Ailsa Ostowitz is accused of using AI in three assignments in two different classes this school year.
“It's mentally exhausting because I feel like I know this is my job,” said Ostowitz, 17. “I know this is my brain putting words and concepts down on paper so other people can understand.”
Ostowitz, a senior at Eleanor Roosevelt High School in the Maryland suburbs of Washington, D.C., told NPR of one of the accusations she received from a teacher. The September message included a screenshot of an AI detection program showing that Ostowitz had a 30.76% chance of using AI on a writing assignment that included a description of the music she was listening to.
“I write about music. I love music. Why use AI to write what I want to talk about?” Ostowitz said.
Ostowitz reached out to teachers about the assignment through the school's online learning platform. “I was serious when I said I wasn't using AI. Can you try a different detector?”
The teacher did not respond and withheld Ostowitz's grade.
Ostowitz's mother, Stephanie Lisk, said her daughter is a high-achieving student who cares about doing well in school, and was surprised that her teacher was quick to judge Ostowitz's efforts so early in the school year.
“Get to know their skill level, so an AI detector might be able to help,” Rizk says.
Rizk told NPR that she met with her teacher in mid-November, but that the teacher never saw her daughter's messages.
The district, Prince George's County Public Schools, said in a statement that Ostowitz's teachers use their own AI detection tools and the district does not pay for the software.
“We advise educators not to rely on these tools during staff training, as multiple sources have documented potential inaccuracies and inconsistencies,” the statement said.
PGCPS declined to interview Mr. Ostowitz's teachers. Rizk told NPR that after the meeting, the teacher no longer believes Ostowitz used AI.
But whatever happened to Ostowitz shouldn't come as a surprise.
More than 40% of sixth- through 12th-grade teachers surveyed used AI detection tools in the past school year, according to a nationally representative poll by the Center for Democracy and Technology, a nonprofit organization that advocates for civil rights and civil liberties in the digital age.
This is despite numerous studies showing that AI detection tools are far from reliable.
“It is now fairly well established in the field of academic integrity that these tools are not fit for purpose,” says Mike Perkins, a leading academic integrity and AI researcher at the UK University of Vietnam.
Perkins found that some of the most popular AI detectors, such as Turnitin, GPTZero, and Copyleaks, flag things as AI that are not AI, and vice versa. When the AI text was manipulated to appear more human-like, the accuracy rate decreased further.
“We've seen some very worrying issues with some of the most prolific AI text detection tools,” he says.
Despite these issues, NPR found that school districts from Utah to Ohio to Alabama are spending thousands of dollars on these tools.
Why one of the largest districts in the U.S. uses AI detection software
Broward County Public Schools, near Miami, is spending more than $550,000 on a three-year contract with Turnitin. The long-established education technology company has historically provided plagiarism detection software to schools. In 2023, it introduced an AI detection function. When educators submit student work to the tool, it generates a percentage that reflects the amount of text that the software determines is likely to have been generated by AI. There is one caveat. According to the company, scores below 20% are unreliable.
“The Turnitin tool helps facilitate conversation and feedback rather than grading,” said Sherry Wilson, director of innovative learning for the Broward School District, one of the largest school districts in the country with more than 230,000 students.
Wilson said the district is “fully aware” of research showing that AI detection tools, including Turnitin, are not 100% accurate or reliable.
Turnitin also acknowledges this. The company's website states, “Our AI writing detection is not always accurate and should not be used as the sole basis for adverse action against students.”
In a statement to NPR, Turnitin wrote that avoiding false accusations of student misconduct is more important than catching every AI post.
Wilson said the Turnitin tool remains valuable and can save teachers time by quickly scanning student work for suspected AI use.
Another reason Broward teachers have access to this tool is because the district participates in academic programs, Wilson said. Some schools, such as the International Baccalaureate (IB), require teachers to authenticate student work before submitting it for external review.
Both the IB and International Education at Cambridge programs offered by Broward told NPR that schools are not required to use AI detection software as part of the certification process. Nevertheless, Broward said in a statement to NPR, “We want our teachers to [Turnitin] As one of the tools to meet your requirements. ”
But Wilson insists that the ultimate authority on whether a student's work is their own is the teacher, not the AI detection tool.
“They use these tools as feedback and have teachable moments with their students,” she says.
Why one teacher uses AI detection tools
Language and literature teacher John Grady said the AI detection tool provides a “starting point” for starting conversations with students who may have used AI.
“It's certainly not foolproof,” he says. “But it gives you something to hang your hat on.”
Grady teaches at Shaker Heights High School, part of the Shaker Heights City School District outside of Cleveland. The district, which has about 4,400 students, paid GPTZero, another AI detection software company, about $5,600 this year for annual licenses for the district's 27 teachers. This tool calculates the percentage chance that a student's work was generated by AI.
Grady said all student essays are submitted through GPTZero. If the tool indicates that there is a 50% or higher chance that AI was used for the assignment, Grady will investigate further. This includes using the revision history tool to see how much time students spent on assignments and the number of edits they made during the writing process. If a student seems to have only made a few edits and spent little time writing, check in with them.
“And I say, 'Hey, this was flagged. Can you talk to me about why?' If it was AI, most of the time, 75% of the time, you'd be like, 'Oh, yeah.' So I thought, 'Okay, now I have to rewrite it with less credit,'” Grady says.
Edward Tian, co-founder and CEO of GPTZero, says this is how educators do it. should using his company's tools.
“We absolutely do not believe this is a means of punishment,” Tian said. “This has to be a tool in the toolkit, not the be-all and end-all.”
He says it's important to understand that a GPTZero probability score of less than 50% means the text was likely generated by a human rather than an AI. He says any score above 50% warrants closer examination. This is the same as Mr. Grady's explanation.
Tian does not dispute research showing that GPTZero is not always reliable. But he notes that some educators, like Grady, still find the information it provides valuable.
He said such tools provide “signals of what's going on in the classroom,” but teachers should always follow up with students if those signals indicate something concerning.
People skeptical of AI detection
“My writing style sometimes seems like AI because of the repetition of words that I use. I think it's because my vocabulary is limited,” said Ji Xi, a junior at Shaker Heights whose first language is Mandarin.
Shi, who is not a student at Grady's, said she is still honing her writing skills and worries that the AI detection software may be biased against non-native English speakers like herself.
Although some educators share this concern, research to date is limited and contradictory.
Shi said an assignment he completed in his English class earlier this fall was flagged by GPTZero as potentially AI-generated. He said his teacher suggested that his use of an online tool called Grammarly may have triggered the detection software. Grammarly uses AI to correct your grammar and generate text when prompted. (A teacher confirmed Shi's account to NPR.)
Shi said he only used Grammarly to organize his writing and that he wrote the assignments himself. “I was really disappointed when I saw the comments about it being reported as AI,” Shi said.
Shi believes AI detectors should be thought of as “smoke alarms, which are signs and warnings.” However, in some cases, it may be a false alarm.
He questions whether districts need to spend thousands of dollars on AI detection software. He says more money could be spent on professional development for teachers.
Carrie Cofer, a high school English teacher in the Cleveland Metropolitan School District, just a few miles from Shaker Heights, agrees.
Last year, as an experiment, she uploaded a chapter of her Ph.D. Paper to GPTZero. “And then 89% or 91% of it was written by AI, and I thought, 'Oh, no, I don't think so, because it's all mine,'” Cofer said.
Cofer is helping develop the district's AI policies and guidelines. She says she will advocate against it because Cleveland schools are not currently paying for the AI detection software.
“I don't think it's a good use of their money,” Cofer said. “The kids are going to get around it one way or another.”
Possible workarounds for students include using AI detection software themselves, doing workshop assignments to avoid being flagged, and using “AI humanizer” programs that claim to make AI-generated sentences appear more human-like.
Ultimately, she says, teachers will need to adapt to AI by changing the way they teach and assess student learning.
Ailsa Ostowitz, a high school senior back in Maryland, is also adjusting. She now runs all her homework assignments through multiple AI detection tools before handing them in.
Although the text is her own, she says she will rewrite the text that the software identifies as potentially AI-generated. This additional step adds approximately 30 minutes to every assignment.
“I think I’ve become more cautious about presenting my work as my own rather than AI,” she explains.
She doesn't want to miss the chance.
This report is Tarbell AI Journalism Center.
Editor: Nicole Cohen
Visual design and developer: LA Johnson
Audio story production: Lauren Migaki
