How AI could close the education inequality gap – or widen it

Follow ZDNET: Add us as a preferred source on Google.

---

Within months of ChatGPT’s launch at the end of 2022, the nation’s largest public school system had banned the chatbot. It would have negative impacts on student learning, the NYC education department said, also citing concerns about the security and accuracy of ChatGPT’s content. 

Across the Hudson River, Franklin School took exactly the opposite approach. The private school in Jersey City, New Jersey, made AI a central part of its curriculum. Opened in 2022, the school began integrating AI tools not to replace teachers, but to enhance their work and deepen student engagement. 

“We looked at the integration on how to enrich the learning for students, but also, at the same time, we wanted to see where we could create efficiencies at our school for our teachers,” said Will Campbell, head of Franklin School. “We have incredible teachers in our building. How can we give them more time to be even better for our students?”

Also: Why AI chatbots make bad teachers – and how teachers can exploit that weakness

Franklin’s early AI experiments included custom chatbots trained on approved course material, designed to function as tutor-like learning aids. Faculty, including Campbell himself, found they could offload routine administrative tasks to AI, freeing up time for instruction and student support. The school also redesigned assessments, allowing students to use AI to tackle more complex problems that emphasize critical thinking. 

A similar philosophy emerged at the university level. Ethan Mollick, Ph.D., a professor at the Wharton School of the University of Pennsylvania, added explicit AI-use guidelines to his syllabus as early as January 2023, permitting students to use the technology in all of his classes.

Three years later, Mollick has become one of the leading voices advocating for AI in education. He has worked with AI companies, including OpenAI, to develop education guides and authored the New York Times bestselling book Co-Intelligence, which examines the role of AI in learning and work.

Also: Can AI save teachers from a crushing workload?

Why? He finds there is extreme value to unlock in AI. 

“We have some early evidence that it’s an incredibly powerful teaching tool,” said Mollick. “There’s a lot of potential to solve a bunch of the huge problems in education.”

On one side are those who say that AI tools will never be able to replace the teaching offered by humans. On the other side are those who insist that access to AI-powered tutoring is better than no access to tutoring at all. The one thing that can be agreed on across the board is that students can benefit from tutoring, and fair access remains a major challenge — one that AI may be able to smooth over.  

“The best human tutors will remain ahead of AI for a long time yet to come, but do most people have access to tutors outside of class?” said Mollick. To evaluate educational tools, Mollick uses what he calls the “BAH” test, which measures whether a tool is better than the best available human a student can realistically access.

“The answer is clearly yes already, and with a little work, it could probably become even better,” he adds. 

The tutoring gap 

Decades of research have shown that students learn more effectively with individualized tutoring, yet personalized, at-home instruction remains out of reach for many families due to staffing shortages and high costs. AI-powered tools now promise a tutor-like learning experience that students can access at home. 

Also: Students are using AI tools instead of building foundational skills – but resistance is growing

Research on the effectiveness of human tutors dates back decades. In 1984, educational psychologist Benjamin S. Bloom found that students who received one-on-one tutoring combined with mastery learning performed up to two standard deviations higher on achievement measures than students taught through conventional classroom instruction. He called it “The 2 Sigma Problem.”

The findings remain striking today, explained Jennifer Steele, Ph.D., a professor in the School of Education at American University. In education research, she explained, even two-tenths of a standard deviation is considered a meaningful effect. But she added that there are very practical challenges preventing students from accessing this educational supplement. 

“It’s very hard to afford a one-to-one tutor for every kid because teachers command professional salaries, and giving one to every kid, when you could give one to every 30 kids, is expensive,” she said. 

A peer-reviewed study conducted by the University of Wisconsin-Madison in 2024 found a correlation between socioeconomic status and access to tutoring. As a result, relatively few students have tutors. The University of Southern California’s Understanding America Study surveyed more than 1,600 households and found that only about 15% of students were receiving any tutoring at all. Fewer than 2% received tutoring that met even a moderate definition of “high quality.” Among students earning grades of C or lower — those who would likely benefit the most — fewer than 4% had access to high-quality tutoring.

How AI could help 

In an ideal world, AI could help close this gap by providing students with a tutor-like resource at home, available at any time. Advances in natural language processing mean AI tools can answer specific questions conversationally, explain complex concepts, and adapt responses in ways that resemble human tutors, often more efficiently than traditional search tools.

As AI models become more sophisticated, students can receive additional forms of support, such as uploading photos of their work or sharing live video with questions. These tools have also proven particularly effective for writing and coding, areas where providing detailed, line-by-line feedback requires significant time and expertise.

Many colleagues have told Michael Hilton, Ph.D., a professor in the School of Computer Science at Carnegie Mellon University and the associate department head for education, that office hours are dropping dramatically — likely because students are turning to AI tools for help. Because AI can efficiently answer simple questions, such as basic Python syntax, students can instead use office hours for higher-level, concept-focused learning.

Also: My top 5 free AI tools for school – and how they can help supercharge your learning

“It’s much more of this ‘teach them how to fish’ paradigm in office hours, which, thankfully, we could do more of because the office hours are less busy,” said Hinton. “A lot of the kind of lower hanging fruit that there’s just a simple answer to, they can get that answer easily, and that’s okay — they don’t need TAs for simple syntax questions or questions that are easily answered by a tool correctly most of the time.”

AI tools that function like a tutor could also help students who don’t have the resources to access a human tutor. A recent Brookings Institution report found that the largest barrier to scaling effective tutoring programs is cost, estimating a requirement $1,000 to $3,000 per student annually for high-impact models.

Because private tutoring often requires financial investment, it can drive disparities in educational achievement. Aly Murray experienced those disparities firsthand. Raised by a single mother who immigrated to the US from Cuba, Murray grew up as a low-income student and later recognized how transformative access to a human tutor could have been. That experience led her to found Upchieve, an edtech nonprofit that provides free, 24/7 online tutoring and college counseling to low-income middle and high school students nationwide.

Human tutors still play a role 

While the Upchieve platform incorporates AI in limited ways, including real-time chat moderation, session summaries shared with teachers, and progress reports designed to boost student confidence, Murray says the organization is not interested in replacing human tutors.

“AI does have lots of exciting applications to education. It’s going to help us accomplish our mission,” said Murray. “However, humans are really great for tutoring specifically, so I think tutoring is actually not one of the applications that I think we should be using AI for.” 

That view is supported by Upchieve’s own research. In partnership with Microsoft and the Gates Foundation, the nonprofit developed and tested UPbot, an AI tutoring chatbot trained on more than 70,000 human math tutoring transcripts and open educational resources. When the tool was introduced to 3,200 students, engagement remained low.

Only one in five students tried the AI chatbot, and just 3% of tutoring sessions were AI-only. By comparison, 92% of sessions involved only human tutors, suggesting a strong student preference for human interaction. When asked about why the students preferred humans, they cited the value of human connection. 

Also: The best free AI courses and certificates for upskilling in 2025 – and I’ve tried them all

Some of this preference reflects how people learn. Pedagogy, the methods and practice of teaching, relies on social interaction, something Rachel Slama, Ed.D., associate director at the Future of Learning Lab at Cornell University, says AI still struggles to replicate. 

“I think another piece is, like, it’s not a warm human,” said Slama. “Learning is a social process.” 

Upchieve’s data only sheds light on students’ tutoring experience, rather than outcomes after the fact, like whether AI tutoring meaningfully improves grades and test scores. It’s still unclear whether academic performance changes after regular AI tutoring.

Limitations – and risks 

The efficacy of these AI tutor interactions was also in question. In the Upchieve study, there was no statistically significant difference in learning or confidence between sessions that were human-only and those that were AI-only, raising concerns about AI’s ability to close gaps on its own.

While Murray acknowledges that AI can help low-income students, she cautions that it may also widen existing inequalities.

“I think it is possible that it could widen the gap because if we find that it’s really the most affluent students who are really good at taking advantage of an incredibly powerful tool like ChatGPT, then they could be accelerating faster than low-income students,” she said. 

This challenge is not unique to AI. Similar disparities in adoption have accompanied nearly every major technological shift. One of the clearest examples is broadband access: A study using City Health Dashboard and American Community Survey data found that high-income neighborhoods had the highest broadband access rate (87.2%) while low-income neighborhoods had the lowest (58.8%).

Also: Some teachers are using AI to grade their students, Anthropic finds – why that matters

“Really good learners are going to surge ahead with AI and speed up their learning and production,” said Slama. “And those who don’t know how to ask good questions, or don’t know how to sort of persist and regulate their own learning, are going to fall behind.” 

Incremental progress 

Despite these concerns, many experts, including Slama, Mollick, Steele, and Murray, agree that while individualized human tutoring remains the gold standard, the more relevant question is whether AI tutoring is better than a student having no support at all.

“I think it’s important that if I could put a chatbot that looks like my teachers into a community that doesn’t have the resources, doesn’t have the experience, the knowledge, etc. — I fundamentally believe that is what we’re meant to be doing for education,” added Campbell.

Edtech companies developing online learning tools are grappling with these same questions. McGraw Hill, founded in 1888, has evolved from a traditional publisher into a provider of educational content, software, and services, and has folded in AI. The technology has been carefully implemented by having teams of educators and psychologists develop a learning experience first, and then use AI only as the technological support to achieve it, explained Dylan Arena, Ph.D., McGraw Hill’s chief data science & AI officer. Concerns about widening disparities are central to that process. 

“If we could produce an AI tutor for the poor kids and then let the rich kids have human tutors, is that an equitable solution? At what point do we say, it’s good enough, it’s better than not doing anything at all, so let’s give it to them. And at what point do we say, wait, wait, no, don’t be satisfied with that. Demand better,” said Arena. 

Also: AI agents have arrived in US classrooms

He added that the real value lies in keeping a human in the loop — one that prioritizes human relationships while using AI to augment human “excellence and performance.” Ideally, he said, this would allow a high-quality human tutor to support 10 students instead of three with the help of AI tools.

Khan Academy, the free online nonprofit learning platform, has taken a similar approach by integrating AI into its offerings through tools such as Khanmigo, its AI-powered tutor. The tool was designed not to simply provide answers, but to scaffold learning and teach students how to use AI productively, according to Kristen DiCerbo, Ph.D., Chief Learning Officer at Khan Academy. 

Deployment results have been positive, she said, particularly for English language learners (ELL), highlighting AI’s ability to help close gaps. These students reported feeling more comfortable asking follow-up questions and requesting clarification, including in their home languages. DiCerbo also sees classroom deployment as a way to address access disparities.

“Part of our mission is to make sure that kids from all kinds of backgrounds are getting this into their hands and kids, where kids are in classrooms,” said Dicerbo. “If we can get these tools into classrooms, that’s where we can start making sure that all students have access to them and they’re able to make the best use of them.” 

Also: ChatGPT’s study mode could be your next tutor – and it’s free

There are also learning tools from nearly every major AI chatbot that are easy to access and understand, focused on the same principle of not just handing over answers but rather using the Socratic method to encourage students to find the answer themselves. Some examples include ChatGPT’s Study Mode, Google Gemini’s Guided Learning, and Anthropic Claude AI’s Learning Mode. 

“All the AI companies now often offer, for free, a learning mode that makes the AI less likely to give you the answer and more likely to act like a normal tutor,” said Mollick. “Those things will result in better outcomes.”