Julianna Brion for NPR For a few weeks last school year, a small group of students in Dayton, Ohio, regularly sat in front of laptops in conversation with a social-emotional learning chatbot named Jordan.

Unlike AI chatbots that are designed to be agreeable — and even sycophantic — Jordan presents as angsty: On the day NPR visited, the chatbot told students, "Lately, I've been feeling like what's the point of any of this? I'm not failing. I just don't care. You ever feel like that?" "Jordan's a little whiner, but I always help him out," said K, a sophomore at the Alternative School at Jackson Center, which is part of Dayton Public Schools.

"I built a little relationship with him," K explained. "So I don't really see it as like an AI on a computer. It's Jordan — like, I'm texting somebody. That's how I see it." Many students at this alternative school would otherwise be suspended or expelled, and staff told NPR that students here often struggle with both behavioral and academic challenges. NPR is using only K's first initial because of the stigma associated with attending a school for students with a history of disciplinary problems.

Dayton schools paid the ed tech company SchoolAI $72,600 last academic year for multiple AI tools, including the Jordan chatbot, which administrators designed in collaboration with the company and an outside consultant. District leaders said they wanted to see whether Jordan could help students open up about their emotions and experiences outside of school, information that could be useful for school therapists and staff. (According to administrators, families and students were notified that the chat logs could be viewed by school employees, an external evaluator and SchoolAI staff.) Instead of directly asking students how they feel, Jordan asks students for advice and then grades their responses for engagement and authenticity.

K, who spent just part of his sophomore year at the alternative school, said Jordan could be frustrating at times, but he also found the chatbot's responses relatable.

"I would say it helped me kind of solve my own problems … just answering Jordan every morning and getting him through his problems," K said. "I will go home … but I won't have nobody to talk to.

So one day … I put myself in a Jordan perspective. I'm just talking to myself, and it's helped ever since." Across the United States, school districts and individual educators are experimenting with AI in the classroom.

Teachers are using it to develop lesson plans special educators are using it to track student progress in some districts, students have access to AI chatbots that give them feedback on their work; and a San Diego charter school recently introduced its students to an AI-powered humanoid robot.

"It feels very Wild West in terms of the randomness," says Robin Lake, director of the Center on Reinventing Public Education, which has been tracking AI adoption in K-12 schools.

Meanwhile, the research around how schools can use AI to help students — and avoid potential harms — is extremely limited, according to a recent Stanford University review of more than 800 academic papers According to that report, some of the existing studies suggest that carefully designed AI tools show more promise than general-purpose chatbots.

There's also no consensus around how schools and educators should be measuring what works and what doesn't. "There are so many questions.

This is bigger than individual schools, even individual districts can figure out on their own. We need state and federal leadership," Lake says.

But states and the federal government have been mostly hands off, which means many schools are forging ahead on their own.

NPR asked Dayton Public Schools Superintendent David Lawrence about his district's decision to experiment with AI on vulnerable students, like the ones at the alternative school that K attended. He said educators must constantly adapt to keep up with the changing times.

"I think when the history is written, it'll say that 'Hey, they were really cutting-edge, innovative and ahead of the curve in the interest of children.'" The challenge of measuring what works and what doesn't In Dayton, administrators say they will continue using AI at the alternative school, and they will more than likely deploy more social-emotional chatbots like Jordan. The district also renewed its contract with SchoolAI at a cost of $78,650 for the current school year.

Mike Kentz, an external consultant the district hired to help design and evaluate the pilot, says the experiment produced enough positive evidence — drawn from chat logs and feedback from students and staff — to continue testing. However, he says he doesn't think it's ready to be rolled out to students across the district.

Kentz says that's partly because it's difficult to measure the success of such an experiment. "Every test, every experiment that I try to measure in some sort of concrete way, I can only get as far as perception," he says, referring to how students and staff feel about it.

Katelyn Schoenhofer, the AI specialist for Wichita Public Schools in Kansas, says they're facing the same problem. "We wrestle with what does success really mean when it comes to AI use? And what metric do we want to use in order to evaluate that? And we have not come to an answer yet," she says. That challenge has prevented Wichita schools from moving forward with a large-scale rollout of student-facing AI tools. But they've slowly begun to introduce the technology in more controlled ways.

In the earlier grade levels, Schoenhofer says, students don't have access to chatbots, but they do learn about the technology. "So it's a robot. It's pattern recognition. It is not your friend," she explains.

In middle school, students have limited access to tools like Canva AI for image generation, and she says that last school year, students were allowed to use the tool's coding feature to create educational games.

And in high school, some students in upper-level math courses can use an AI math tutor from the ed tech company Edia.

Until they figure out how to measure the value of these AI initiatives, Schoenhofer says, these experiments will remain small, but they will continue.

"We are still going ahead with caution but with optimism with our students because it is something that is going to impact their lives.

It is something they need to be educated about," she says. Some school communities are pushing back on experiments Mark Beehler, superintendent of the Salamanca City Central School District in western New York, had planned an AI experiment of his own, until community outrage halted it.

Beehler's district is located on territory that belongs to the Seneca Nation. The district, which serves many Native American students, paid a tech company called Realbotix nearly $58,000 for an AI-powered humanoid robot named Sally.

Beehler told NPR in July that he planned to bring Sally to some of the district's high school classrooms this fall. He said many students struggle to ask for help from teachers and friends and said, "I'm anticipating that some students will be more likely to interact with a humanoid robot than they would with an actual human." He added that the robot could help teachers as well.

"When they need to know something, they can just ask Sally, you know, 'Hey, what's the history of this? I'm not exactly sure.' And boom, you know, in 30 seconds to a minute, it spits it out." But as news of the planned pilot spread, teachers, parents and even state education leaders raised concerns around student data privacy, among other things. By the end of July, the district had put the Sally plan on hold Beehler did not respond to NPR's follow-up requests for comments.

In July, he said: "I don't mind being a little bit on the forefront of how we can make this work for a school and ultimately benefit the students." The pressure to proceed … with caution Rebecca Winthrop, a senior fellow who researches AI and education at the Brookings Institution, says there's good reason to be cautious.

"I don't think you should just sort of wildly experiment, especially with really marginalized communities," Winthrop says. She says these kinds of experiments could have unintended consequences for students, like "cognitive stunting, diminishing their learning potential" and shifting how students are socialized. Winthrop worries that students could become reliant on AI to do their thinking or navigate relationships for them.

That's why she says she doesn't believe anyone under age 18 should have access to general-purpose AI chatbots like ChatGPT and Claude.

But she does support educators experimenting with AI in narrowly targeted ways — for example, using a chatbot trained on a teacher's grading rubric to provide feedback to students.

"Educators should really be at the forefront [of AI experimentation], and it should be [for] a problem they're facing," Winthrop says. But even as experimentation continues, some districts are leaning away from AI.

The country's largest district, New York City Public Schools, issued a one-year ban on student-facing AI beginning this school year for its preschool through 8th-grade students, while allowing only limited AI pilots for a small number of high school classrooms. And the second-largest district, the Los Angeles Unified School District, recently followed suit banning students from using AI on district devices. As LAist has reported, the district previously allowed students 13 and older to use AI tools under school supervision and after they had completed a lesson on "digital citizenship." Lake, of the Center on Reinventing Public Education, agrees that AI poses significant risks to students. But she says there are also risks in not engaging with the technology.

"If we really close off AI experimentation, if we prevent kids from learning about AI, learning how to use it really well, thinking about how teaching could be transformed, we will set our kids back from being able to navigate a world that is changing outside of schools faster than we can possibly imagine." To keep up, Lake says, the world inside schools also needs to change.

Edited by Nicole Cohen Audio story produced by Lauren Migaki Visual design and development by LA Johnson