The state government created a position overseeing AI in K-12 education statewide over a year ago, when the Utah State Board of Education hired Matthew Winters as its full-time AI specialist. Agnew said Utah can collect stronger data about how AI is used in schools and get a clearer picture of what is and isn’t working.

Dixie Technical College staff have recently discussed the issues related to AI development in education. Utah Tech University Provost Michael Lacourse spoke at length on this subject during the university’s convocation ceremony last month, shortly before a new cybersecurity degree program was unveiled for this semester. 

Utah Tech Associate Professor of Computer Science Curtis Larsen told St. George News that it is “always risky” to speculate about the forms AI may take in the future. However, he said he believes that a few forms of AI seem likely, even though they may feel unfamiliar at first.

The first likely form, Larsen said, is truly personal agents.

“Not a generic chatbot, but a trusted assistant that learns your preferences, filters the firehose, summarizes what matters and takes routine actions across your apps,” he said.

“Think of it as a private curator aligned to your goals, not an ad-driven feed,” he added.

Larsen said he thinks that a second likely form will be clinical copilots and surgical robotics in the medical field. He said AI could be somewhat involved in procedures such as a laparoscopy, where a fiber-optic instrument is inserted into the body to examine certain organs.

“I don’t see AI replacing surgeons, but I do see systems that plan procedures, guide instruments, watch for safety issues and automate narrow tasks in laparoscopy and beyond — always with clinicians directing and reviewing the work,” he said. “Health care has higher stakes and tighter regulation than self-driving cars, so the pace will be measured, but task-level autonomy is plausible.” 

Larsen said another likely form is what he called “backstage autonomy” across the economy. This means that company warehouses, labs, customer support and campus operations would be run by specialized AI agents operating “backstage” while humans would handle exceptions that require additional attention.

“These systems are less visible than chatbots but potentially more transformative,” he said, adding that, “across all of this, the guardrails matter.”

Those guardrails include privacy, provenance, liability and clear human oversight. 

“The promise is productivity and focus; the risk is over-reliance,” Larsen said. “We should design for the former and protect against the latter.”

Lacourse had previously spoken about how the university will soon be using agentic AI to help students.

When asked about how he thinks this should be implemented, Larsen said he doesn’t think the answer is “a single, one-size-fits-all university chatbot.” 

“The near-term path is specialized, domain-specific agents embedded in the tools students and faculty already use,” Larsen said. “General models are excellent for broad help, but they often miss program-level details. Coding agents are the clearest early example: they follow well-defined workflows and boost productivity, but they still require a developer to direct and review the work.”

Larsen said he expects agentic AI to show up at Utah Tech in three ways. 

The first way will be student-facing course agents that turn out syllabi, rubrics and datasets that will be delivered to students through the university’s learning management systems. He hopes that this will “free students to spend more time on the creative and conceptual work — without replacing fundamentals.” 

Secondly, there will be AI faculty assistants that operate on real files by editing slides, aligning outcomes and assessments, proposing better exercises and flagging issues with pacing or flow. 

Thirdly, operational agents are expected to help with routine university workflows, with humans verifying the results.

“None of this is ‘set it and forget it,’” Larsen said. “Effective agents will be specialized, supervised and integrated with clear guardrails for data privacy and academic integrity.”

Larsen said that most of his colleagues in Utah Tech’s computing department use AI in some form, although they do so “primarily as a tutor or assistant, not an ‘easy button.’” 

“We encourage students to use these tools to learn faster, and we’re redesigning assessments — especially in first-year courses — to verify understanding,” he said. “Because many traditional assignments can now be completed by AI, we’re shifting toward work that requires students to explain their reasoning, show their process and defend their code.”

He said that “AI lets us raise the bar” in upper-division classes, where students can accomplish more in the same amount of time with an AI assistant at their side, but are still accountable for the concepts and for every line they submit.  

“They may type less, but they must understand more,” Larsen said. “As for misuse, it happens — just as with any new technology — but we’re keeping pace by emphasizing assessment design over detection. Oral check-ins, in-class problem-solving, version-control history, and code reviews make it harder to pass off AI’s work as your own and easier for us to confirm genuine learning.”

Blueprints and warning signs

Larsen said that there are three things that he wishes more people understood about AI.

First, he wants people to know that today’s AI systems are powerful pattern recognizers, but they are not minds.

“They generate fluent text by predicting likely words from training data; they don’t have intent, self-awareness or human-level understanding,” he said.

Secondly, he would like people to treat conversations with chatbots as if they are willingly sharing their personal data with the chatbot. 

“Depending on the provider and your settings, prompts and files may be stored and used to improve the service — so don’t paste sensitive or confidential information,” he said. 

Thirdly, Larsen would like to remind people that “anyone speaking with certainty about AI’s long-term trajectory is making an educated guess — including me.” 

“The field is moving fast, and timelines and impacts remain uncertain,” he said.

So, in light of that, what are the drawbacks of AI that people should be worried about or prepared to face in the future?

“AI delivers real value, but the risks are just as real,” Larsen said, adding that one of the risks with AI is quality, since “these systems can be fluent and confident while being wrong, and they can reflect biases in their training data.”  

He said that subject-matter expertise and verification are non-negotiable tools, since “you need to tell the jewels from the junk.”

Another concern with AI is data ownership.

“Prompts and files can be stored or reused depending on the service, and source attribution and copyright remain unsettled,” he said. “Treat anything you paste as shared data and favor tools with clear enterprise controls.”

In the long term, there is a risk of online content degrading after years of relying on inferior AI too much. This would mean that, if you look up the answer to a problem online, the AI-generated answers or search results you receive may be much less reliable, since the answers themselves could be based on data compiled sloppily with AI.

“As more online material is AI-generated, models can end up training on their own outputs — a feedback loop that degrades quality unless we curate high-quality, human-verified datasets,” Larsen said.

The final risk of AI that Larsen pointed out is over-reliance.

“If we let AI drive the process, we can slowly give up judgment and skill,” he said. “The safeguard is straightforward: keep AI in an assistive role, require human review for consequential decisions, and design workflows that keep people — and their expertise — in charge.”