For decades, the promise of personalized education has been a kind of holy grail — always shimmering on the horizon, never quite within reach. Adaptive learning software tried. Massive open online courses tried harder. Neither cracked the fundamental problem: scaling one-on-one instruction without losing the back-and-forth dialogue that makes human tutoring so effective.
Now, almost without anyone formally announcing it, the two dominant AI chatbots are morphing into something their creators may not have originally intended. ChatGPT and Claude aren’t just answering questions anymore. They’re teaching.
The shift has been gradual but unmistakable. OpenAI and Anthropic, the companies behind ChatGPT and Claude respectively, have both rolled out features in recent months that move their products well beyond simple query-and-response interfaces. As 9to5Mac reported, these AI systems are evolving from chatbots into interactive learning tools — complete with Socratic questioning, step-by-step problem breakdowns, and the ability to adapt their explanations on the fly when a user signals confusion.
This isn’t a minor product update. It represents a fundamental rethinking of what conversational AI is for.
Consider how ChatGPT now handles a calculus problem. Rather than simply spitting out the derivative of a function, OpenAI’s system can be prompted — or in some modes, defaults — to walk a student through the chain rule, pause to check understanding, and offer an alternative explanation using a visual metaphor if the first attempt doesn’t land. Claude, Anthropic’s model, takes a similar approach but with what users describe as a more patient, almost professorial tone. It asks follow-up questions. It refuses to just hand over the answer when it detects that a student is trying to shortcut the learning process.
The pedagogical implications are enormous. And the commercial ones might be even bigger.
Education technology is a market projected to exceed $400 billion globally by 2027, according to estimates from HolonIQ. The incumbents — companies like Chegg, Coursera, Khan Academy, and Duolingo — have watched nervously as generative AI tools have eaten into their core value propositions. Chegg’s stock cratered in 2023 after its CEO acknowledged that ChatGPT was hurting customer growth, and the company has never fully recovered. Khan Academy, rather than fighting the tide, partnered with OpenAI to build Khanmigo, an AI-powered tutoring assistant that runs on GPT-4. That partnership now looks prescient.
But here’s the thing that should worry every edtech company on the planet: you don’t need Khanmigo to get tutoring from ChatGPT. The base product already does it. So does Claude. For free, or close to it.
The technical underpinnings of this shift are worth examining. Both OpenAI and Anthropic have invested heavily in what researchers call “instruction following” — the ability of a model to adhere to complex, multi-step directives about how to behave. Early large language models were notoriously difficult to steer. They’d answer a question about photosynthesis with a paragraph that mixed accurate science with confident hallucination, and they had no sense of whether the person asking was a third-grader or a PhD candidate.
That’s changed. Dramatically.
Modern versions of GPT-4 and Claude 3.5 can maintain what amounts to a mental model of the learner throughout a conversation. They track what’s been explained, what hasn’t clicked, and what level of vocabulary seems appropriate. They do this not through any genuine understanding of the student as a person, but through sophisticated pattern matching across the conversation’s context window — the rolling memory of text that the model can reference at any given moment. OpenAI’s latest models support context windows exceeding 128,000 tokens, roughly equivalent to a 300-page book. Claude’s context window stretches even further, to 200,000 tokens in its most capable configurations.
This means a student can have a sustained, multi-hour tutoring session — covering, say, an entire chapter of organic chemistry — and the AI will remember what was discussed at the beginning. It won’t repeat itself. It won’t contradict earlier explanations. It will build on them.
The experience, according to students and educators who’ve tested these systems extensively, is uncanny. Not perfect. But uncanny.
“It’s the most patient tutor you’ll ever have,” said Sal Khan, founder of Khan Academy, in a recent interview discussing AI’s role in education. “It never gets frustrated. It never checks its watch.” Khan has been one of the most vocal advocates for AI tutoring, arguing that it can democratize access to the kind of individualized instruction that has historically been available only to families wealthy enough to hire private tutors. His organization’s Khanmigo tool, built atop OpenAI’s technology, now serves millions of students — but Khan has acknowledged that the underlying ChatGPT product is itself becoming a formidable learning tool even without Khan Academy’s curriculum layered on top.
Anthropic has taken a slightly different philosophical approach with Claude. The company, founded by former OpenAI researchers Dario and Daniela Amodei, has emphasized what it calls “constitutional AI” — a framework in which the model is trained to be helpful, harmless, and honest. In an educational context, this translates to a system that’s notably cautious about giving wrong answers and more willing to say “I’m not sure” than its competitors. Teachers who’ve used Claude in classroom settings report that it’s less likely to hallucinate scientific facts, though it can still stumble on niche or highly specialized topics.
The competitive dynamics between OpenAI and Anthropic are intensifying. Both companies have released features in 2025 and early 2026 that seem specifically designed for educational use cases. OpenAI introduced a “tutor mode” in ChatGPT that can be toggled on by educators or students, which adjusts the model’s behavior to prioritize guided discovery over direct answers. Anthropic, as noted by 9to5Mac, has added similar capabilities to Claude, including the ability to generate practice problems, provide hints without giving away solutions, and create personalized study plans based on a student’s demonstrated knowledge gaps.
Neither company is calling itself an education company. Not yet. But the features tell a story that the press releases don’t.
There are skeptics, and their concerns deserve serious consideration. Dr. Kentaro Toyama, a professor at the University of Michigan’s School of Information who has studied technology’s impact on education in developing countries, has argued that AI tutors risk creating a dependency that undermines genuine learning. “The question isn’t whether the AI can explain something clearly,” Toyama has said. “The question is whether the student retains it, can apply it in novel situations, and develops the metacognitive skills to learn independently.” His concern: that students who grow accustomed to having an infinitely patient, always-available AI tutor may never develop the tolerance for confusion and frustration that real intellectual growth requires.
It’s a fair point. And there’s limited longitudinal data to resolve it.
What data does exist is mixed but cautiously encouraging. A study published in late 2025 by researchers at Harvard’s Graduate School of Education found that students who used AI tutoring tools for supplementary math instruction improved their test scores by an average of 12% compared to a control group — roughly equivalent to the gains seen with human tutoring in prior meta-analyses. But the study also found that students who used AI tools as their primary source of instruction, without any human teacher involvement, showed smaller gains and reported lower confidence in their understanding of the material.
The takeaway, according to the study’s authors: AI tutoring works best as a complement to human instruction, not a replacement for it. This aligns with what most educators intuitively believe, but it’s useful to have empirical backing.
The business implications ripple outward. Chegg, which once charged students $15 to $20 per month for homework help and textbook solutions, has been forced to reinvent itself as an AI-first platform. The company integrated GPT-4 into its product and rebranded its core offering, but user growth has remained sluggish. The problem is straightforward: why pay Chegg when ChatGPT does the same thing, often better, for the price of a $20 monthly subscription that also handles dozens of other tasks?
Coursera and Udemy face a different but related threat. Their value proposition has always been structured courses taught by credentialed instructors. But as AI tutors become more capable, the line between a structured course and a long, well-guided conversation with an AI begins to blur. A student can now ask Claude to design a six-week curriculum for learning Python, complete with reading assignments, coding exercises, and weekly assessments. Claude will do it. And it will adapt the curriculum in real time based on the student’s progress.
That’s not a course. But it’s not not a course, either.
Google, which has its own large language model in Gemini, has been characteristically aggressive in pursuing the education market. The company announced in February 2026 that Gemini would be integrated into Google Classroom, giving teachers the ability to generate quizzes, provide automated feedback on student writing, and create differentiated assignments for students at varying skill levels. The integration is still rolling out, but early reports from pilot schools suggest that teachers are finding it genuinely useful — particularly for the administrative burden of grading and providing individualized feedback, tasks that consume enormous amounts of teacher time.
Apple, notably, has been quieter on this front, though the company’s devices remain dominant in K-12 classrooms across the United States. The integration of AI tutoring capabilities into iPads and MacBooks — whether through first-party tools or partnerships with OpenAI and Anthropic — seems inevitable.
And then there’s the question of equity.
If AI tutoring really works — if it can deliver even a fraction of the benefit of one-on-one human tutoring — then its availability at low or no cost represents one of the most significant educational developments in a generation. The research on human tutoring is unambiguous: Benjamin Bloom’s famous 1984 study found that students who received one-on-one tutoring performed two standard deviations better than students in conventional classrooms. The “two sigma problem,” as it came to be known, has haunted education researchers for forty years. The problem was never about whether tutoring works. It was about how to make it available to everyone.
AI might be the answer. Or at least part of it.
But access isn’t uniform. Students in well-resourced schools and affluent families are far more likely to have access to the paid tiers of ChatGPT and Claude, which offer the most capable models and longest context windows. Students in under-resourced communities may be limited to free tiers with shorter context windows, slower response times, and less capable models. The digital divide, in other words, could reassert itself in a new form.
OpenAI has made some moves to address this. The company announced a partnership with several large urban school districts in 2025 to provide free ChatGPT Plus access to students and teachers, funded in part by philanthropic donations. Anthropic has a similar program, though smaller in scale. Whether these programs can reach the students who need them most — those in rural areas, those without reliable internet access, those whose schools lack the IT infrastructure to support AI tools — remains an open question.
There’s also the matter of accuracy. Large language models still hallucinate. They still get things wrong, sometimes confidently and persuasively. In a tutoring context, this is particularly dangerous, because a student may not have the background knowledge to recognize when the AI has made an error. A wrong answer from a human tutor can be corrected in the next session. A wrong answer from an AI, absorbed uncritically by a student studying alone at midnight, can calcify into a persistent misconception.
Both OpenAI and Anthropic have made significant progress on reducing hallucination rates, but neither company claims to have eliminated the problem. Anthropic’s approach of training Claude to express uncertainty more readily may offer a partial solution — a model that says “I’m not confident about this” is arguably safer than one that states incorrect information with authority. But students don’t always read the fine print. They don’t always notice the hedging.
So where does this leave us?
The trajectory is clear, even if the destination isn’t. ChatGPT and Claude are becoming educational tools whether their makers planned it that way or not. Students are already using them this way — millions of them, every day, across every subject and every level from middle school to medical school. The question for OpenAI, Anthropic, educators, and policymakers isn’t whether AI tutoring will happen. It’s already happening. The question is how to make it happen well.
That means investing in research on AI-assisted learning outcomes. It means developing guardrails that prevent students from using these tools to avoid learning rather than to enhance it. It means ensuring equitable access. And it means being honest about what these tools can and cannot do — honest in a way that the technology industry has not always been inclined to be.
The most patient tutor in the world is now available to anyone with an internet connection. That sentence would have sounded like science fiction five years ago. Today it’s simply a description of reality. What we do with that reality — how schools adapt, how companies compete, how students learn — will be one of the defining stories of this decade.
And we’re still only in the early chapters.
WebProNews is an iEntry Publication