Artificial intelligence often induces fear, awe or some panicked combination of both for its impressive ability to generate unique human-like text in seconds. But its implications for cheating in the classroom — and its sometimes comically wrong answers to basic questions — have left some in academia discouraging its use in school or outright banning AI tools like ChatGPT.
That may be the wrong approach.
More than 8,000 teachers and students will test education nonprofit Khan Academy’s artificial intelligence tutor in the classroom this upcoming school year, toying with its interactive features and funneling feedback to Khan Academy if the AI botches an answer.
The chatbot, Khanmigo, offers individualized guidance to students on math, science and humanities problems; a debate tool with suggested topics like student debt cancellation and AI’s impact on the job market; and a writing tutor that helps the student craft a story, among other features.
First launched in March to an even smaller pilot program of around 800 educators and students, Khanmigo also allows students to chat with a growing list of AI-powered historical figures, from George Washington to Cleopatra and Martin Luther King Jr., as well as literary characters like Winnie the Pooh and Hamlet.
Khan Academy’s Chief Learning Officer Kristen DiCerbo told CNN that Khanmigo helps address a problem she’s witnessed firsthand observing an Arizona classroom: that when students learn something new, they often need individualized help — more help than one teacher can provide all at once.
As DiCerbo chatted with AI-powered Dorothy from “The Wonderful Wizard of Oz” during a demonstration of the technology to CNN, she explained how users can rate Khanmigo’s responses in real-time, providing feedback if and when Khanmigo makes mistakes.
“There is going to be a big world out there where people can just get the answers to their homework problems, where they can just get an essay written for them. That’s true now too on the Internet,” DiCerbo said. “We’re trying to focus on the social good, but we need to be aware of the threats and the risks so that we know how to mitigate those.”
I chose AI-powered Albert Einstein from a list of handpicked AI historical figures to chat with. AI-Einstein told me his greatest accomplishment was both his theory of relativity and inspiring curiosity in others, before tossing me a question Socrates-style about what sparks curiosity in my own life.
Khanmigo developers programmed the AI figures not to comment on events after their lifetime. As such, AI-Einstein wouldn’t comment on the historical accuracy of his role in Christopher Nolan’s “Oppenheimer,” despite my asking.
Some figures from the list are not as widely praised as Einstein. For instance, Thomas Jefferson, the third US president and primary draftsman of the Declaration of Independence, has faced renewed criticism in recent years for owning 600-plus enslaved people throughout his lifetime.
Khanmigo’s Thomas Jefferson will not shy away from scrutiny. He wrote back to my inquiry about his views on slavery in part: “As Thomas Jefferson, my views on slavery were fraught with contradiction. On one hand, I publicly expressed my belief that slavery was morally wrong and a threat to the survival of the new American nation […] Yet I was a lifelong slaveholder, owning over 600 enslaved people throughout my lifetime.”
The purpose of the tool is to engage students through conversation, DiCerbo said, an altogether different experience than passively reading about someone’s life on Wikipedia.
“The Internet can be a pretty scary place, and it can be a pretty good place. I think that AI is the same,” DiCerbo said. “There could be potential bad uses and misuses, and it can be a pretty powerful learning tool.”
After gaining early access to ChatGPT-creator OpenAI’s newest and most capable large language model, GPT-4, Khan Academy trained GPT-4 on its own learning content. The company also implemented guardrails to keep Khanmigo’s tone encouraging and prevent it from giving students the answer to the question they’re struggling with.
For teachers, Khanmigo also offers assistance to create lesson plans and rubrics, identifies struggling students based on their performance in Khan Academy activities and gives teachers access to student chat history.
“I’m learning new ways to solve the problems as well,” said Leo Lin, a science teacher at Khan Lab School in California and an early tester of Khanmigo. Khan Lab School is a separate nonprofit founded by Khan Academy CEO Sal Khan.
Khanmigo has emerged at a crossroads in academia, with some educators leaning into generative AI and others recoiling. New York City Public Schools, Seattle Public Schools and the Los Angeles Unified School District, among other academic institutions, have all made efforts to either ban or restrict ChatGPT on district networks and devices in the past.
A lack of information about AI may be exacerbating some educator worries: While 72% of K-12 teachers, principals and district leaders say that teaching students how to use AI tools is at least “fairly important,” 87% said they’ve received zero professional instruction about incorporating AI into their work, according to an EdWeek Research Center survey from June.
Khan Academy’s in-the-works AI learning course “AI 101 for Teachers,” created in partnership with Code.org, ETS and the International Society for Technology in Education, offers a path toward AI literacy among teachers.
Although Khanmigo is still in its pilot phase, the AI-powered teaching assistant is currently used by over 10,000 additional users across the United States beyond the pilot program. They agreed to pay a donation to Khan Academy to test the service.
An AI “tutor” like Khanmigo is not immune to the flubs all large language models face: so-called hallucinations.
“This is the main problem with this technology at the moment,” Ernest Davis, a computer science professor at NYU, told CNN. “It makes things up.”
Khanmigo is most commonly used for math tutoring, according to DiCerbo. Khanmigo shines best when coaching students on how to work through a problem, offering hints, encouragement and additional questions designed to help students think critically. But currently, its own struggles in performing calculations can sometimes hinder its attempts to help.
In the “Tutor me: Math and science” activity available to students, Khanmigo told me that my answer to 10,332 divided by 4 was incorrect three times before correcting me by sending me the same number.
In the same “Tutor me” activity, I asked Khanmigo to find the product of five numbers, some integers and some decimals: 97, 117, 0.564322338, 0.855640047, and 0.557680043.
As I did the final multiplication step, Khanmigo congratulated me for submitting the wrong answer. It wrote: “When you multiply 5479.94173 by 0.557680043, you get approximately 33.0663. Well done!”
The correct answer is about 3,056.
Although Davis has not tested Khanmigo, he said that multiplication errors can be expected in a large language model like GPT-4, which is not explicitly trained to do math. Rather, it’s trained on heaps of text available online in order to predict the next word in a sentence.
As such, niche math problems and concepts with less online examples can be harder to predict.
“Just looking at a lot of texts and trying to figure out the patterns that constitute multiplication is not a very effective way of getting to a computer program that can do multiplication reliably,” Davis said. “And so it doesn’t.”
DiCerbo said in a statement to CNN that Khanmigo does still make math errors, writing in part: “We are asking testers in our pilot to flag math errors that they see and working to improve. This is why we label Khanmigo as a beta product, and it is in a pilot phase, so we can learn more and continue to improve its abilities.”
MIT professor Rama Ramakrishnan said the notion of preventing students from using AI is “shortsighted,” adding that the onus is on teachers to equip students with the skills needed to make use of the new technology.
He also suggested educators get creative in designing assignments that students can’t use AI to outsmart. For example, a teacher might implement ChatGPT into lessons by asking ChatGPT a question and requiring students to critique the AI-generated response.
“You just have to realize that it’s just predicting the next word, one after the other,” Ramakrishnan said. “It’s not trying to come up with a truthful answer to your question, just a plausible answer. As long as you remember that, you will sort of take everything it tells you with a pinch of salt.”