Evan, a high school sophomore in Houston, was stuck on a calculus problem. He fired up Answer AI on his iPhone, took a photo of the problem from his advanced math textbook, and ran it through the Homework app. Within seconds, Answer AI generated the answer, along with instructions on how to solve the problem.
A year ago, Evan would comb through lengthy YouTube videos to work through his homework assignments, and he also had private tutors who charged $60 an hour. Now, the rise of AI bots is threatening established tutoring franchises like Kumon, a 66-year-old Japanese tutoring giant with 1,500 locations and about 290,000 students across the US.
“The hourly rate for a tutor is roughly the same as the price of a year's subscription to Answer AI,” Evan told me. “So I thought, [in-person] “I'm a tutor.”
Answer AI is one of a handful of popular apps that are leveraging the rise of ChatGPT and other large-scale language models to help students with everything from writing history papers to solving physics problems. According to May 21 data from Data.ai, five of the top 20 education apps in the U.S. App Store, including Answer AI, are AI agents that help students with their schoolwork.
There's been a constant debate about what role AI should play in education. The benefits of AI tutoring are clear: it would make after-school tutoring much more equitable. A $60 hourly tutor in Houston is already much more affordable than tutoring in more affluent, academically competitive areas like the Bay Area, where tutoring can be three times more expensive, Answer AI founder Rick Chou told me.
Zhou, a serial entrepreneur, also suggested that AI could enable more personalized instruction, which is difficult to achieve in a class of 20 students. A chatbot teacher who remembers a student's study habits and never gets grumpy about answering a question could replace the private tutors that wealthy families hire. Myhan, a high school senior in Houston, said her math grades improved from 85 to 95 within six months of using generative AI to study.
For now, AI tutors are primarily limited to text-based interactions, but in the near future they could be able to speak to students in a way that optimizes each student's learning style, such as a more empathetic, humorous, or creative style. OpenAI's GPT-4o has already demonstrated that AI assistants capable of generating voice responses in a range of emotive styles are within reach.
When AI doesn't help you learn
The vision of AI-powered, unbiased learning has yet to be fully realized. Like other apps that forward API calls to LLM, the AI tutor suffers from hallucinations and sometimes spits out wrong answers. Answer AI tries to improve its accuracy through search augmentation generation (RAG), a way to fine-tune LLM with specific domain knowledge (in this case, a large set of problems). But it still makes more mistakes than previous generations of homework apps that match user queries against a library of existing practice problems, because these apps don't try to answer questions they don't already know.
Some students are aware of AI's limitations: Evan frequently cross-checks Answer AI results on ChatGPT, and Myhanh uses Answer AI in after-school study sessions to exchange ideas with peers. But while Evan and Myhanh are self-starters and more likely to use AI to help them learn, some of their peers don't learn anything from entrusting their homework to AI.
Screen capture feature with Answer AI Chrome extension / Image: TechCrunch Image credit: Answer AI
For now, educators are unsure of how to handle AI. Some U.S. public school districts have banned access to ChatGPT on school devices, but a blanket ban on generative AI would be difficult to enforce as soon as students leave school grounds.
In reality, it is impossible for teachers and parents to stop their children from using AI for their studies. It is hard to tell if a student has memorized a math problem from the answers they write, and there are (currently) major flaws in detecting if an essay was written by AI. So rather than banning AI altogether, it may be more effective to teach children that AI is an imperfect assistant that sometimes makes mistakes.
China's Advantage
As of May, the two most popular AI helpers in the U.S. are both owned by Chinese companies. Year-old Question AI is the brainchild of the founders of Zuoyebang, a popular Chinese homework app that raised about $3 billion in equity over the past decade. Gauth, meanwhile, was launched in 2019 by TikTok parent ByteDance. Since its launch, Question AI has been downloaded 6 million times in the U.S. on Apple's App Store and Google Play Store, while rival Gauth has garnered twice as many installs since its launch, according to data from market research firm SensorTower. (Both are publicly traded in the U.S. by Singaporean companies, a popular tactic amid growing Western scrutiny of Chinese technology.)
ChatGPT solves arithmetic sequence problem / Image: TechCrunch
The success of Chinese homework apps is the result of consistent efforts to target the American market in recent years. In 2021, China imposed rules to crack down on the burgeoning tutoring sector, which focuses on the country's public school curriculum. Since then, many service providers, including brick-and-mortar tutoring centers and online learning apps, have pivoted to overseas users. Given their sheer size, the United States is unsurprisingly their most coveted international market.
The fact that tutoring apps likely use similar underlying AI technology levels the playing field for foreign players, who can overcome language and cultural barriers by invoking AI to study user behavior. As Eugene Wei writes in his well-established analysis of TikTok's global success:[A] Highly responsive and accurate machine learning algorithms can pierce the veil of cultural ignorance.”
And because they rely on the same LLM group, it's hard for these learning apps to differentiate themselves on the quality of their answers alone. Some established companies, like Zuoyebang and PhotoMath, are able to combine generative AI with searching a vast library of problem sets to improve accuracy. New entrants will need to find other ways to differentiate, such as by beefing up user personalization features.
“AI agents need to actively engage with students and tailor their answers to their individual learning needs,” Zhou says. “Raw language models are not an out-of-the-box AI agent, so we differentiate by fine-tuning our AI to teach more effectively. For example, after providing an answer, our AI bot will prompt students for follow-up questions, encouraging deeper learning rather than just having them copy the results.”
Updated to correct text that said AI can be used to detect essays written using AI. This cannot be done.