A Harvard physics class ran an experiment in 2023. Half the students learned from an AI tutor at home. Half sat through one of the university’s best active-learning sessions in person.
The AI group learned more than twice as much. In less time. Published in Scientific Reports in 2025, the Kestin et al. study offered rigorous proof that a well-designed AI tutor can beat a strong human classroom on raw content acquisition.
That result now shapes how millions of students prepare for exams. Cambridge International qualifications — sat by close to a million learners a year — expose exactly where the AI advantage runs out.
Knowing the Content Isn’t the Same as Answering the Examiner
Cambridge mark schemes don’t reward comprehension. They reward precise, structured responses to specific command words: describe, evaluate, justify, compare. A student can understand a topic deeply and still lose marks for writing the wrong kind of answer.
That’s a pattern-recognition problem. Pattern recognition is exactly what large language models do well. McKinsey research puts a number on the payoff: teachers can redirect 20% to 40% of their time away from routine content delivery when AI handles adaptive instruction. That frees human expertise for the part of teaching AI still can’t touch.
Where AI Genuinely Helps
AI tutoring tools do three things well in exam prep. They flag weak topics from past-paper performance. They generate unlimited variations of a question type. They give instant feedback on factual accuracy.
Math-focused tools built for step-by-step problem solving show this clearly — the kind of scan-and-solve, show-your-work approach reviewed recently around tools like Solvely AI works well for building fluency fast.
Where It Doesn’t
An AI tool can confirm a photosynthesis answer is factually correct. It can’t reliably judge whether that answer earns 4 marks or 6 under the exact wording of a June 2026 mark scheme. That judgment depends on examiner conventions that shift between syllabus updates — conventions no training dataset fully captures.
The Harvard study is a useful caution here, not just a headline number. Kestin’s AI tutor worked because his team built it with expert-authored scaffolding for one specific course. Nobody deployed a generic chatbot and hoped. Off-the-shelf AI tutoring, without that syllabus-specific engineering, doesn’t reproduce the result.
What This Means for Cambridge Preparation
For a qualification built entirely around board-specific assessment objectives, generic AI tutoring hits a ceiling fast. A student running an Economics essay through ChatGPT gets grammatically sound feedback. It has no idea what a CAIE examiner report from the last series flagged as the gap between a Level 3 and a Level 4 evaluation.
This is where specialist human input still does work AI can’t replicate. Cambridge online tutors coach students on mark scheme logic directly — pairing AI-style repetition and instant feedback with the examiner-specific judgment the research shows machines still lack. The most effective prep right now isn’t AI versus tutor. It’s AI for volume, human for precision.
Students serious about closing that gap should start with the primary source. The Cambridge programmes and qualifications portal publishes the syllabus documents, specimen papers, and Principal Examiner Reports that define what a mark scheme actually looks for — the material any AI tool or tutor should build around, not replace.
The Trust Paradox Nobody’s Resolving Yet
Two-thirds of higher education institutions worldwide now have, or are developing, formal guidance on AI use, according to UNESCO. Not because AI fails. Because it works unevenly, and unsupervised use produces uneven results.
Pew Research data on how many teens now reach for chatbots during schoolwork puts a real number behind that unease — and shows why institutions moved from banning AI outright to writing policy around it instead.
That’s the paradox worth sitting with. The technology proven to double learning gains in a controlled Harvard trial is the same technology education leaders are moving to regulate, not remove. For Cambridge students, the takeaway isn’t which side wins. The students pulling ahead right now use AI for what it measurably does well, and human expertise for what it still doesn’t.
Related: What Tasks Is Generative AI Actually Good For? A Practical Guide
