Wendy Graham’s two kids first encountered Amira, a popular artificial intelligence tutoring platform, in 2025, during a summer reading program in New Mexico. She also used it at home with her daughter, now a fourth grader.

But she found that the purple-haired avatar with the big, round glasses didn’t seem to understand where her daughter was struggling. When she stumbled over “usually,” the “learning agent” wanted her to repeat the word and spend time defining it. 

“She knew what the word meant; she just didn’t decode the word,” Graham said. Amira, in her view, “got hung up on all the wrong things.”

Her son, on the other hand, never gave Amira a chance.

“He just looked at me and looked at the screen like, ‘This isn’t happening,’ ” Graham said.

Her kids’ lack of enthusiasm for the program reflects what researchers have concluded so far about AI tutors. While some states are spending millions on Amira and other virtual tutors, most kids aren’t sticking with them long enough to benefit. For now, high-dosage tutoring models proven to be effective are human-led and depend on “a relationship that keeps students showing up and engaged,” according to from the National Student Support Accelerator, a Stanford University center that conducts tutoring research. 

“We don’t have solid research showing that AI tutoring can work in the U.S. at scale,” said Susanna Loeb, the center’s executive director. Researchers’ “biggest finding is that the take-up is very low.” 

Amira’s leaders, however, say their platform was based on extensive prior research showing AI tutoring can be effective . 

“No other scalable instructional intervention has been interrogated as many times, by as many independent teams, with this consistency of positive impact,” CEO Mark Angel said in a statement to ĂŰĚŇÓ°ĘÓ.

The brief comes as Amira, which records students’ voices to test their reading progress, faces increased scrutiny from districts in New Mexico. Parents expressed over how long the recordings are stored and what Amira does with the data. The New Mexico Public Education Department to use the program three times a year to assess K-2 students’ literacy skills. But in response to the opposition, it is allowing districts to , and Amria says it does not use the recordings for . 

Last month, the Las Cruces Public Schools, where Graham’s children attend and where she also teaches high school history, told parents they have disabled the recording feature and that they could request a paper assessment instead of Amira’s digital version. Officials also said they would discontinue use of Amira’s tutoring program.

“Teachers,” the memo said, “will continue to monitor progress and provide support for students to become proficient readers.”

‘Supercharge teachers’ efforts’

New Mexico parents’ reactions to Amira reflect broader reservations about the use of AI in schools. In from the Century Foundation, a left-leaning think tank, 84% of Americans said they’re worried about tech companies using students’ personal data to make a profit, and nearly half want AI use in the classroom kept to a minimum. 

The vast majority of survey respondents, 81%, expressed concern that teachers are being pressured to use AI even if it hasn’t been proven to help students learn. 

But Angel, with Amira, said the tool “can supercharge teachers’ efforts” by allowing them to listen to every child read aloud and identify struggling readers before third grade.

In 2024, the Duval County Public Schools in Florida signed a contract with Amira to help about 2,600 students in grades second through fourth reading below grade level and then tied half of the $100,000 contract to students’ progress. According to a of the work, more than 1,200 of the students met their goals for oral reading fluency, which exceeded the expectations of the contract.

Five hundred students made gains on state tests and about 300 reached the proficiency level, but the report doesn’t identify whether those results met the targets Amira was expected to reach.  

Graham, who previously taught fifth graders, appreciates what districts using Amira hope to accomplish.

“I think they’re trying to take things that have historically been done by teachers and streamline it to free up teachers’ time,” she said. But she still thinks states and districts aren’t getting their money’s worth. “Kids don’t do their best for robots; kids do their best for people.” 

Tutoring programs, however, can benefit from AI, the Stanford team concludes. A 2024 study, also from Stanford, showed that human tutors can become more effective with prompts from an AI tool. Tasks like scheduling, tracking attendance, analyzing notes from sessions and giving students additional practice outside of tutoring sessions are other ways that AI can support tutoring, according to the researchers.

The North Little Rock School District in Arkansas used , an AI platform that measures kids’ reading accuracy and fluency, as part of a research study. Students were expected to read to the AI program for 20 minutes, read a physical book independently for 20 minutes and then write about the topic, said Carmen Langston, the district’s literacy coordinator. 

On the screen, students would find digital books that teachers assigned or could create their own version, picking animals or scenery they prefer. 

“The kids like to put their names in there,” Langston said, and those who used it for the recommended number of minutes saw progress in phonics. But the district, she said, is trying to limit screen time. “We tell our teachers that if you are doing [an] intervention, it has to come from a human teacher.”

‘Engagement was thin’

The Stanford center’s research summary follows its of an AI tutoring program showing that students averaged about two minutes per week in one district and about five minutes in another — well under the two 30-minute sessions they were expected to complete each week.

A released last month tested the use of Khanmigo, Khan Academy’s AI tutor that relies on OpenAI’s technology. During the 2024-25 and 2025-26 school years, researchers Philip Oreopoulos and Nina Low randomly assigned students in 18 Hamilton County, Tennessee, middle schools to use the program for math practice. But while nearly all students with Khanmigo used it at least once, “engagement was thin,” they wrote. 

Students only consulted Khanmigo a third of the days they practiced math and just 17% of the times they made a mistake. Learning gains were about the same as they would have been had students not used the program, the researchers wrote. 

The results disappointed Sal Khan, who told that for students, the program’s release was a “non-event.”

Education Secretary Linda McMahon recently discussed the use of AI during a , saying there are not yet “a lot of metrics” to determine its value in the classroom.

“Are we seeing better outcomes in schools? And if we’re not, then they should be pulled out,” she said, adding that she thinks a parent who is dissatisfied with an ed tech tool “has every right to go in and say, ‘No, I want something different for my child.’ ”

The secretary, however, has repeatedly praised , a chain of for-profit AI-focused schools. Students spend two hours learning from AI and spend the rest of their days learning “life skills” like running a food truck or producing a musical, according to the website.

“The beauty of that is that the student almost has a one-on-one tutor,” she said. “The teacher can monitor to see if the child is a bit behind and repeat information for them. If the child is really excelling, then their programs get sped up a bit, so it’s a great tool.”

State contracts

Several states feel the same way about Amira and have spent millions to make it available to districts. New Mexico’s contract costs about $2.7 million a year. Iowa in 2024 to make licenses available to nearly 250 schools across the state and boosted the contract by to expand access to more schools.

The North Dakota Department of Public Instruction has a $2.6 million to implement Amira and points to showing that the more kids use it, the better their performance gets. But the data was self-reported by Amira and didn’t include a comparison group. The department declined to comment on research showing low student engagement. 

Utah doesn’t have a statewide contract with Amira and didn’t endorse it. An of interactive software programs showed that students considered at risk for falling behind in reading who used Amira outperformed a similar group of students who didn’t use the program. But of the 8,036 first through third graders deemed at risk, only 249 met the recommended usage time.

“Therefore, these results should be interpreted with sample size in mind,” the Utah evaluation said.

Amira also commissioned an analysis of the AI tutor’s use in Louisiana schools. showed better results for students who spent more time on the platform. That makes sense, Stanford’s Loeb said, but doesn’t prove the “causal effect of Amira.” The state spent in COVID relief funds on the program and in December, authorized another for additional AI tools, including Khanmigo and Writable.

Ted Beasley, a spokesman for the Louisiana Department of Education, said high-dosage tutoring has helped drive recent improvement in reading on the National Assessment of Educational Progress. 

“As we’ve expanded tutoring statewide, AI programs like Amira give school systems another option to complement that work,” he said. “A number of school systems have chosen to utilize this tool and have found it useful.”

As new research continues to emerge, other states are exercising caution. North Carolina’s recently for fiscal 2027 includes $500,000 for a dual enrollment math course using Khanmigo, as well as $2.5 million for Khan Academy partnerships with schools. Khanmigo can be turned off if districts decide. Lawmakers, however, originally recommended for the AI program.

Loeb still sees AI tutoring as promising, but said purchasing Amira or Khanmigo is not the same as “buying tutors for everybody.” She stressed that careful execution matters. That’s also what a on “intelligent tutoring systems,” one of the studies that inspired Amira’s founders, concluded. Positive results depend in part on “the adequacy of program implementations,” the researchers wrote.

“If you choose a tool that has some evidence or that you think could be effective, you have to figure out how to actually get kids to use it,” Loeb said. â€œIt’s going to take a teacher, a parent or some school person setting up the context where the students feel motivated to do the work.”