Your child’s essay comes back cleaner, the worksheet gets finished in half the time, and the chatbot did most of the thinking. The OECD’s Digital Education Outlook 2026, the organization’s flagship review of generative AI in classrooms, lands on an uncomfortable distinction: AI reliably improves the work students turn in, which is not the same as improving the student. In the sharpest study the report draws on, high schoolers using a standard AI chatbot lifted practice scores by 48 percent, then scored 17 percent worse than classmates who never touched it once the tool was taken away. So as kids head back to school with an AI helper one tab away, the question worth asking is not whether the homework got better. It is whether your child did.
The OECD’s Digital Education Outlook 2026 draws a line most coverage misses: AI that improves the homework is not AI that improves the learner. Here are the questions parents are asking.
Common questions
Should I let my child use ChatGPT for homework?
How do I tell whether an AI tool is helping my child learn?
Are AI tutors a replacement for human help?
Should schools ban AI tools?
A standard AI chatbot lifted practice scores 48%, then students scored 17% worse once it was gone. The OECD's 2026 outlook leaves parents one question: better homework, or a better learner?
What the OECD found
Released in January 2026, the Digital Education Outlook, subtitled Exploring Effective Uses of Generative AI in Education, gathers the emerging research on chatbots and AI tutors in schools. Its core finding splits the technology in two. General-purpose chatbots, the kind most students use at home, tend to hand over direct answers, a design the report links to shortcuts and offloaded thinking. Purpose-built educational tools sit at the other pole: in the report’s description, they incorporate curriculum alignment, learner models, tutoring logic and safeguards, and they showed learning gains that held up, strongest in collaborative settings.
The evidence behind that split is concrete. In a randomized experiment at a Turkish high school, Wharton researchers Hamsa Bastani, Osbert Bastani, Alp Sungu and colleagues gave close to a thousand students in grades 9 through 11 one of three setups for math practice: a ChatGPT-style interface, a version with tutoring safeguards that hints instead of answering, or no AI at all. During practice, the chatbot group improved 48 percent and the safeguarded tutor group improved 127 percent. Then came an exam with every tool closed. Students who had practiced with the standard chatbot scored 17 percent worse than students who never used AI. The safeguarded version largely erased that harm, though it produced no lasting advantage either. The authors’ verdict: students were using the chatbot as a crutch.
None of this reads as a case against the technology itself. Dialogue-based tutoring systems showed the strongest results of any category the report reviewed, and in one randomized trial it cites, students working with an AI tutor inside a carefully designed online course outperformed peers in comparable in-person classes. The dividing line ran through design, not through the presence of AI.
When we look only at performance metrics, AI interventions appear to have large positive effects. When we isolate genuine learning outcomes, the effects are small and sometimes negative.” (Dragan Gašević, learning scientist, on the report’s findings)
Laura Lurns · Learning Success expert
Task performance is not learning
Most coverage of AI in schools runs on one of two frames: the adoption race, where any district that hesitates is failing the future, and the cheating panic, where the technology is a plagiarism machine to be banned. The OECD’s evidence supports neither. Across the studies it reviews, the variable that decided outcomes was whether a tool produced performance on the task or thinking in the child, and those two outcomes came apart the moment the AI left the room.
Dragan Gašević, a learning scientist featured at the report’s launch, put the pattern plainly: “When we look only at performance metrics, AI interventions appear to have large positive effects. When we isolate genuine learning outcomes, the effects are small and sometimes negative.”
Gamified learning apps made the same promise for years: points and streaks that made children feel like they were learning while the skill quietly failed to arrive. Our longstanding position at Learning Success is that a points system makes a child feel like they are learning, while real mastery makes them feel capable. An answer-giving chatbot runs the same shallow loop with far more horsepower. And the struggle it removes was never the obstacle. Productive struggle, difficulty engaged with rather than avoided, is part of how learning happens, so work that gets easier without the child getting stronger deserves a second look. The underlying processing skills learning runs on, attention, memory, the speed and accuracy of taking information in, are built through practice a child does, not practice a model does for them.
Key takeaways
- Practice gains, exam losses: Chatbot users improved practice 48 percent, then scored 17 percent worse unassisted.
- Design decides the outcome: Tutors built with learning safeguards avoided the harm answer-giving chatbots produced.
- The unassisted test: Close the tool and watch what your child produces alone.
The kitchen-table test
The Turkish experiment hands parents a test that costs nothing: separate the practice from the proof. Let your child use whatever helper the school allows, then close the tool and ask them to work one similar problem, or explain the idea in their own words, on their own. That unassisted minute tells you more than any polished worksheet ever will.
The tradeoffs run in both directions, and honesty about them helps. A well-designed tutor that questions, hints and withholds the final answer showed genuine promise in the research the report gathers, including for families without access to human tutoring. A general-purpose chatbot in homework mode showed the opposite. The useful move is not banning the technology or embracing all of it; it is sorting tools by design. Ask your child’s school which tools the classroom uses, whether they were built with learning safeguards, and how anyone would notice if assisted work stopped translating into unassisted skill.
AI is arriving in classrooms either way, and state guidance documents are being written this school year. Parents who ask about unassisted performance are asking the question the OECD’s evidence says matters most, and the children whose thinking stays their own will be the ones this technology genuinely served.
Nobody notices whether a child is getting stronger sooner than the parent watching them work. The obstacle here is an idea: the assumption that better output means better learning, baked into tools designed to finish tasks rather than build children. Keep the thinking in your child’s hands, and make sure the foundations that thinking runs on are solid. The Learning Success All Access membership coaches you through short daily sessions that build those underlying skills, with your encouragement, not a points system, doing the motivating.
See what All Access gives your child
See MTSS for schools
Is your child struggling in school?
Get your free personalized learning roadmap
You describe what you see at home. We turn it into a plan you start this week.
- Answer 5 short questionnaires about what you already notice, 30–45 minutes at your own kitchen table
- Your child sits no test and gets no score: nothing to schedule, nothing for them to dread
- You do the answering, the AI does the writing, and a person reviews it before it reaches you
- Access all 40+ courses instantly: reading, math, focus, processing and more, with new ones added regularly
Why we use AI, plainly: it writes from a knowledge base our team maintains and audits. We work through it line by line and pull anything the evidence stops supporting. The roadmap you get on Tuesday reflects what we corrected on Monday, and a human still reads it before you do.
Your school district must evaluate your child free of charge if you ask in writing, whatever your income and whatever the outcome (US, 34 CFR 300.111 and 300.301(b)). That route takes time and answers a different question than you do. This one starts today, from what you already know.
Your answers stay yours. We do not sell your personal information, and we do not hand identifiable assessment data to outside AI companies to train their models.
A screener is a starting point, not a diagnosis. If your child might need formal accommodations (an IEP or 504 plan), or you suspect a vision, hearing or medical cause, pursue a professional evaluation too. That is the only route to those supports.
References
- OECD - Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education
- Bastani, Bastani, Sungu, Ge, Kabakci & Mariman (2024) - the randomized field experiment behind the practice and exam figures (SSRN 4895486)
- CIDDL - summary of the OECD Digital Education Outlook 2026
- Reda Sadki - coverage of the report's launch discussion
