Research · Education · Practice

An AI assistant can help someone solve a problem correctly in seconds. But does that mean the person has learned how to solve it? A set of three randomized experiments involving 1,222 US-based adults suggests that the answer depends on what the AI does for the learner—and what happens when the help disappears.

The paper, “AI Assistance Reduces Persistence and Hurts Independent Performance” by Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker and Rachit Dubey, was first submitted to arXiv on 6 April 2026 and last revised on 3 October 2026. It is a research preprint, so its findings should be treated as important but not yet as a settled verdict about AI-assisted education.

The central distinction: performance is not the same as learning

When a learner uses AI, the immediate task score can improve because the system supplies an explanation, a hint or even the answer. That score alone cannot tell us whether the learner has built knowledge they can retrieve independently later. To test this distinction, the researchers measured performance both while AI was available and after it was removed.

The paper reports three randomized experiments: two using fraction problems and one using reading-comprehension questions. In total, 1,222 people were recruited through the online research platform Prolific. Participants were adults in the United States—not a sample of schoolchildren—and each task session lasted roughly 10–15 minutes.

What the experiments found

In the first experiment, participants worked on fraction problems with or without an AI assistant. The AI condition used GPT-5 in a sidebar, with the system pre-prompted with each problem and its solution so that participants could request direct answers. After the practice phase, the assistant was removed and participants attempted three further problems independently.

• While AI was available: participants with AI solved more problems correctly and skipped fewer questions.
• After AI was removed: the analysed AI group solved a lower share of the final fraction problems than the control group—57% versus 73%. The difference was statistically significant, with a reported Cohen’s d of −0.42.
• Persistence also mattered: in that first experiment, the AI group skipped more of the final questions—20% versus 11%. This difference was statistically significant, though the effect was smaller than the performance difference.

The researchers then repeated the fraction experiment with methodological changes intended to address a potential confound in the first experiment. The independent solve rate was again lower in the AI condition: 71% versus 77%, a smaller but statistically significant difference (Cohen’s d = −0.19). The overall difference in skipping questions was not statistically significant in this second experiment.

A third experiment tested reading comprehension. After the AI assistant was removed, the analysed AI group answered fewer of the final questions correctly than the control group—76% versus 89%—and skipped more questions. Both differences were statistically significant in this experiment.

These are not proof that AI permanently damages skills. They are evidence that, in these short experimental tasks, access to a system that readily supplies answers was followed by weaker performance on a small set of unaided questions.

Why the type of AI help may matter

One of the paper’s most educationally relevant analyses examined how participants used the assistant. In the second fraction experiment, 61% of AI-condition participants said they mainly used it to obtain direct answers, 27% said they mainly sought hints or clarifications, and 12% reported not using it. The authors found that poorer independent performance and some of the persistence differences were concentrated among those who sought direct solutions.

This comparison needs careful interpretation: participants were not randomly assigned to “answer-seeking” or “hint-seeking” styles, so this part of the analysis cannot establish that one prompt style caused the difference. People who chose different styles may also have differed in other ways. Nevertheless, it offers a useful hypothesis for educators to test: AI that completes the cognitive work may have different consequences from AI that prompts learners to do the work themselves.

How strong is the evidence—and what are its limits?

The study has several strengths. It uses randomized comparisons, includes a replication of the fraction task with improved controls, and tests a second domain—reading comprehension. The findings therefore go beyond a survey asking people whether they feel more or less capable after using AI.

But the boundaries of the evidence are just as important:

• Short exposure and short follow-up: the experiments measured performance immediately after brief AI-assisted sessions. They do not establish whether any difference persists over weeks or months.
• Adult online participants: the findings cannot be transferred directly to children, teenagers, classroom teaching or every educational subject.
• A specific form of assistance: the AI was configured to make direct answers readily available. This is not a test of every AI tutor, guided practice system or classroom implementation.
• Small final test sets: the unaided outcome was measured using only a few questions after the assistance phase, so the size and generality of the effect require further study.
• Not every result was consistent: the overall persistence difference was not statistically significant in the second fraction experiment, even though the performance difference was.

The right conclusion is therefore narrower than “AI makes students less intelligent.” The study suggests a risk: when a learner repeatedly receives solutions instead of practising the reasoning, immediate success may not translate into independent performance once the support is gone. A language-learning example is explored in our guide to AI vocabulary learning and independent word use.

How this fits with the wider research

The result is consistent with a more directly educational field experiment published in Proceedings of the National Academy of Sciences in 2025. In a study involving nearly 1,000 high-school mathematics students, Bastani and colleagues compared an unrestricted GPT-4 interface with a tutor designed with learning safeguards. Access to AI improved performance during practice, but students using the unrestricted version performed worse on a subsequent unaided test than students who had not received that assistance. The scaffolded tutor produced a different pattern, underscoring that the design of the help matters. Read the PNAS study.

The OECD’s Digital Education Outlook 2026 makes a related point: general-purpose generative AI can improve task performance without necessarily producing learning gains when it is used without pedagogical guidance. Taken together, these sources do not show that AI is inherently harmful. They point to a more practical question: is the tool designed to finish the task, or to help the learner become able to do it?

For a broader overview of cognitive offloading and critical thinking, see Prowad’s earlier analysis of AI and critical thinking. For classroom routines that protect retrieval and independent recall, see Prowad’s guide to AI and student memory. For the product-design side of this question, read our analysis of AI tutors and study agents. This article focuses more narrowly on the evidence from randomized tests of what happens after assistance is removed.

What educators can do now

The evidence supports cautious design principles rather than a blanket ban on AI. Teachers and learning designers can test the following approaches:

1. Ask for an attempt before revealing the answer. Have learners write down a first step, prediction or explanation before asking AI for help.
2. Prefer graduated hints to complete solutions. Start with a question or small clue, and reveal more only when the learner is genuinely stuck.
3. Fade the assistance. After a worked example, give learners a similar problem to solve without AI. This checks whether they can transfer the method independently.
4. Make reasoning visible. Ask learners to explain why a step works, identify a possible error, or compare two solution paths—not simply submit the final answer.
5. Assess learning without the tool as well as with it. Use short retrieval questions, oral explanations, or a brief unaided task alongside AI-assisted work. The goal is to measure both supported performance and independent competence.
6. Evaluate the intervention, not just satisfaction. If a school introduces an AI tutor, compare learning outcomes and retention against an appropriate alternative, rather than relying only on usage, speed or student enthusiasm.

These are design implications informed by the evidence, not interventions all directly tested by the 1,222-person study. Their effectiveness should be evaluated in the relevant classroom, subject and age group.

The takeaway: design AI around the learner’s next independent step

The most useful question is not whether AI can produce a correct answer. It is whether the interaction leaves the learner better prepared to reason when the answer is no longer available. The new experiments provide a warning about immediate-answer assistance, while the wider evidence suggests that guided, pedagogically purposeful use may lead to different outcomes.

For schools and universities, this means judging educational AI not only by what students accomplish with it, but also by what they can explain, remember and do independently afterwards.

Sources and further reading

• Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv preprint, submitted 6 April 2026; version 5 revised 3 October 2026. https://arxiv.org/abs/2604.04721
• Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
• OECD (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. https://doi.org/10.1787/062a7394-en

Editorial note: The lead study is an arXiv preprint. Its latest version is dated 3 October 2026; the article does not treat it as a newly published journal paper or as definitive evidence of long-term effects.

As generative AI becomes part of everyday study, a practical question is becoming harder for schools and universities to avoid: how can educators tell whether students are genuinely AI-literate? A student may feel confident using ChatGPT, produce a polished assignment with it, and still struggle to verify a claim, recognize a fabricated reference, decide when AI assistance is inappropriate, or explain the reasoning behind a submitted answer.

This distinction matters because AI literacy is not simply familiarity with a chatbot or the ability to write an effective prompt. It includes understanding what a system can and cannot do, judging the quality of its output, making responsible decisions about its use, and retaining responsibility for learning. Recent research on performance-based testing and assessment-specific AI literacy suggests that education needs more precise ways to measure these capabilities.

This article reviews evidence on AI literacy assessment, considers what current measurement tools can and cannot tell us, and proposes practical principles for educators. It is an evidence-based synthesis of published research, not a report of a new experiment.

The measurement problem: confidence is not competence

Many AI literacy questionnaires ask learners to rate their own knowledge or confidence. Self-reports are useful: they can reveal attitudes, anxiety, perceived ability, and whether students believe they need support. But they do not directly demonstrate that a learner can perform a task successfully.

A systematic review in npj Science of Learning found that, although available AI literacy scales generally showed promising structural validity and internal consistency, relatively few had been tested for content validity, construct validity, responsiveness, or cross-cultural validity. The review also identified far fewer performance-based measures than self-report instruments. This points to a measurement gap, not proof that self-report tools are useless.

A later systematic review in Computers and Education: Artificial Intelligence examined 31 empirical studies using six AI literacy instruments across 14 countries. It found that AI literacy was associated with factors such as self-efficacy, attitudes, motivation, and digital competencies. It also reported differences between relationships measured with self-report scales and those measured with a performance-based test, raising the possibility that self-perceptions can distort how literacy is estimated.

The educational implication is straightforward: a survey can tell a teacher what students think they can do, but a performance task can provide evidence of what they actually do in a defined situation. Strong assessment systems should use each method for the questions it can answer. This distinction builds on the broader question explored in our guide to what AI literacy actually means and complements our analysis of the limits of AI detectors: neither confidence nor a detector score alone can establish what a student understands.

What performance-based assessment adds

One important contribution is the Generative AI Literacy Assessment Test (GLAT), reported in Computers and Education: Artificial Intelligence in December 2025. The researchers developed a 20-item multiple-choice test and evaluated it with 355 higher education students. They reported acceptable reliability and structural validity; GLAT scores also predicted performance on GenAI-supported tasks better than self-ratings such as perceived ChatGPT proficiency.

The finding supports a shift from asking only “How confident are you?” to also asking students to demonstrate knowledge and judgment. However, the evidence should be interpreted within its limits. The study involved a particular instrument and a sample of higher education students. A 20-item test cannot capture every real-world AI skill, and further validation is needed across age groups, languages, disciplines, and changing AI tools.

Performance-based assessment does not have to mean a long, standardized examination. A teacher might ask students to compare two AI-generated explanations, identify unsupported claims, verify a reference, explain why one answer is stronger, or revise an output while documenting what they changed and why. Such tasks make the underlying reasoning more visible than a confidence rating alone.

A second development: assessing AI literacy within academic assessment

A study published in Frontiers in Education on 10 September 2026 developed and validated the Generative AI Assessment Literacy Scale (GAA-LS) for higher education students. Rather than measuring general AI knowledge, it focuses on decisions students face when AI is used in assessed work.

The researchers began with an item pool and used expert review and pilot testing before analysing data from two student samples: 486 students in the first study and 798 in the second. The final 18-item instrument supported five dimensions:

• Assessment criteria awareness: understanding what the assignment requires and how work will be judged.
• AI-task appropriateness: deciding when AI assistance is suitable or unsuitable for a particular task.
• Verification and evidence checking: checking claims, references, and reasoning instead of accepting outputs at face value.
• Ethical attribution and academic integrity: following disclosure rules and taking responsibility for submitted work.
• Feedback uptake and revision: using AI-supported feedback critically to improve work rather than accepting it uncritically.

The authors reported satisfactory to strong internal consistency and evidence supporting the scale’s validity, including measurement invariance across gender, discipline, and AI-use-frequency groups in their samples. Scores were positively associated with feedback engagement and intention to maintain academic integrity.

These results are promising, but the distinction between association and causation is essential. The study used a cross-sectional scale-development design; it does not show that raising a student’s score will necessarily improve academic integrity or learning outcomes. The instrument is also focused on higher education and needs further testing across languages, cultures, institutions, and observable behaviours. Because it uses self-report items, it complements rather than replaces performance-based assessment.

What should schools actually measure?

Taken together, the research suggests that AI literacy assessment should cover more than prompt-writing. A practical framework for classroom use can focus on five observable capabilities.

1. Understand the tool and its limits

Can the student explain, in appropriate terms, that generative AI may produce plausible but inaccurate answers? Can they identify situations where a model’s output needs independent checking? The goal is not to require every learner to understand the technical details of a large language model, but to develop an accurate mental model of its strengths and limitations.

2. Evaluate evidence, not just fluency

Give students an AI-generated response containing a mix of accurate statements, unsupported claims, and a questionable reference. Ask them to verify selected claims using credible sources and explain which evidence changed their judgment. A fluent answer should not receive full credit simply because it sounds authoritative.

3. Make appropriate decisions about delegation

Students should be able to distinguish between using AI to brainstorm, request feedback, or practise a skill and handing over the central intellectual work that an assignment is intended to assess. The boundary depends on the learning objective and the teacher’s rules. Assessment instructions should therefore specify permitted, restricted, and prohibited uses instead of relying on vague expectations.

4. Disclose use and retain responsibility

Where disclosure is required, students should be able to describe how AI contributed to their work. They should also be able to defend the reasoning, sources, and final decisions in their submission. Disclosure alone does not prove learning, but it can make the process more transparent and support fairer assessment.

5. Use feedback to improve, not merely to finish

Ask learners to compare AI feedback with a rubric or teacher feedback, decide which suggestions to accept, and explain why they rejected or adapted others. This makes evaluative judgment visible and reinforces the idea that AI-generated feedback is a resource to assess—not an instruction that must automatically be followed.

A practical assessment design for teachers

A balanced approach can combine three kinds of evidence rather than relying on one score:

• Knowledge check: a short quiz on AI capabilities, limitations, privacy, bias, and responsible use.
• Authentic performance task: students critique an AI output, verify evidence, improve the answer, and explain their decisions.
• Process reflection: students record what they delegated, what they checked, what they changed, and what they learned.

A simple rubric might assess the accuracy of verification, the quality of reasoning, the appropriateness of AI use, transparency about assistance, and the student’s ability to explain the final work. The rubric should be shared before the task and aligned with the actual learning objective. It should not reward elaborate prompting if prompting is not the skill the course intends to teach.

This approach also helps avoid a false choice between banning AI and allowing unrestricted use. Different assignments serve different purposes. If the objective is unaided recall or foundational practice, AI may need to be restricted. If the objective is evaluating sources, revising arguments, or using tools responsibly, AI can be included explicitly in the assessment design.

Equity and validity: the cautions that matter

Assessment design must account for unequal access to paid tools, differences in language proficiency, disability accommodations, and students’ prior experience with technology. A test that inadvertently measures subscription access, English fluency, or familiarity with one interface may misrepresent the intended competence.

Tools also change quickly. A test tied to a particular product interface may become obsolete after an update. Durable assessment should focus on transferable judgment: checking evidence, recognizing uncertainty, selecting appropriate support, protecting private information, and explaining decisions.

Finally, no single score should be treated as a complete measure of a student’s AI literacy. Self-report questionnaires, knowledge tests, performance tasks, process evidence, and teacher observation each capture different aspects. Their combination is more informative than any one instrument used in isolation.

What educators and institutions should do next

For teachers, the immediate step is to define the capability a task is meant to assess, then design a short activity that makes that capability observable. For institutions, the priority is to provide consistent rules for acceptable AI use, teach students how to verify outputs, and evaluate assessment methods for fairness and validity.

For researchers, the agenda includes testing instruments across age groups and languages, comparing self-reports with actual performance, and studying whether assessment results predict learning over time. New scales should be evaluated not only for reliability but also for whether they measure the intended construct and work fairly across different student populations.

UNESCO’s guidance on generative AI in education and research provides a broader human-centred framework, emphasizing privacy, age-appropriate use, validation, and the development of educators’ and learners’ competencies. These principles support an approach in which assessment is designed around learning and responsibility rather than around the presence or absence of a particular tool.

Key takeaways

• AI literacy is broader than prompt-writing or familiarity with a chatbot; it includes evaluation, verification, appropriate use, ethics, and responsibility.
• Self-report questionnaires can reveal students’ perceptions, but they do not by themselves demonstrate practical competence.
• The GLAT study, involving 355 higher education students, found that a performance-based test predicted GenAI-supported task performance better than self-ratings in that study.
• A September 2026 study developed an 18-item scale focused on AI literacy in assessment contexts, covering criteria, appropriate use, verification, attribution, and feedback-based revision.
• Both instruments have limits and need further validation across contexts; neither should be treated as a universal measure of every AI-related capability.
• Teachers can begin with authentic tasks that require students to verify, justify, revise, and explain their use of AI.

Editorial conclusion: The central challenge for education is no longer only teaching students how to use AI. It is establishing credible evidence that they can use it thoughtfully, verify what it produces, and remain responsible for their learning. Assessment should make those capabilities visible. Confidence and polished output may be useful signals, but neither is a substitute for demonstrated judgment.

Sources and further reading

• Jin, Y., Martinez-Maldonado, R., Gašević, D., & Yan, L. (2025). GLAT: The generative AI literacy assessment test. Computers and Education: Artificial Intelligence, 9, 100436. https://doi.org/10.1016/j.caeai.2025.100436
• Nie, J., Zhang, Z., Lu, X., Zhang, Y., & Zhang, M. (2026). Development and validation of the generative AI assessment literacy scale for higher education students: psychometric evidence and associations with feedback engagement and academic integrity. Frontiers in Education, 11, 1934632. https://doi.org/10.3389/feduc.2026.1934632
• How AI literacy correlates with affective, behavioral, cognitive and contextual variables: A systematic review. (2025). Computers and Education: Artificial Intelligence, 9, 100493. https://doi.org/10.1016/j.caeai.2025.100493
• Lintner, T. (2024). A systematic review of AI literacy scales. npj Science of Learning. https://doi.org/10.1038/s41539-024-00264-4
• UNESCO. (2023). Guidance for generative AI in education and research. UNESCO guidance.

Editorial note: This article synthesizes published research available as of 10 October 2026. Findings from higher education samples should not be assumed to generalize automatically to school-age learners or every educational setting.

On October 7, 2026, a new independent assessment raised a difficult question for schools and families: can a chatbot designed to help teenagers learn also be trusted to protect them when conversations become risky? The Youth AI Safety Institute at Common Sense Media published an evaluation of OpenAI’s ChatGPT for Teens, reporting that some advertised safeguards did not work reliably in its tests. OpenAI disputed aspects of the testing. The disagreement matters because educational AI is moving beyond occasional homework help toward tools that may become part of students’ daily routines.

The issue is not whether AI can be useful for learning. It is whether the safety claims surrounding products used by minors are supported by evidence—and whether schools, parents, and policymakers have enough information to make informed decisions.

What changed: AI companies are designing for teenage learners

OpenAI introduced ChatGPT for Teens on August 18, 2026. The company describes it as an experience with age-appropriate protections and learning-oriented features, including Study Mode, step-by-step guidance, quizzes, reminders, and optional parental controls. In an October 7 update, it also announced additional learning and college-planning features for teenagers.

The educational rationale is understandable. A conversational system can explain a concept in different ways, offer practice questions, and help a student work through a problem when a teacher or family member is unavailable. These capabilities may widen access to support. But a product used by adolescents is not simply a tutoring tool: it is also a conversational environment that may encounter disclosures about distress, health, relationships, or personal safety.

That combination makes the reliability of safeguards a central educational question, rather than a secondary technical detail.

What the independent assessment found—and what it did not prove

In its October 7 risk assessment, Common Sense Media’s Youth AI Safety Institute said it tested more than 4,000 prompts before and after the teen experience was introduced. The organization reported that some protections worked, including refusals of explicit sexual role-play, while other measures were less reliable. Its concerns included missed crisis referrals, parental alerts that did not consistently trigger in its tests, the ease of moving from Study Mode to direct answers, and responses that could make the chatbot seem more person-like than intended.

The Institute rated the product an “Unacceptable Risk” for children and teenagers and recommended restricting access until the problems were addressed. That rating is the Institute’s conclusion, not a universal scientific consensus. The assessment is a structured product evaluation, not a randomized study of students’ learning outcomes or a peer-reviewed clinical trial. Its results identify concrete failure modes worth investigating, but they do not by themselves establish how often those failures occur across all users, settings, versions, or real-world conversations.

OpenAI has said the testing does not accurately represent how its safeguards work in practice. Reporting by the Associated Press also notes that the organization conducting the assessment receives funding from industry, including technology companies it evaluates. That funding relationship is relevant context for readers, but it does not automatically invalidate the results; the appropriate response is to examine the methods, disclose limitations, and seek independent replication.

The most defensible conclusion is therefore narrower than either “the chatbot is safe” or “the chatbot is unsafe in every circumstance”: the assessment raises credible questions about whether particular safeguards work as advertised, and those questions require transparent, repeatable testing.

Why this matters for learning—not only for online safety

1. Safety is part of educational quality. A learning tool should be judged not only by how clearly it explains algebra or summarizes a text, but also by how it responds when a student is vulnerable, confused, or seeking help beyond its competence. A failure in those moments can undermine trust in the entire learning environment.

2. Study support and academic shortcutting can coexist. OpenAI presents Study Mode as a way to encourage hints, guided reasoning, and knowledge checks rather than simply supplying answers. The Institute reported that some tested routes still allowed students to obtain direct answers. This does not show that guided learning is ineffective; it shows why claims about how a feature behaves should be tested in realistic conditions. Teachers should continue to design tasks that require explanation, reflection, and evidence of a student’s own reasoning.

3. Parental controls need clear limits. A setting that encourages Study Mode is not necessarily a hard restriction, and a safety notification is not a guarantee that every concerning conversation will be detected. Families and schools need plain-language descriptions of what controls do, what they do not do, what data is shared, and when a human must intervene.

4. Human relationships remain essential. Conversational AI can sound supportive, but it is not a teacher, counsellor, parent, or emergency service. Systems should make that distinction clear and should direct users toward appropriate real-world help when necessary. Educators should avoid positioning a chatbot as a substitute for trusted adults or professional support.

A useful framework already exists

UNESCO’s Guidance for Generative AI in Education and Research argues for a human-centred approach, including data privacy, age-appropriate use, validation of systems, and the development of educators’ and learners’ AI competencies. Published in 2023 and updated online in 2026, the guidance predates this particular product dispute, but its core principle is directly relevant: educational institutions should not assume that a widely available AI tool is pedagogically or ethically suitable without evaluating it.

For schools, this suggests a practical procurement and use checklist. Before recommending an AI tool to minors, ask:

• Has its safety performance been evaluated independently, using transparent methods and age-appropriate scenarios?
• Are the product’s age rules, privacy practices, parental controls, and escalation procedures explained clearly?
• Can teachers distinguish a learning-support feature from a control that actually prevents certain actions?
• Does the tool encourage students to reason, verify, and reflect—or does it routinely complete the task for them?
• Is there a clear human support pathway for sensitive or high-risk situations, and a process for reporting failures?
• Will the school review the tool again after updates, rather than treating a single assessment as permanent proof of safety?

What evidence should come next?

The next step should not be to settle the debate through competing headlines. AI companies should publish measurable, age-specific safety results and explain how tests translate into product changes. Independent evaluators should publish methods, limitations, and reproducible findings. Researchers should examine not only whether a system refuses dangerous requests, but also whether it recognizes when a student needs human support, whether controls behave as users understand them to behave, and how repeated use affects learning and help-seeking over time.

Schools, meanwhile, can take proportionate precautions without abandoning educational experimentation: use approved tools with clear policies, teach students to question AI outputs, avoid sharing sensitive personal information, keep teachers involved in assessment, and never treat a chatbot as the sole support for a student in crisis.

Key takeaways

• AI tutoring features may expand access to individualized explanations and practice, but usefulness does not automatically establish safety.
• A Common Sense Media evaluation published on October 7, 2026, reported weaknesses in some ChatGPT for Teens safeguards; OpenAI disputed aspects of the testing.
• The evaluation is important evidence about possible failure modes, but it is not a peer-reviewed study of student outcomes or proof that every interaction is unsafe.
• Schools should assess privacy, age-appropriateness, crisis pathways, parental controls, and learning design—not just product features.
• The most credible path forward is transparent company reporting, independent replication, and continued human responsibility for students’ wellbeing and learning.

Editorial conclusion: The emerging issue is not whether teenagers will encounter AI, but whether the systems marketed to support their learning can demonstrate that their protections work under realistic conditions. For education, responsible adoption requires evidence, clear limits, and human oversight—not confidence based on product claims alone.

Sources and further reading

• OpenAI. Introducing ChatGPT for Teens: Built for learning, backed by protections (August 18, 2026).
• Common Sense Media, Youth AI Safety Institute. ChatGPT for Teens Risk Assessment (updated October 7, 2026).
• Associated Press. Some guardrails on ChatGPT for Teens don’t work as promised, watchdog group says (October 2026).
• UNESCO. Guidance for Generative AI in Education and Research (2023; updated online in 2026).

The latest wave of educational AI is moving beyond the chatbot. OpenAI has just expanded ChatGPT for Teens with college-planning support, flashcards and quizzes, while Microsoft is positioning its Study and Learn Agent around guided practice rather than simply supplying answers. Anthropic is also building education-specific capabilities for teachers. The direction is clear: major AI companies are trying to turn general-purpose models into AI tutors for students.

But the important educational question is not whether these systems can explain a concept, generate a quiz or answer a student’s question. They can. The harder question is whether giving students an AI tutor actually produces more learning.

Recent evidence gives a much more cautious answer than the current technology race might suggest.

The trend: from AI answers to AI study agents

On 7 October 2026, OpenAI reported early usage data from ChatGPT for Teens and announced new learning-oriented features. Students can turn notes into flashcards and interactive quizzes, while a forthcoming College Planner is designed to help high-school students organize applications, deadlines and financial-aid tasks. OpenAI also says that, among teens using its learning tools, common requests include developing ideas, receiving feedback and checking evidence. OpenAI’s announcement.

Microsoft is pursuing a similar direction with its Study and Learn Agent. The company describes it as a learning-focused system that asks students to think before it tells them, using scaffolding, quizzes, flashcards and other activities. It is explicitly designed around the principle that the learner does the thinking. Microsoft’s documentation.

This is a significant shift in product design. The educational AI market is moving from the model of “ask the chatbot” toward “work with an AI tutor”: a system intended to question, prompt, scaffold, diagnose misconceptions and provide practice.

That shift is educationally important because the quality of an AI interaction depends not only on what the model knows, but also on what the student is required to do.

What does the research actually show?

The strongest reason for caution is that access to an AI tutor is not the same thing as effective tutoring.

A two-year cluster-randomized experiment in 18 Tennessee middle schools provides one of the clearest recent tests. Students using Khan Academy with the AI tutor Khanmigo achieved gains in mathematics of roughly 0.06–0.08 standard deviations over a school year. However, the researchers found that these gains were comparable to those produced by Khan Academy practice without AI. Although 96% of students tried Khanmigo at least once, the median student used it on only about one third of practice days and in only 17% of sessions in which they made a mistake. The researchers therefore identify student engagement with the tutor, rather than mere access, as a major constraint. Oreopoulos & Low, NBER Working Paper 35620 (2026).

Even more striking is a randomized university study published in October 2026. The study involved 2,379 undergraduate students and 30 instructors across multiple disciplines. Access to a course-integrated generative-AI tutor was associated with a 0.37 standard-deviation reduction in final grades and a 0.90-standard-deviation reduction in learning-management-system participation. The researchers suggest that the AI tutor may have displaced participation in existing learning activities rather than improving learning. The estimated losses were also larger among first-generation students. Liu et al., EdWorkingPapers (2026).

These two studies should not be treated as contradictory proof that AI tutoring is either good or bad. They show something more useful: the educational effect depends heavily on implementation, student behavior, task design and the surrounding course.

The wider evidence points in the same direction

A 2026 systematic review and meta-analysis of generative AI in STEM education examined 85 eligible studies, with 49 studies and 59 effect sizes included in the quantitative synthesis. It found an overall positive effect on cognitive learning outcomes, but the heterogeneity was extremely high (I² = 96.32%), with a very wide prediction interval. In other words, the average result hides substantial differences between contexts and interventions. Boolzen, Kuhn & Flegr (2026).

The OECD’s Digital Education Outlook 2026 reaches a similar conclusion from a policy and evidence perspective. General-purpose GenAI can improve the quality of students’ immediate outputs without producing equivalent learning gains when it removes the cognitive work that students need to perform themselves. By contrast, AI designed or used with a clear pedagogical purpose has greater potential to support learning. The OECD specifically highlights the possibilities of AI tutors that question, nudge and adapt their support rather than simply deliver solutions. OECD (2026).

This distinction is fundamental. Better performance on an AI-assisted task is not automatically evidence of better learning. A student can submit a better answer while acquiring less of the knowledge or skill required to produce that answer independently.

Why the design of the AI tutor matters

Learning requires cognitive activity. Students need to retrieve information, explain ideas, compare alternatives, make mistakes, receive feedback and revise their thinking. A system that removes all of those difficulties may feel efficient while quietly removing the very processes through which learning occurs.

This is why the most promising AI tutor is not necessarily the one that produces the best answer. It may be the one that produces the best learning sequence.

For example, an educational AI system can:

• ask a student to attempt a problem before revealing a solution;
• identify a misconception and ask a diagnostic question;
• request an explanation in the student’s own words;
• generate progressively harder practice questions;
• compare two possible arguments and ask the student to defend one;
• provide feedback without rewriting the student’s work for them;
• ask the student to verify an AI-generated claim against a source.

These designs preserve what educational researchers call productive cognitive effort. They also align with the OECD’s emphasis on pedagogical purpose and with the emerging direction of purpose-built tools from major AI companies.

What could this mean for students?

1. More individualized support

The strongest promise of AI tutors is availability. A student who gets stuck at home does not have to wait until the next lesson to ask a question. AI can provide explanations, examples and practice at a pace that fits the learner.

For students who have limited access to tutoring, this could be particularly valuable. But access must not be confused with effectiveness: the evidence shows that students still need to engage actively with the learning process.

2. A new risk of educational dependency

If students learn that every difficult step can immediately be outsourced, they may become less willing to struggle with unfamiliar problems. The danger is not simply “cheating”; it is the gradual weakening of independent problem solving.

The relevant question for educators therefore becomes: What intellectual work remains with the student?

3. AI literacy becomes part of learning itself

Students now need to evaluate not only books, websites and teachers’ explanations, but also machine-generated explanations. They need to ask: Is this claim accurate? What evidence supports it? What might the system have misunderstood?

The OECD and European Commission’s 2026 AI Literacy Framework places this capacity at the center of education: learners should understand AI, critically evaluate its outputs and use it ethically and creatively. OECD/European Commission (2026).

4. The AI divide may become a learning-behavior divide

The problem may no longer be simply who has access to AI. It may be who knows how to use it well.

Recent Microsoft Research work involving 318 university students found that proactive and critical AI engagement was associated with better learning outcomes, while learner background differences were partly explained by differences in how students interacted with AI. This suggests that the ability to use AI productively could itself become an educational advantage. Microsoft Research (2026).

A particularly important lesson for Morocco

The issue is already highly relevant to Moroccan students. According to the OECD’s PISA 2025 country findings, 45% of students in Morocco reported using AI chatbots to help them learn at least once a week, almost identical to the OECD average of 46%. Thirty-six percent reported using AI for preliminary research, 30% for summarizing texts and 29% for drafting writing assignments. Only 9% said they never or almost never used AI for the schoolwork tasks examined by PISA. OECD, PISA 2025: Morocco.

At the same time, the Moroccan data reveal an important inequality: the gap between advantaged and disadvantaged students in opportunities to assess the quality of AI-generated information was among the widest observed across participating systems. This suggests that the educational challenge is not simply to give students access to AI. Schools also need to teach students how to interrogate AI.

For Moroccan education, this means that banning AI alone is unlikely to be a sufficient long-term strategy. A more productive approach would distinguish between AI use that replaces learning and AI use that strengthens learning.

What should teachers do now?

Teachers do not need to turn every lesson into an AI lesson. Instead, they can establish a few clear principles:

1. Require an attempt before AI assistance. Students should demonstrate initial thinking before receiving extensive support.
2. Use AI for feedback rather than substitution. Ask the system to identify weaknesses or pose questions rather than rewrite the entire answer.
3. Make verification part of the task. Students should check important AI claims against textbooks, academic sources or reliable data.
4. Assess the process. Drafts, oral explanations, reflections and problem-solving steps can reveal learning better than a final AI-assisted product alone.
5. Teach AI literacy explicitly. Students need to understand hallucinations, bias, uncertainty, privacy and the limits of automated reasoning.

The real question is not whether students will have AI tutors

The latest announcements from OpenAI, Microsoft and other AI companies suggest that AI tutoring will become increasingly embedded in everyday student life. The technological trend is therefore unlikely to be reversed.

But the research does not support a simple claim that AI tutors automatically improve education. Some interventions show positive effects. Others show little added value. A recent randomized university study even found substantial negative effects on grades and participation. The strongest common lesson is that pedagogy mediates technology.

The future of AI in education will therefore not be decided by which company builds the most capable tutor. It will be decided by whether educational systems can design environments in which students still have to think, retrieve, explain, question, verify and create.

An AI tutor should make the student more capable, not make the student less necessary.

Research note

This article is an evidence-based synthesis of recent company announcements, randomized studies, systematic research and international policy evidence. It does not present original empirical research. The findings cited here vary in population, intervention and research design; therefore, causal claims should not be generalized beyond the contexts studied.

Selected references

• OpenAI. (2026). Helping teens learn, plan, and shape the future of AI.
• Oreopoulos, P., & Low, N. (2026). One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment. NBER Working Paper 35620. DOI: 10.3386/w35620.
• Liu, J., et al. (2026). The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial. EdWorkingPaper 26-1598. DOI: 10.26300/y3f8-vh05.
• Boolzen, C., Kuhn, J., & Flegr, S. (2026). Evidence of impact and interpretational limits of generative AI in STEM education. Artificial Intelligence Review, 59, 223. DOI: 10.1007/s10462-026-11665-9.
• OECD. (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. DOI: 10.1787/062a7394-en.
• OECD/European Commission. (2026). Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education. DOI: 10.1787/65cd27d4-en.
• OECD. (2026). PISA 2025 Results: Morocco.

Artificial intelligence is entering a new phase in education. The question is no longer simply whether students, teachers, and researchers use generative AI. Increasingly, institutions are asking a harder question: Can we determine how AI was involved in producing a piece of work?

On October 5, 2026, OpenAI announced a new approach to text provenance for ChatGPT and Codex. The company said it would begin introducing an invisible, machine-readable watermark into eligible AI-generated text for users in the European Union. The technology, called textGrain, creates a statistical signal through model word choices that can later be detected by an authorized system. The move responds to transparency requirements under the EU AI Act. OpenAI’s announcement.

From AI detection to AI provenance

An AI detector generally asks: Does this text look as if it was generated by AI? A provenance system asks a narrower question: Does this text contain a signal associated with a particular AI system?

OpenAI’s textGrain belongs to the second category. It does not insert visible labels, hidden characters, or unusual punctuation. It subtly influences statistical choices among possible words or word pieces, creating a pattern that a detector can search for.

This distinction matters enormously in education. A watermark may provide evidence that an AI system contributed to a passage. It does not establish how much of the work came from AI, what instructions the student provided, what the student understood, or whether the final submission represents meaningful human reasoning.

Why AI watermarking matters for assessment

Generative AI has challenged a basic assumption behind traditional assessment: the submitted product is evidence of the student’s learning. That assumption becomes weaker when students can use AI to generate essays, solve problems, summarize readings, write code, translate text, and produce presentations.

But replacing traditional assessment with AI detection creates another problem. Consider two students: one copies an AI-generated essay without understanding it; another researches independently, uses AI for feedback, challenges its claims, verifies sources, and substantially develops the final argument. Both may have interacted with AI, but their learning processes are completely different.

A binary question—“Was AI used?”—cannot adequately distinguish between them.

The scientific limitation: watermarks are not perfect

OpenAI reports that detection becomes more difficult with shorter passages and subjects where there is less flexibility in word choice, such as mathematics. Editing can also substantially weaken the watermark. In one evaluation, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%.

These findings have an immediate implication for schools and universities: a watermark should not be treated as a definitive verdict about authorship. A positive result may indicate that an AI system generated or processed some text. An absence of a watermark does not prove human authorship.

OpenAI also says that a watermark does not measure human contribution, establish ownership or responsibility, identify the user, or verify content accuracy.

Assessment may need to move from products to processes

The most important educational consequence of AI watermarking may be the acceleration of a transition from product-based assessment toward process-aware assessment.

Educators can increasingly evaluate research processes, drafts and revisions, oral explanations, source evaluation, problem-solving steps, reflective commentary, project development, classroom discussion, and the student’s ability to defend conclusions.

This does not mean abandoning essays or examinations. It means recognizing that the final text alone may no longer provide sufficient evidence of learning when generative AI is widely available.

AI literacy becomes part of learning

If students are going to work with AI, education needs to teach them how to work critically with AI. AI literacy includes evaluating AI-generated information, identifying unsupported claims, verifying sources, recognizing uncertainty, protecting sensitive information, understanding model limitations, disclosing AI assistance when appropriate, distinguishing assistance from substitution, and taking responsibility for final work.

The objective should not simply be to teach students how to use AI. It should be to teach them how to think when AI is available.

What teachers should do

AI prohibited

Students complete the task without generative AI.

AI assisted

Students may use AI for brainstorming, language assistance, or feedback but must produce and verify the substantive work themselves.

AI integrated

Students are explicitly asked to use AI and critically evaluate its output. For example: Use an AI system to explain a scientific problem, identify weaknesses in its answer, verify its sources, and produce a corrected explanation.

In this model, AI becomes not a shortcut around learning, but an object through which learning takes place.

What universities should avoid

Universities should be cautious about turning AI detection into an automated disciplinary mechanism. A detection signal should not automatically lead to a finding of misconduct. Students should have an opportunity to explain their work, demonstrate understanding, provide drafts where relevant, and respond to the evidence.

The purpose of assessment is ultimately not to identify machines. It is to understand learning.

What are we actually assessing?

A polished essay can demonstrate writing ability, but it may also demonstrate access to tools. A correct mathematical answer can demonstrate understanding, but it may also demonstrate the ability to obtain an answer from a machine. A programming assignment can demonstrate coding competence, but it may also demonstrate the ability to supervise an AI coding system.

None of this makes assessment meaningless. It means that assessment design must evolve alongside the technology.

What OpenAI’s watermark really changes

Text watermarking does not solve the AI cheating problem, create a perfect AI detector, or tell a teacher whether a student has genuinely learned something. What it represents is a move toward a world in which AI-generated content can carry provenance signals.

For education, this could support transparency—but provenance should be combined with better assessment. The strongest response is likely to combine provenance + human judgment + AI literacy + process-based assessment + clear institutional policy.

Conclusion

OpenAI’s textGrain announcement may be an early step toward a new norm of digital transparency. But education should resist the temptation to equate AI origin with misconduct.

The real challenge is bigger: schools and universities need to redesign learning environments in which students can use powerful AI systems without outsourcing the cognitive abilities education is supposed to develop.

The goal should not be classrooms where AI is impossible to use. It should be classrooms where using AI does not make thinking unnecessary.

Research basis

This article is an analytical synthesis of current reporting and official regulatory material. It does not present original empirical research.

• OpenAI — Our approach to EU text provenance rules, October 5, 2026.
• European Commission — Code of Practice on Transparency of AI-generated Content.
• European Commission — Transparency obligations under Article 50 of the AI Act.
• The Verge — OpenAI is adding text watermarking in ChatGPT and Codex, October 5, 2026.
• TechCrunch — OpenAI will start watermarking ChatGPT’s text in the EU, October 5, 2026.

Related reading: Generative AI in Education: What the Research Says About Learning and Assessment · Do AI Detectors Actually Work? What the Evidence Shows

Generative AI in education has moved from a question of technological possibility to a question of educational design. The central issue is no longer simply whether students and teachers will use tools such as large language models, but how those tools should be integrated into teaching, learning and assessment without weakening the purposes education is meant to serve.

This article reviews recent peer-reviewed research and international guidance to identify what the evidence currently supports, where findings remain uncertain, and what the shift means for educators and institutions. It is a research synthesis, not an original empirical study.

Why generative AI changes the educational question

Generative AI can produce text, explanations, examples, code, images and other forms of content on demand. That changes the relationship between a learner and a conventional educational task. A student may now obtain a plausible explanation, draft an argument or receive feedback before submitting work that previously required more independent production.

UNESCO’s guidance on generative AI in education and research frames this development as a question of human-centred educational design. It emphasizes human agency, inclusion, equity, data protection, appropriate validation of AI systems and the need to consider how AI affects what is taught, how it is taught and how learning is assessed. UNESCO, 2023.

This distinction matters. The educational value of a generative AI system cannot be inferred from the quality of its output alone. A fluent answer may help a learner understand a difficult concept, but the same answer can also bypass the cognitive work that an assignment was designed to elicit.

What recent research tells us about learning

A 2025 systematic review in the International Journal of Educational Technology in Higher Education examined 63 peer-reviewed studies of AI-based learning tools in higher education. The review identified assessment and evaluation, personalized feedback and recommendations, and intelligent tutoring among the main roles of AI-based tools. It also examined cognitive, skill-based and affective learning outcomes rather than treating “learning” as a single measure.

Another 2025 systematic review in TechTrends synthesized research on pedagogical applications of generative AI in higher education. Taken together with other reviews, the literature indicates that generative AI is being used across teaching, learning support, feedback, content generation and assessment-related activities, but the research base remains heterogeneous.

A 2025 meta-analysis of 57 studies and 97 effect estimates reported positive effects of generative AI on several university learning outcomes, including language skills, academic achievement, affective-motivational outcomes and higher-order thinking. However, the same analysis found no statistically significant effect on metacognition and identified substantial variation associated with learners, tools, roles, rules and learning contexts. The implication is important: positive average effects do not mean that every use of generative AI produces better learning.

In practical terms, the strongest interpretation of the current evidence is conditional. AI can function as a scaffold, tutor, feedback partner or source of examples, but the educational design still determines whether the learner is thinking, checking, explaining and applying—or simply accepting an answer.

The assessment problem is bigger than plagiarism

The most visible educational debate has focused on academic misconduct: whether a student used AI to produce work that was expected to be independently authored. That issue is real, but it is only one part of a broader assessment problem.

If a conventional assignment can be completed largely by generating a plausible response, the institution may lose information about the student’s reasoning process. The question therefore becomes: What exactly is the assessment intended to measure?

A 2025 systematic review of the early impact of AI on higher-education curriculum, instruction and assessment found that the emergence of generative AI was putting pressure on established assessment practices and on how academics design curriculum and instruction. A later systematic review focused specifically on university assessment similarly identified academic integrity, AI literacy, ethics and institutional governance as connected issues.

Research published in 2026 on generative AI and 21st-century skills assessment adds another dimension. Its systematic review argues that skills such as critical thinking, creativity, communication and collaboration remain relevant, while suggesting that assessment methods need to capture processes and evidence of learning rather than relying exclusively on conventional outputs.

What assessment can look like in an AI-enabled environment

The evidence does not imply that every assessment should become an oral examination or that written assignments should disappear. Instead, it points toward greater variety and stronger alignment between the learning objective and the evidence collected.

• Process evidence: drafts, revisions, research notes, problem-solving steps or reflective explanations can make learning more visible.
• Authentic tasks: students can be asked to apply concepts to unfamiliar cases, local problems or discipline-specific situations.
• Oral or dialogic components: short discussions can test whether a student can explain and defend the reasoning behind submitted work.
• AI-use disclosure: where AI is permitted, students can document how they used it, what they accepted or rejected, and how they verified outputs.
• Portfolio assessment: a collection of work over time can provide more evidence of development than a single generated product.

These approaches do not eliminate the need for academic integrity policies. They change the emphasis from trying to detect every possible instance of AI assistance toward designing assessments in which meaningful evidence of learning is harder to outsource.

AI literacy becomes part of the learning objective

Generative AI introduces a new literacy requirement. Students need to understand not only how to operate an AI system, but also how to evaluate its outputs, recognize uncertainty, protect sensitive information and decide when using AI is educationally appropriate.

UNESCO’s framework is explicitly human-centred: it treats AI competency as connected to human judgement, ethics and responsible participation rather than technical tool use alone. Recent research likewise identifies AI literacy and educator preparation as important conditions for responsible integration.

This changes the role of the teacher. The teacher is not simply a gatekeeper deciding whether AI is allowed. In many contexts, the teacher becomes a designer of tasks in which students learn how to question, verify and use AI without surrendering responsibility for the final intellectual product.

What institutions should change—and what they should not assume

Institutions considering generative AI policies should avoid two opposite assumptions. The first is that prohibition alone can restore pre-AI assessment conditions. The second is that introducing AI automatically modernizes teaching.

The research instead points toward several institutional priorities:

1. Define permitted and prohibited uses clearly. Policies should distinguish brainstorming, translation, feedback, coding assistance and content generation rather than treating all AI use as one category.
2. Align assessment with learning outcomes. If critical reasoning is the objective, the assessment should generate evidence of critical reasoning.
3. Train educators. Faculty need practical knowledge of AI capabilities, limitations, assessment redesign and responsible use.
4. Protect privacy and data. Institutions should evaluate what information students and staff enter into external AI systems.
5. Evaluate evidence continuously. Generative AI systems and educational practices are changing quickly; policies should therefore be revisited as evidence develops.

Importantly, the research base still has limitations. Studies differ in discipline, educational level, duration, AI system, learning task, outcome measure and research design. Many findings are therefore context-dependent. Systematic reviews can identify patterns across studies, but they cannot remove the underlying heterogeneity of the evidence.

What the evidence means for educators

For educators, the practical question is not “How do I stop students from using AI?” in isolation. A more useful sequence is:

What should students learn? What evidence would demonstrate that learning? Which parts of the task can appropriately involve AI? Which parts require independent human reasoning? And how will students be taught to verify AI-generated information?

That approach is consistent with the direction of recent scholarship. Generative AI can be integrated into education, but its educational value depends on the surrounding pedagogy, the assessment design and the quality of human oversight.

Conclusion

The emerging evidence on generative AI in education is more nuanced than either technological optimism or technological alarm. Recent systematic reviews report meaningful opportunities for learning support, feedback, personalization and some learning outcomes, while also identifying unresolved challenges around assessment, metacognition, academic integrity, AI literacy, privacy and institutional governance.

The central educational task is therefore not to choose between “AI” and “no AI.” It is to decide where AI adds educational value, where it undermines the purpose of a task, and how assessment can continue to provide credible evidence of what a learner knows and can do.

For institutions, this means treating generative AI as a curriculum and assessment question—not merely as an IT policy question.

Research basis and references

1. Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. Source.
2. Luo, J., Zheng, C., Yin, J., et al. (2025). “Design and assessment of AI-based learning tools in higher education: a systematic review.” International Journal of Educational Technology in Higher Education, 22, 42. DOI.
3. Qian, Y. (2025). “Pedagogical Applications of Generative AI in Higher Education: A Systematic Review of the Field.” TechTrends, 69, 1105–1120. DOI.
4. Belkina, M., Daniel, S., Nikolic, S., et al. (2025). “Implementing generative AI (GenAI) in higher education: A systematic review of case studies.” Computers and Education: Artificial Intelligence, 8, 100407. DOI.
5. Liang, J., Stephens, J. M., & Brown, G. T. L. (2025). “A systematic review of the early impact of artificial intelligence on higher education curriculum, instruction, and assessment.” Frontiers in Education, 10, 1522841. DOI.
6. Chen, S., & Cheung, A. C. K. (2025). “Effect of generative artificial intelligence on university students learning outcomes: A systematic review and meta-analysis.” Educational Research Review, 100737. DOI.
7. Nyagoga, L. M., Malik, L., & Lukoko, H. O. (2026). “Generative AI’s implications on higher education students’ 21st-century skills assessment: insights from a systematic literature review.” Artificial Intelligence in Education, 2, 109–126. DOI.

Editorial note: This article synthesizes published research and international guidance. It does not present original empirical data. Findings should be interpreted in light of differences in study design, population, discipline, intervention and assessment context.