An English teacher in Sheffield marks 120 essays a week. By Wednesday evening she is skimming. Not because she doesn’t care, but because the third stack of the night blurs into the fourth. That specific exhaustion is where artificial intelligence in education has found its most honest use case so far: not replacing her judgement, but buying back the hours she needs in order to apply it.
Strip away the conference keynotes and the vendor promises, and something more modest is happening in real classrooms. AI is showing up in three places: practice software that adjusts to each student within seconds, tools that give feedback before a teacher ever sees the work, and accessibility features that make lessons usable for children who were previously left behind.
Where AI Has Genuinely Earned Its Place
Practice that adjusts in real time
Adaptive platforms such as Carnegie Learning’s MATHia and DreamBox look at how a student solves a problem, not just whether the final answer is right. Misread a negative sign twice and the next question isolates that skill. Get five in a row and the system moves on instead of assigning twenty more of the same. In a class of thirty, that kind of tuning can only happen for one student at a time, which is precisely why it never happened before.
Feedback on the first draft
Second-language learners and first-year writers benefit most. A tool that flags a missing counter-argument, or a paragraph with no evidence behind it, gives a student something to act on at 9pm when no teacher is awake. The teacher still marks the final version, but the draft landing on her desk has already been through two rounds of repair.
Access rather than acceleration
Roughly five million students in US public schools are English learners, and live translation has changed what a science lesson looks like for them. Captions, speech-to-text and text-to-speech are the least glamorous applications of AI in education, and arguably the ones with the largest effect on who gets to participate.
The Tutoring Gap Nobody Could Afford to Close
In 1984, educational psychologist Benjamin Bloom published a finding that has haunted schools ever since. Students taught one-to-one performed about two standard deviations above students taught in a conventional classroom. That is the gap between a middling pupil and one in the top few percent of the class. The catch was always cost. Private tutoring runs from $40 to $100 an hour in most Western cities, which puts it out of reach for exactly the families who need it most.
AI tutoring does not reproduce Bloom’s result. These tools still lose the thread of an argument, flatter weak answers, and now and then invent a formula. But they are available at 11pm, in any language, for a fraction of the price, and that is enough to take them seriously. Georgia Tech’s Jill Watson, an assistant that answered student questions in an online course back in 2016, was a curiosity. Khan Academy’s Khanmigo and similar assistants now run in thousands of schools daily. If you want a grounded view of which deployments are holding up, this piece on AI in education and what’s actually working in classrooms right now is a sensible starting point.
Universities Are Rebuilding the Curriculum Around It
Higher education moved faster than most people expected, though not in the direction predicted. Blanket coursework bans in 2023 gave way to redesign. Duke built one programme specifically for students who will never train a model but will manage teams that use them, and the Duke AI for Everyone course structure shows how that plays out across a semester. The University of Helsinki went the other way, releasing a free online course that has drawn well over a million enrolments worldwide; Elements of AI remains one of the cheapest ways to get genuinely literate in the subject.
Working professionals are a separate market, and a more expensive one. Executives and mid-career engineers are enrolling in short certificate programmes on the assumption that AI fluency will matter at promotion time. Whether the price is justified depends on the provider. This breakdown of the Oxford artificial intelligence programme covers the real curriculum and time commitment, and there is a similarly blunt review of MIT Professional Education for anyone weighing cost against career return.
The companies that once sold study help are feeling the shift. Chegg, which built a business on textbook solutions and on-demand expert answers, has announced successive rounds of layoffs since 2023 as subscribers drifted toward free chatbots. Pearson has restructured repeatedly. The demand for explanation never disappeared. It moved.
What Still Goes Wrong
The confident wrong answer
Language models produce fluent nonsense, and fluency is the problem. A student revising for a history exam has no easy way to separate a real citation from a fabricated one. In 2023 a New York lawyer was sanctioned for filing a brief containing cases invented by ChatGPT. That failure mode scales down to a fourteen-year-old’s homework perfectly well.
Detectors that accuse the innocent
AI detectors carry a false-positive problem that nobody has solved. Stanford researchers found in 2023 that detectors flagged a large share of essays written by non-native English speakers as machine-generated, while human-written essays from native speakers mostly passed. Accusations built on a probabilistic guess have reached disciplinary panels, and a few have ended up in court.
Data, contracts and quiet collapses
Los Angeles Unified School District launched a district chatbot called Ed in early 2024 and pulled it within months after the vendor behind it ran into serious financial trouble. Student data is not a good thing to misplace. Any deployment worth the name needs a clear answer about where conversation logs live, how long they are kept, and which humans can read them.
Five Questions Worth Asking Before You Sign Anything
- What happens to the data? Where are student conversations stored, for how long, and can a school have them deleted?
- Who checks the output? A named person should review anything the AI sends directly to students without a teacher seeing it first.
- Does it reduce or add workload? Some tools save an hour a week. Others create a review job nobody budgeted for.
- Can it explain a mark? If a system cannot say why it judged an answer wrong, it can neither help a student learn nor help a teacher defend a grade.
- What is the plan for when it fails? Assume it will. Decide in advance what students and staff are told to do.
Where This Goes Next
Watch the national pilots rather than the product launches. Estonia announced a programme in early 2025 to put AI tools in front of 20,000 high school students and train 3,000 teachers to use them, treating the technology as public infrastructure instead of something each school buys on its own. Similar arrangements are being negotiated elsewhere, and they will set the expectations that individual schools inherit.
The more interesting question is not whether AI belongs in classrooms. It is what schools stop doing once it is there. When calculators arrived, arithmetic drills did not vanish, but the purpose of a maths lesson shifted toward modelling and reasoning. A comparable shift is coming for writing, research and assessment, and it will be uncomfortable for institutions that measure themselves by how much marking their staff still do by hand.
Teachers who use these tools well tend to describe the same outcome. Less time on the mechanical part of the job, more time with the students who need a human in the room. It is a modest claim, and it is the one most likely to outlast the current round of hype.