Despite worries that the artificial intelligence program could make cheating easier, half of K-12 teachers report using it. Edtech companies are scrambling to release their own tools.
Diego Marin used to work late into the night at the end of each trimester submitting grades and comments for his 70 eighth grade math students at a Chicago public school. No more. Now he has a virtual assistant, ChatGPT, that cuts the time he spends on report card comments to an hour, by rapidly putting his individualized comments on each student into smooth, clean prose.
“As teachers, it feels like we’re short on time for everything,” Marin says. “I’ve had some nights where I’ve stayed up until midnight submitting comments.” He also uses the free artificial intelligence chatbot to help draft emails to parents, complete Individualized Education Programs for some students (such as those with learning disabilities) and even as a teaching tool.
Admittedly, Marin, 30, is into social media and the latest tech—he has 1.4 million followers on TikTok. But his embrace of AI isn’t unusual among teachers. Despite immediate fears after ChatGPT’s release to the public last November that the service would upend education by making cheating easier, more teachers seem to be using it to their advantage than worrying about that risk.
In a February survey of 1,000 kindergarten through 12th grade teachers nationwide, 51% said they had used ChatGPT, with 40% reporting they used it weekly and 10% using it daily. (ChatGPT is free, as long as users create an account with OpenAI. To access the latest version of the bot, built atop a more advanced version of AI—GPT-4, instead of GPT-3.5—users must pay $20 a month for a ChatGPT Plus subscription.)
About a third of teachers in the survey, commissioned by the Walton Family Foundation, said they use ChatGPT for lesson planning and coming up with creative ideas for classes.
Middle school and high school teachers were more likely than those teaching lower grades to say they use the AI. Twenty-two percent said they use it to communicate with parents, students and colleagues. It’s easy to see the appeal of the tool, particularly for out-of-classroom tasks: a 2022 EdWeek survey found the typical teacher works 54 hours per week, with five hours spent on planning and prep, three hours on administrative work and two hours communicating with parents.
Of those teachers who have used ChapGPT, 88% said it’s having a positive impact on education. Even teachers who haven’t tried it themselves are more likely to say it has had no impact than a negative one—44% to 10%, respectively. But just 9% of nonusers ascribed a positive impact to the tool.
Moreover, only one in ten teachers reported catching a student using the chatbot without their permission. It’s unclear whether that’s because cheating is fairly rare or because teachers are missing the signs of students inclined to cheat (or more charitably, cut corners) using ChatGPT. But overall, students have actually been slower than teachers to adopt the tool, according to a February survey of 1,000 students aged 12 to 17 also commissioned by the foundation, funded by members of the billionaire Walton family. Only a third of surveyed students said they had tried ChatGPT at all, with younger students in the sample actually reporting more use than older ones.
Not all school systems have embraced the AI tool. In January, the New York City education department, which oversees the nation’s largest school district with more than 1 million students, blocked the use of ChatGPT by both students and teachers, citing concerns about safety, accuracy and negative impacts to student learning. The Los Angeles Unified School District, Seattle Public Schools, the Fairfax County Public Schools in Virginia, and Montgomery County schools in Alabama have also blocked access to the AI.
In addition, another smaller survey shows more teacher skepticism—at least among those working at private schools. The National Association of Independent Schools (an association for private schools) recently surveyed about 200 teachers from its member schools and found 37% of them thought that AI in schools was a negative development, compared with only 33% who said it was positive. Another 30% of teachers said they thought it was neutral.
But some private school teachers, too, are in the vanguard of experimenting with the chatbot. Phillips Academy Andover in Massachusetts, one of the nation’s fanciest college-prep schools, does not have any blanket policies about using ChatGPT. “Andover is a really big private school, and we have plenty of faculty members that are excited by this, and there are many other people that are very nervous about it,” says Nick Zufelt, a computer science and math teacher.
Zufelt is in the excited category. He first introduced ChatGPT to his students as subject material. “I was having conversations with my students about what it was doing and why it’s doing what it’s doing,” Zufelt says.
Now he’s beginning to use it as a tool in a new computer science seminar he’s teaching, that includes both students who have never written a line of code before and some with several years of coding experience. Zufelt is allowing students to have ChatGPT write a first draft of the code they’re working on. “I’m more interested that students end up with an understanding of the code than an ability to write it from first principles,” he explains. “Learning to think like a programmer is a lot harder than learning to program.”
Patrick Powers, an English teacher at Navo Middle School in Denton, Texas, has also proactively introduced ChatGPT to students; he’s encouraged them to use it to practice for debate assignments, develop business proposal templates for a “Shark Tank”-like pitching simulation and gather information about historical figures. “[Students] appreciate it being an interactive tool instead of just using Google or instead of using the lesson plan,” Powers says. “They’re just more engaged with the lesson and content itself.” He also touts ChatGPT’s clean writing as a “strong work example” that his eighth graders can learn from.
Like Marin, Powers finds ChatGPT’s writing strong enough to use when reaching out to families and says it enables him to increase the frequency of those communications, even while saving hours of email writing time a week. “Instead of being reactive and saying, ‘Hey, Johnny had a bad day in class,’’’ Powers says, he now uses the chatbot to draft regular emails to parents about what’s going on in class and how students are succeeding.
Cheating on essays has been one of the biggest concerns for educators, but Powers isn’t too worried about it. “I had a good feel for the voice of my students, so I knew the type of work to expect from them,” Powers says. “And before I introduced it, I sent out a letter to parents letting them know about the expectations in the classroom and how this tool could help us learn greater content in an innovative way.” (For those wondering, he drafted the letter in his own voice, without ChatGPT.) That said, Powers does believe a program that detects AI-written work would be useful. (GPTZero is so far the most well-known service that can identify ChatGPT-generated text).
As teachers experiment on their own, edtech providers (like those in other industries) are racing to build services on top of GPT-4. Online learning nonprofit Khan Academy debuted a closed beta version of its new AI tutor, Khanmigo, earlier this month. While students can use the older version of ChatGPT to write essays or solve problems for them, Khanmigo will act only as a coach, says Sal Khan, founder and CEO of the operation that carries his name.
“If you could go to two different tutors, and one tutor says ‘Gimme your homework, I’ll just do it for you. Here—turn this in,’ and the other tutor says, ‘Okay, let’s think about this together. How would you approach that?’ The second tutor is going to be far better for the student,” Khan says. Notably, Khanmigo won’t just give students the answer, even if they ask.
Khanmigo is also better at math than the free version of ChatGPT. One of the well-known limitations of that free version built on GPT-3.5 is that it very often gives incorrect answers to basic math problems, even if it accurately explains the concepts behind the solution. Because Khanmigo is built using GPT-4, its math skills are much improved.
Marin has occasionally witnessed GPT-3.5’s math fails during class and now uses them as a learning opportunity. “When I introduced ChatGPT to my students … we had ChatGPT open on the projector and it was going through a sequence of how to solve a system and it actually came up with the wrong answer,” Marin says. “For a second I thought my kids were just pulling my leg to be honest—because that’s what they like to do—but they said ‘No, Mr. Marin, it’s telling us that this is the answer, but it’s not.’ So that’s where we went back and found a mistake.” Now, he routinely asks ChatGPT to solve problems incorrectly and asks students to point out where the bot went wrong and why.
Teachers interested in piloting Khanmigo can sign up for a waitlist, and once selected, are asked to donate $20 monthly to use the service. Running Khanmigo is expensive—it costs Khan Academy about $0.05 for every 600-700 words that Khanmigo generates, and Khan anticipates that the AI will cost about $10 to $15 per student per month to run. The nonprofit has spent millions to build Khanmigo and hopes to secure extra philanthropic support for the project.
“Right now, it’s not cheap to be running this stuff,” Khan says. “The reason why we’re moving so aggressively on this, but in a safe way, is we think those costs are going to come down dramatically. And so we’re going to make it a lot more accessible in the coming months.”
Journalists, bloggers and other users have revealed how easy it can be to get around GPT’s guardrails. A NewYork Times tech columnist published a disturbing conversation with Microsoft’s Bing chatbot—also built using GPT-4—that showed the AI, which called itself Sydney, fantasizing about hacking computers, spreading misinformation, and harming humans, despite guardrails that are supposed to prevent Bing from talking about such topics. (Additional guardrails have since been added.)
Khan is well aware of this, and has built additional safety measures into Khanmigo to keep students from straying into questionable, non-educational conversations with the chatbot. All Khanmigo chats are logged, and if a student runs up against one of Khanmigo’s guardrails, teachers are notified.
ChatGPT also has a tendency to make things up—called hallucinations in AI-speak. Joe Welch, a history teacher at North Hills Middle School in Pittsburgh, saw this first hand, when he was seeking specific information on a niche topic.
A student working on a report shared that he had found conflicting information from different data sources about the extent of damage and casualties caused by a series of tornadoes that hit Western Pennsylvania in 1985. “I asked ChatGPT for a table of casualties, towns, time of the tornado, and monetary value of damage and the table came back with extremely inaccurate numbers,” says Welch. In other words, while a middle-schooler recognized the numbers he was finding were in conflict, the chatbot admitted to no such uncertainty, producing a definitive (but wrong) table.
Despite this, ChatGPT has saved Welch tons of time. He uses the chatbot, among other things, to create tables for student study guides, which would have previously used up at least an hour of prep time.
“It’s just cleaning up a lot of tasks that would’ve been arduous for me to complete,” he says.