Holmes]: a question. [Chair Kathleen James]: Do we have any other questions? Rep Campbell? Rep Kleffner? [Bram Kleppner]: What, what keeps you up at night relative to AI and Vermonte? [Mark Holmes]: So Josiah and I are are we're we're a good pairing in the sense that he's he likes the policy side of things. So he he probably stays awake thinking about the the policy sides on more of the what if we could do this person. And so I think that there's a lot of opportunities that we don't yet discuss because and it's important. I'm not saying we've neglected them, but the AI problem set comes with, let's deal with the risk first. And I think there's opportunities out there on the how we could guide AI to assist the state that maybe we need to have more conversations as well, not in any way ignoring, you know, ethical use and and guidelines to help us get there. But, you know, I I fear that while we're ahead of the game for most states in terms of recognizing the risk and developing policy, I don't I think Vermont has an opportunity to be ahead of the game on the other side too, like smart agriculture, things like that for the state. I don't have a specific idea in mind, but just throwing out areas that I know that, you know, we try to help our our business community develop in these areas that we can help. So [Desiree Shanks]: I think a lot about being fixed right now. Yeah. He's the the risk guy. So, yeah, being the risk guy. And I remember having a conversation with my pastor about deepfakes, and he was like, wait. So somebody can make a video of me doing something with one of our congregants, and, like, no one would tell it was fake. Yeah. Something that could like, that's something that could happen. And so, like, how we operate in that world, how we make sure that we have protections for folks who are impacted by that. So I've been working with other states. New Jersey has some really good draft legislation on this. And one of the things that because I'm a risk guy, I've been talking about with our council about new speak some language forward for next year. So if you if you have any deepfake stuff that you I could wanna talk about, I'm happy to talk about that. I'd like but that's the biggest one for me right now. [Mark Holmes]: So far it's okay. I have to buy lunch now. [Chair Kathleen James]: We did just high low I'm sorry. Go ahead. [Mark Holmes]: Let's go ahead. [Bram Kleppner]: Those are relatively mild. Right? I mean, there are a lot of people who catastrophize and say, hey. If someone's gonna give AI instructions to make paperclips and not tell it not to use human bodies as raw material for the paperclips, then it's gonna eat us all up and turn us into paperclips. [Mark Holmes]: And [Bram Kleppner]: you aren't touching anything anywhere near those. AI is gonna call buildings to blow up or any [Mark Holmes]: of that stuff. [Desiree Shanks]: The reason for that is I think that sometimes that catastrophizing distracts us from, like, the real things that are right here. And, like, there are real real bad things that that are happening already right now that we can do something about. Some of those farther out ones are, a, maybe aren't as likely and certainly aren't things that we can take a lot of action, like, at this level meaningfully right now. And so that's why the the ones I'm focused on are for this. Thank you. [Chair Kathleen James]: Anybody else? Alright. Thank you so much for joining us.

Play Clip Play Full Video Pause

SmartTranscript of House Energy and Digital Infrastructure 2025-05-06 2:30pm

This is useful for seeing what happened at the meeting without wading through the whole transcript: