John Hult-SD Searchlight
When he’s not at the Capitol, South Dakota state Rep. Al Novstrup helps oversee his amusement parks in Sioux Falls and his hometown of Aberdeen.
He’s not an expert on off-label prescription drug use.
The Republican professed as much during a Feb. 5 speech to the House of Representatives, delivered in support of a bill to expand the availability of ivermectin and hydroxychloroquine. The two drugs are championed in certain circles as COVID-19 remedies, but not FDA approved as treatments for the disease.
“I’m not qualified to know the answer,” Novstrup said, “but here’s some of the things I’ve learned by just doing some quick artificial intelligence.”
Novstrup then proceeded to read parts of an AI-generated summary of a Google search on off-label prescription drug use.
“You’ve always got to be skeptical,” he said of the summary, from Google’s AI assistant, Gemini, “but you’ve also got to give it consideration.”
In an interview, the 24-year lawmaker told South Dakota Searchlight he’s used Gemini and a handful of other generative AI tools to build arguments for or against bills, likening it to pre-AI internet searches, and to pre-internet sourcing for floor speeches.
“Prior to Google searches, we relied on the Dewey Decimal System,” Novstrup said, referencing a cataloging framework used by libraries to sort books and other physical media. “Before that, we relied on our friends and neighbors, or the newspaper or the radio.”
The term “generative AI” refers to tools that can produce or summarize content, as opposed to simply finding it. When users search for information on Google, Gemini now places summaries of the content above the links to that content. Gemini can also be used to draft emails in Google’s Gmail platform, among other tasks. Tools like ChatGPT, Grok or Claude can return summaries of queries with links to sources, and can also be used to draft emails, write stories, essays, scripts or, during the recently concluded 2026 South Dakota legislative session, draft or refine bills and talking points on bills.
Novstrup said he’s aware of the potential pitfalls of generative AI — news coverage of its rapid proliferation are peppered with cautionary tales of factual errors, hallucinated sources or failures to put facts into context when summarizing articles — but he said lawmakers and the public have always had to consider the source of information their lawmakers use to support their arguments.
“The main thing is it’s a source,” Novstrup said. “Sometimes any one of them can be wrong.”
AI proliferates
Several lawmakers interviewed by Searchlight said they’d used some form of artificial intelligence to craft arguments or research talking points.
Others went beyond that. Rep. Kent Roe, R-Hayti, said he runs all his bill drafts through Grok, a generative AI tool developed by the Elon Musk-owned company xAI, to refine his ideas. Asking questions more than once, refining search queries and cross-checking the answers are key to getting trustworthy results, said Roe.
The more human interaction, he said, the better the outcome.
“You just recycle it 10 times, and if it comes up with 10 different answers, you know it’s not working,” said Roe, who also uses AI at his day job as an appraiser to speed up reporting on property values. “If you come up with nine out of 10 answers that are very similar, well, then you know it’s doing its job.”
Roe hasn’t used Grok to draft bills from scratch, but said it helps him improve his proposals. Like Novstrup, he sees the use of generative AI as a more evolved version of prior lawmaking practice. Lawmakers have long pulled bill language from other states, or from draft proposals pitched by industry or political interest groups.
“I think it’s safe to say there are no original ideas,” Roe said.
The speed of generative AI can be a blessing when lawmakers are crunched for time, said Amber Hulse, a Republican state senator and lawyer from Hot Springs.
This legislative session, some of her colleagues on the Senate Judiciary Committee asked her to review other states’ laws requiring people to contact authorities when a person threatens suicide. Aberdeen Republican Sen. Carl Perry had drafted a bill that would have criminalized a failure to do so in South Dakota.
She had two days to get the answer. Generative AI summaries told her where to look quickly enough for her to still attend to her other responsibilities as a senator.
“I think it’s really helpful when people need to do 50-state surveys of code,” said Hulse, whose research helped Perry craft amendments to his bill. “I was paid handsomely in law school to do that type of research for all different types of organizations. And before now, you had to do all that manually, look up every single state’s code and dig into the code. And now you can type into ChatGPT and say, ‘Hey, what do other states do for this area of law?’”
The bill ultimately failed, but that was because of “policy concerns,” Hulse said, not worries about whether the legislation would conflict with other laws or put South Dakota out of step with other states.
AI and bill drafting
Lawmakers like Roe argue that generative AI should help do that kind of work upfront.
Rep. John Hughes, R-Sioux Falls, told the House Taxation Committee on Feb. 12 that, with Roe’s help, he’d used Grok to turn his first draft of a bill about a new class of taxable property into a proposal that would square with the state’s existing property tax framework.
Like Hulse, Hughes is a lawyer, but his practice doesn’t typically involve deep dives on the ins and outs of property tax law.
Need to get in touch?
Have a news tip?
CONTACT US
Roe, Hughes told the committee, “ran this bill through AI and I am just astounded. Astounded,” Hughes told the committee. “Wow. It even makes me sound like I think perhaps I might make sense.”
The committee rejected Hughes’ bill that morning, but shortly thereafter advanced one from Rep. Scott Odenbach that sought to level out countywide property valuations for taxation purposes with multi-year averages.
The Spearfish Republican offered a sheepish admission about AI as he introduced it. Not about his use of it, but about how little he used it.
“I have used a little AI to help me with this, but maybe not enough,” Odenbach said.
Roe is a member of the taxation committee. When he asked Odenbach whether he’d done his due diligence to make sure his bill would work within the state’s existing tax structure, Odenbach said he “probably should have” run his bill through an AI “constitutional stress test.”
The bill was eventually amended and folded into a compromise property tax package between lawmakers and Gov. Larry Rhoden’s office near the end of the 2026 legislative session, but failed to earn the support of the state Senate.
AI guardrails
Odenbach is an enthusiastic backer of AI for lawmaking. He’s asked Grok to create first drafts of bills, though he said he couldn’t recall which ones.
When asked if the public should be concerned about AI writing bill drafts, Odenbach said anyone who might be terrified about the prospect will have a lot to be terrified of in the coming years.
“I think that terror can apply to all aspects of our society that are going to be transformed by AI very, very soon, in the next few years,” he said. “That terror can be handled by the extent to which we still require a human to interface with it, to have trust in it.”
John McCullough, head of the South Dakota Legislative Research Council, sounded similar notes on the topic, but he doesn’t share Odenbach’s enthusiasm.
The council manages the Legislature’s day-to-day operations, helps lawmakers draft bills and is responsible for updating the state’s code of laws. No proposal makes it onto the session’s list of bills without a review by the council’s team of legal experts.
McCullough’s not convinced any AI tool can perform trustworthy analysis of a bill’s potential conflicts with existing laws, and he doesn’t see a future where it replaces the human eyes of trained council staff.
“By design, we publish the law,” McCullough said. “Implicit in our mission or duty is to be the caretaker of the law. So you have to have a centralized drafting agency that is attempting to make the law be consistent throughout.”
Which is not to say that AI is strictly off-limits at the council in every context. Most legal research tools, including those used by the council, now incorporate AI to help lawyers find what they’re after more quickly.
South Dakota has the smallest state legislative staff in the U.S., McCullough said, so “if we can use technology to help us, then we’ll use technology to help us.”
But the help it can offer has limits, in McCullough’s experience. He recently asked a generative AI tool a question about “some obscure legislative issue” and got an answer that cited a South Dakota law that doesn’t exist. He told the AI tool as much, and it apologized and told him the law actually came from North Dakota.
That law didn’t exist, either.
“It was basically hallucinating,” McCullough said.
Concerns, quality standards
The council doesn’t have a formal policy on generative AI use. Lawmakers interviewed by Searchlight were largely confident that the legislative process would weed out errors, citing the council’s review of proposals and the multiple votes a bill needs to make it to the governor’s desk.
One former lawmaker, however, is concerned that might not be enough.
“I think it can be a very significant problem,” said R. Shawn Tornow, a Republican former state representative from Sioux Falls. “AI is not obviously old fashioned legal research and reasoned analysis, because one small trigger of incorrect information can greatly skew and make a legal position absolutely incorrect.”
Tornow was in Pierre in March to testify, as a private citizen, against bills he felt had been poorly written, which is an issue with human authors he said predates generative AI.
Generative AI, he said, is “at bare level, an inch-deep level starting point” in terms of crafting defensible laws.
Sen. Liz Larson, D-Sioux Falls, said her generative AI use has been largely limited thus far to drafts of communications with constituents and electoral messaging. Even then, she said, the work has to be refined into something that uses her own voice and style before it goes out.
Larson was troubled this year by what she saw as inadequate due diligence before AI conclusions appeared in testimony or in speeches.
“Those standards of rigor have gone down in the past year,” Larson said. “People aren’t even ashamed to say ‘I just looked this up.’ If something is on the Senate floor and up for final passage, you probably should’ve done the research before it got to that point.”