On 26 August 2026, OpenAI released a report on how students and educators use ChatGPT. Drawing on what it described as privacy-preserving analysis, the company said that people across all age groups have about 70 million conversations each week devoted to testing what they know, including checking misconceptions and requesting further practice. In the United States, prompts related to classwork and homework peak at more than 460 million messages a week during the school year and remain above 180 million a week during summer. OpenAI presented these patterns as evidence that AI is extending learning beyond the classroom and fixed study hours, allowing students to seek explanations, feedback and practice when a question arises.
The figures show that AI has become an everyday part of the educational environment, but they do not directly show that learning is continuously taking place. The report measures conversations, timing and classified uses, not knowledge retention, transfer or independent judgement. A learner may ask a question, receive the right answer and even submit a better piece of work without having learned the underlying material. The philosophical issue is therefore not whether AI can provide help at any time. It is the relation between continuous access to assistance and the continuing formation of knowledge.
At least four things need to be separated. Information availability means that an answer can be found. Task performance means that a person meets the requirements of a particular exercise. Learning means that practice, feedback and correction have produced a change in the person's capacity that can be retained. Knowledge requires that this changed structure can be called upon, tested and revised in relevant situations. The first two may occur almost entirely on the tool's side. The latter two must alter the learner's relation to the problem. AI can produce a result for someone or help that person form a capacity. Both may look like a successful conversation, while their epistemic structures are quite different.
The OECD's 2026 Digital Education Outlook draws a similar distinction from the emerging research. General-purpose generative AI often improves students' performance while the tool is available, yet that advantage can disappear, and sometimes reverse, when the tool is removed in an examination. The OECD therefore argues that without pedagogical purpose and process constraints, outsourcing cognitive tasks to a chatbot may improve immediate output without producing genuine learning gains. This does not mean cognitive offloading is inherently harmful. Books, calculators, search engines and teachers have long served as external epistemic resources. The decisive question is whether the external resource supports the formation of a capacity or replaces the discrimination, reasoning and correction through which the learner would acquire it.
Nor does this risk imply that AI tutoring is ineffective. A randomised controlled trial published in Scientific Reports in 2025 compared AI tutoring with in-class active learning in an undergraduate physics course at Harvard University. The AI group achieved larger immediate test gains in less time and reported greater engagement and motivation. The conditions behind that result matter. Students were not simply given a general chatbot and invited to request finished answers. The researchers built pedagogical principles into the system, provided sequential scaffolding, required active participation and supplied detailed worked solutions to reduce hallucinations. They also stated that the experiment concerned immediate outcomes from two physics lessons and did not establish that the method would outperform classroom teaching in every setting, particularly those requiring complex synthesis and higher-order critical thought.
The two bodies of evidence are compatible. Together they show that AI's educational effect is not determined by the bare fact that AI was used. It depends on how the interaction organises cognitive activity. A system that immediately supplies a finished answer shortens the path to task completion, but may also remove the learner's opportunity to identify an error, try alternatives and explain why a conclusion follows. A well-designed tutor can deliberately delay the answer, using questions, prompts, counterexamples and timely feedback to return the cognitive work to the learner. Greater convenience does not automatically produce better learning. In some cases, well-designed resistance is one of the conditions under which learning occurs.
Sustenesis Theory makes this distinction more precise. Difference is the starting point at which a learner can distinguish a problem, an existing understanding and new evidence. Constraint consists of the conditions that prevent arbitrary acceptance and require an understanding to answer to facts, logic, the task and corrective feedback. Sustained Coherence is not the fluency of a single response. It is the continued operation of an epistemic structure after later retrieval, exposure of error and correction. AI can quickly produce linguistic coherence, but learning occurs only when that coherence becomes part of a structure the learner can preserve, call upon, test and revise.
Knowledge need not be defined narrowly as content stored entirely inside an individual's brain. A person working with books, databases and AI can form a combined system with greater operational knowledge than unaided memory could provide. If a doctor can frame a question accurately, identify its conditions of application, check sources, detect a model's mistake and incorporate the result into professional responsibility, AI has become part of the structure through which knowledge operates. If the user can only copy a conclusion while the model is present, cannot explain why it holds and cannot recognise when changed conditions make it fail, what exists is access to a knowledge resource rather than possession of the corresponding capacity. External knowledge can participate in human cognition, but a frictionless interface does not automatically turn it into personal competence.
My judgement is therefore that AI can make assistance continuously available, but cannot by itself make learning continuous. If “continuous learning” means only that a dialogue box can be opened at any hour, service availability has been mistaken for epistemic change. Genuine continuity should appear as an observable change in the learner after repeated interactions. Can the learner explain the concept again without the original prompt? Can the method be transferred to a new problem? Can a plausible but false response be identified? Can the learner continue as support is gradually reduced? These are better indicators of learning than the number of conversations.
This distinction also changes how educational assessment should be understood. Measuring only what students complete while AI is available confuses the joint performance of person and tool with the student's learning. An examination that bans AI entirely measures unaided capacity, but cannot show whether a student can use the external knowledge systems present in ordinary life. A more adequate assessment would examine what a learner can do without AI, with pedagogically structured AI, and with unrestricted general-purpose AI, then determine which capacities remain after support is withdrawn or the problem changes. Only then can personal knowledge, human–AI joint capability and simple substitution by a tool be distinguished.
The present evidence does not settle the long-term question. OpenAI's usage figures come from the company's own analysis and are not a study of learning outcomes. The Harvard trial measured short-term results in a specific course. The OECD synthesis itself describes much of the evidence as emerging. What is still needed includes independent research using delayed tests, transfer across contexts, error detection, fading support and different age groups. The conclusion justified now is limited but clear. Answers can be generated instantly, but knowledge must still be formed. AI's educational value lies not in eliminating that process, but in whether it supplies better differences, constraints, feedback and correction through which the process can continue.
References
OpenAI, Learning never stops: How AI makes learning continuous, 26 August 2026
https://openai.com/index/learning-never-stops/
OECD, OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education, 19 January 2026
https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html
Kestin et al., AI tutoring outperforms in-class active learning, Scientific Reports, 3 June 2025
https://www.nature.com/articles/s41598-025-97652-6
Brookings Institution, What the research shows about generative AI in tutoring, 27 January 2026
https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.