One morning this spring, a dozen or so students huddled in front of communal classrooms, staring at math lessons on their laptops.
The sixth graders at Khan Lab School, an independent school with an elementary school campus in Palo Alto, California, worked on quadratic equations, graph functions, Venn diagrams. But when they encountered questions, many did not immediately call their teacher for help.
They used a text box next to their lessons to ask for help from Khanmigo, an experimental chatbot tutor for schools using artificial intelligence.
The tutoring bot quickly responded to a student, Zaya, by asking her to identify specific data points on a graph. Then Khanmigo persuaded her to use the data points to solve her math question.
“It’s really good at taking you through the problem step by step,” Zaya said. “Then it congratulates you every time it helps you solve a problem.”
Khan Lab School students are among the first college kids in the United States to try out experimental conversational chatbots that aim to simulate one-on-one human tutoring. The tools can respond to students in clear, fluid sentences and are specifically designed for school use.
Based on AI models that underpin chatbots like ChatGPT, these automated study aids could herald a profound shift in classroom teaching and learning. Simulated tutors could make it easier for many self-directed students to hone their skills, dive deeper into topics that interest them, or tackle new topics at their own pace.
Such unproven computerized tutoring systems can also make mistakes, encourage cheating, reduce the role of teachers or hinder critical thinking in schools – forcing students to test subjects for what amounts to an experiment in teaching by algorithms. Or, like a legion of promising tech tools before them, the bots may simply be doing little to improve academic results.
Khanmigo belongs to the wave of new AI-powered learning tools. It is developed by Khan Academy, a non-profit education giant whose video tutorials and practice problems have been used by tens of millions of students.
Sal Khan, the founder of Khan Academy — and of Khan Lab School, a separate nonprofit — said he hoped the chatbot would democratize students’ access to individualized tutoring. He also said it could greatly help teachers with tasks such as lesson planning, allowing them to spend more time with their students.
“It will enable every student in the United States, and ultimately on the planet, to effectively have a world-class personal tutor,” said Mr. Khan.
Hundreds of public schools are already using Khan Academy’s online classes for math and other subjects. Now the nonprofit, which introduced Khanmigo this year, is testing the tutoring bot with districts, including Newark Public Schools in New Jersey.
Khan Academy developed the crash barrier bot for schools, Mr. Khan said. These include a monitoring system designed to alert teachers if students using Khanmigo seem fixated on things like self-harm. Mr Khan said his group was studying the effectiveness of Khanmigo and planned to make it widely available to the districts this fall.
Thousands of US schools are already using analytical AI tools such as plagiarism detection systems and adaptive learning apps designed to automatically adapt lessons to students’ reading levels. But proponents see the new AI-assisted tutoring systems as game-changers in education, behaving more like student-employees than inert pieces of software.
The AI’s facility with language has prompted some enthusiasts to declare that simulated tutors will soon be able to respond to students as individually as human tutors.
“The AIs will achieve that ability, to be the best tutor they can be,” Bill Gates, Microsoft’s co-founder and philanthropist, said at a recent conference for educational technology investors. (Khan Academy has received more than $10 million in grants from the Bill & Melinda Gates Foundation.)
Whether the bots can provide the kind of empathetic support and genuine encouragement that can make human teachers particularly effective is not yet known.
For more than a century, educational entrepreneurs have envisioned classroom devices being programmed to automatically test and instruct students.
As education writer Audrey Watters relates in her book “Teaching Machines,” researchers in the 1920s began to argue that computerized learning devices would revolutionize education. The machines, they promised, would free teachers from mind-numbing work and allow students to work at their own pace and receive automated feedback.
Over the decades, schools scrambling to adopt the latest computerized education technologies often found the systems finicky or faulty. Some concluded that the automated tools did little to improve student results.
Now, new chatbots are fueling a renewed campaign for automated learning tools. Khanmigo underscores the technology’s educational promise and potential drawbacks.
Khan Academy began developing chatbot tutoring software last fall with the goal of assessing AI’s potential to improve learning. The system uses GPT-4, a large language model created by OpenAI, the research lab behind ChatGPT.
Mr. Khan said he wanted to create a system to guide students, rather than simply give them answers. So Khan Academy developers developed Khanmigo to use the Socratic method. It often asks students to explain their thinking as a way to encourage them to solve their own questions.
Khanmigo offers help in a wide variety of subjects: elementary school math, high school American history, high school civics, and college-level organic chemistry. It also has features that invite students to chat with fictional characters like Winnie-the-Pooh or simulated historical figures like Marie Curie.
AI systems based on large language models can also invent false information. That’s because the models are designed to predict the next word in a sequence. They don’t stick to the facts.
To improve Khanmigo’s accuracy in math, Khan Academy developers created a multi-step process: the system works out answers to a math problem behind the scenes and then compares it to a student’s answer. Still, Khan Academy’s tutoring system displays a warning at the bottom of the screen: “Khanmigo sometimes makes mistakes.”
Khan Lab School, where annual tuition costs more than $30,000, provides an ideal test bed for guiding bots. The Silicon Valley school has small class sizes and an entrepreneurial philosophy that encourages children to pursue their passions and learn at their own pace. The technically skilled students are used to tinkering with digital tools.
One morning this spring, Jaclyn Major, a STEM specialist at Khan Elementary School, watched as her students playfully tested the limits of the bot.
A student asked Khanmigo to explain a math problem using song lyrics. Another asked for help with math in “Gen Z jargon.”
“Do you want to do me one more favor and explain everything in Korean?” said a third in a text conversation with the chatbot.
Khanmigo dutifully obliged. It then pushed each student back to the math task at hand.
Ms. Major said she appreciated how the system interacted with her students in an engaging way.
“Khanmigo can connect with them and be on their level if they want to,” she said. “I think it could be useful in any classroom.”
It is too early to say whether Khanmigo will appeal equally to other target groups, such as public schools with larger class sizes or students who are not used to directing their own learning.
In class, Zaya, the sixth grader, had encountered a glitch. Khanmigo had asked her to explain how she came up with the answer to a problem with a dataset. Then the bot falsely suggested that she might have made a “little mistake” in her calculations.
She promptly admonished the AI chatbot: “19 + 12 is 31 khanmigo,” she wrote.
“Sorry for my mistake earlier,” Khanmigo replied. “You are indeed right.”
That may prove to be one of the most important lessons for schoolchildren using promising new tutor bots: Don’t believe every AI-generated text you read.
“Remember, we’re testing it,” Ms. Major reminded her students. “We learn – and it learns.”
