Article by
Milo
ESL Content Coordinator & Educator
ESL Content Coordinator & Educator
All Posts
Ultimate Teacher Planner #
The ultimate all-in-one education management system in Notion.
Learn More
Ultimate Teacher Planner #
The ultimate all-in-one education management system in Notion.
Learn More
Table of Contents #
In spring 2026, Brown University economics professor Roberto Serrano faced a result that looked impressive but raised immediate doubts. The average score on a take-home midterm was 96 percent, far above the course’s historical range of 65 to 80 percent. Serrano suspected that many students had used ChatGPT or similar tools to generate answers. He replaced the planned take-home final with an in-person exam. Eighteen students dropped the course, nine remained enrolled but did not sit the final, and the final-exam average fell to 48.6 percent. Nineteen students ultimately failed.[1]
The episode does not prove that every high-scoring student cheated. It does, however, expose a central problem created by generative AI: academic output can now look stronger while the underlying knowledge is weaker. A student may submit a coherent essay, proof, or explanation without developing the ability to reproduce, defend, or apply it independently.
That distinction matters because education is not merely the production of answers. It is the gradual construction of memory, judgement, analytical habits, and domain expertise. ChatGPT can support that construction, but it can also bypass it.
The Central Risk: Confusing Task Completion with Learning #
People have always used tools to reduce mental effort. Notes preserve information. Calculators automate arithmetic. Search engines locate facts. Cognitive scientists call this cognitive off: moving part of a mental task into the external environment.[2] Off is not automatically harmful. It becomes risky when the outsourced process is the same process a learner needs to practise.
For example, a student learning statistical analysis must struggle with selecting methods, interpreting results, and explaining limitations. If ChatGPT performs those steps and the student edits the wording, the assignment may be completed, but the analytical skill has received little training. The same applies to essay structure, mathematical reasoning, source evaluation, coding, and problem-solving. A large field experiment published in the Proceedings of the National Academy of Sciences illustrates this distinction. Nearly 1,000 students were given access to different GPT-4-based mathematics tools. A standard ChatGPT-like interface improved performance while the tool was available. Yet when access was removed, those students performed 17 percent worse than students who had never received AI assistance. A guarded tutor that offered structured help rather than simply supplying answers largely reduced this learning penalty.[3]
Research involving university students has produced a related warning. One study found that frequent academic use of ChatGPT was associated with greater procrastination, reported memory difficulties, and lower academic performance. These findings should not be treated as proof that any use of ChatGPT causes cognitive decline. They do suggest that patterns of excessive reliance can become self-reinforcing, particularly when students use AI under workload and deadline pressure.[4]
Productive AI Use Keeps the Student Inside the Reasoning Process #
The practical question is not whether students should use ChatGPT. The technology is already embedded in study and work. The relevant question is which parts of learning should remain under the student’s control.
A productive workflow begins with an independent attempt. The student first explains the concept, solves the problem, or drafts the argument without AI. ChatGPT can then be asked to identify missing assumptions, present a counterargument, explain an error, generate practice questions, or compare alternative methods. Finally, the student should close the tool and reconstruct the answer from memory.
This sequence matters. Retrieval practice, in which learners actively recall information rather than repeatedly rereading it, has consistently produced stronger conceptual learning and longer retention.[5] ChatGPT can help create retrieval prompts, but it should not perform the retrieval on the student’s behalf.
The difference can be expressed simply. Asking ChatGPT to “write my answer” transfers authorship and reasoning to the system. Asking it to “challenge my answer, identify weaknesses, and ask me questions without giving the solution” uses the system as structured feedback. In the first case, AI replaces cognition. In the second, it increases the amount and quality of cognitive work.
Transcripts and Summaries Solve a Different Problem #
Recorded lectures create a separate challenge. A two-hour video is difficult to search, annotate, compare with readings, or review quickly before an exam. Students may waste significant time locating one definition or replaying a section they partly understood. In this context, transcription and summarisation can remove mechanical friction without necessarily removing intellectual effort.
A transcript turns spoken material into searchable text. A summary provides an initial map of the lecture. Neither should be treated as the final object of study. The student still needs to check the summary against the source, distinguish central claims from examples, write explanations in their own words, and test what they can recall without assistance.
This is where a canvas video down can fit into a responsible study workflow. Canvas Assistant is a browser extension that can save supported lecture videos from platforms such as Canvas LMS, Blackboard, Moodle, and Panopto, then help convert them into transcripts and summaries. Its role is primarily organisational: it makes recorded material easier to access, search, and revisit. Students who need platform-specific instructions can consult its guide explaining how to download video from Canvas.
Used well, this workflow creates more opportunities for active study. A student can extract the lecture’s main claims, turn headings into questions, compare the transcript with personal notes, generate a practice quiz, and return to the original recording when a summary omits nuance. Used poorly, the same tools can become another shortcut: read the summary, copy several points, and assume the lecture has been learned.
The technology is identical. The cognitive behaviour is not.
Assessment Will Increasingly Test Visible Reasoning #
The Brown case also shows why universities are reconsidering assessment. When polished written work can be generated instantly, a final submission provides less evidence of what a student can actually do. In-person exams, oral defences, staged drafts, annotated sources, practical demonstrations, and questions about the reasoning process are likely to become more important.
For students, this changes the risk calculation. Outsourcing weekly assignments may produce acceptable grades temporarily, but it leaves little preparation for situations in which knowledge must be demonstrated without assistance. It also weakens the expertise required to detect when ChatGPT is confidently wrong. UNESCO’s guidance on generative AI argues for a human-centred approach in which technology remains subordinate to human agency, ethical judgement, and meaningful learning.[6] That principle is useful at the individual level. Students should automate access, formatting, transcription, initial organisation, and low-value repetition where appropriate. They should be far more cautious about automating interpretation, argumentation, evaluation, and original problem-solving.
The Choice Is Becoming More Consequential #
AI will not divide students simply into users and non-users. A more important divide is emerging between students who use AI to avoid cognitive work and students who use it to create better conditions for cognitive work.
The first group may become highly efficient at producing assignments while remaining dependent on tools they cannot evaluate. Their performance becomes fragile because it relies on continuous access, reliable outputs, and assessments that do not require independent explanation. The second group builds knowledge and judgement while using AI to reduce friction, obtain feedback, organise information, and practise more deliberately.
ChatGPT, transcription tools, and Canvas Assistant can all contribute to either pattern. None of them determines the outcome. The decisive factor is whether the student remains responsible for understanding.
The final question is therefore not whether AI belongs in education. It already does. The question is whether students will use it to escape the work that develops competence, or to remove unnecessary friction so they can do more of that work.
Endnotes #
[1] Emma Whitford, “Brown Professor Suspects Majority of His Class Used AI to Cheat,” Inside Higher Ed, 8 July 2026.
[2] Evan F. Risko and Sam J. Gilbert, “Cognitive Off,” Trends in Cognitive Sciences, vol. 20, no. 9, 2016, pp. 676–688.
[3] Hamsa Bastani et al., “Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics,” Proceedings of the National Academy of Sciences, vol. 122, no. 26, 2025.
[4] Muhammad Abbas, Fawad Ali Jam, and Tariq Iqbal Khan, “Is It Harmful or Helpful? Examining the Causes and Consequences of Generative AI Usage Among University Students,” International Journal of Educational Technology in Higher Education, vol. 21, 2024.
[5] Jeffrey D. Karpicke and Janell R. Blunt, “Retrieval Practice Produces More Learning Than Elaborative Studying with Concept Mapping,” Science, vol. 331, no. 6018, 2011, pp. 772–775.
[6] Fengchun Miao and Wayne Holmes, Guidance for Generative AI in Education and Research, UNESCO, 2023.
In spring 2026, Brown University economics professor Roberto Serrano faced a result that looked impressive but raised immediate doubts. The average score on a take-home midterm was 96 percent, far above the course’s historical range of 65 to 80 percent. Serrano suspected that many students had used ChatGPT or similar tools to generate answers. He replaced the planned take-home final with an in-person exam. Eighteen students dropped the course, nine remained enrolled but did not sit the final, and the final-exam average fell to 48.6 percent. Nineteen students ultimately failed.[1]
The episode does not prove that every high-scoring student cheated. It does, however, expose a central problem created by generative AI: academic output can now look stronger while the underlying knowledge is weaker. A student may submit a coherent essay, proof, or explanation without developing the ability to reproduce, defend, or apply it independently.
That distinction matters because education is not merely the production of answers. It is the gradual construction of memory, judgement, analytical habits, and domain expertise. ChatGPT can support that construction, but it can also bypass it.
The Central Risk: Confusing Task Completion with Learning #
People have always used tools to reduce mental effort. Notes preserve information. Calculators automate arithmetic. Search engines locate facts. Cognitive scientists call this cognitive off: moving part of a mental task into the external environment.[2] Off is not automatically harmful. It becomes risky when the outsourced process is the same process a learner needs to practise.
For example, a student learning statistical analysis must struggle with selecting methods, interpreting results, and explaining limitations. If ChatGPT performs those steps and the student edits the wording, the assignment may be completed, but the analytical skill has received little training. The same applies to essay structure, mathematical reasoning, source evaluation, coding, and problem-solving. A large field experiment published in the Proceedings of the National Academy of Sciences illustrates this distinction. Nearly 1,000 students were given access to different GPT-4-based mathematics tools. A standard ChatGPT-like interface improved performance while the tool was available. Yet when access was removed, those students performed 17 percent worse than students who had never received AI assistance. A guarded tutor that offered structured help rather than simply supplying answers largely reduced this learning penalty.[3]
Research involving university students has produced a related warning. One study found that frequent academic use of ChatGPT was associated with greater procrastination, reported memory difficulties, and lower academic performance. These findings should not be treated as proof that any use of ChatGPT causes cognitive decline. They do suggest that patterns of excessive reliance can become self-reinforcing, particularly when students use AI under workload and deadline pressure.[4]
Productive AI Use Keeps the Student Inside the Reasoning Process #
The practical question is not whether students should use ChatGPT. The technology is already embedded in study and work. The relevant question is which parts of learning should remain under the student’s control.
A productive workflow begins with an independent attempt. The student first explains the concept, solves the problem, or drafts the argument without AI. ChatGPT can then be asked to identify missing assumptions, present a counterargument, explain an error, generate practice questions, or compare alternative methods. Finally, the student should close the tool and reconstruct the answer from memory.
This sequence matters. Retrieval practice, in which learners actively recall information rather than repeatedly rereading it, has consistently produced stronger conceptual learning and longer retention.[5] ChatGPT can help create retrieval prompts, but it should not perform the retrieval on the student’s behalf.
The difference can be expressed simply. Asking ChatGPT to “write my answer” transfers authorship and reasoning to the system. Asking it to “challenge my answer, identify weaknesses, and ask me questions without giving the solution” uses the system as structured feedback. In the first case, AI replaces cognition. In the second, it increases the amount and quality of cognitive work.
Transcripts and Summaries Solve a Different Problem #
Recorded lectures create a separate challenge. A two-hour video is difficult to search, annotate, compare with readings, or review quickly before an exam. Students may waste significant time locating one definition or replaying a section they partly understood. In this context, transcription and summarisation can remove mechanical friction without necessarily removing intellectual effort.
A transcript turns spoken material into searchable text. A summary provides an initial map of the lecture. Neither should be treated as the final object of study. The student still needs to check the summary against the source, distinguish central claims from examples, write explanations in their own words, and test what they can recall without assistance.
This is where a canvas video down can fit into a responsible study workflow. Canvas Assistant is a browser extension that can save supported lecture videos from platforms such as Canvas LMS, Blackboard, Moodle, and Panopto, then help convert them into transcripts and summaries. Its role is primarily organisational: it makes recorded material easier to access, search, and revisit. Students who need platform-specific instructions can consult its guide explaining how to download video from Canvas.
Used well, this workflow creates more opportunities for active study. A student can extract the lecture’s main claims, turn headings into questions, compare the transcript with personal notes, generate a practice quiz, and return to the original recording when a summary omits nuance. Used poorly, the same tools can become another shortcut: read the summary, copy several points, and assume the lecture has been learned.
The technology is identical. The cognitive behaviour is not.
Assessment Will Increasingly Test Visible Reasoning #
The Brown case also shows why universities are reconsidering assessment. When polished written work can be generated instantly, a final submission provides less evidence of what a student can actually do. In-person exams, oral defences, staged drafts, annotated sources, practical demonstrations, and questions about the reasoning process are likely to become more important.
For students, this changes the risk calculation. Outsourcing weekly assignments may produce acceptable grades temporarily, but it leaves little preparation for situations in which knowledge must be demonstrated without assistance. It also weakens the expertise required to detect when ChatGPT is confidently wrong. UNESCO’s guidance on generative AI argues for a human-centred approach in which technology remains subordinate to human agency, ethical judgement, and meaningful learning.[6] That principle is useful at the individual level. Students should automate access, formatting, transcription, initial organisation, and low-value repetition where appropriate. They should be far more cautious about automating interpretation, argumentation, evaluation, and original problem-solving.
The Choice Is Becoming More Consequential #
AI will not divide students simply into users and non-users. A more important divide is emerging between students who use AI to avoid cognitive work and students who use it to create better conditions for cognitive work.
The first group may become highly efficient at producing assignments while remaining dependent on tools they cannot evaluate. Their performance becomes fragile because it relies on continuous access, reliable outputs, and assessments that do not require independent explanation. The second group builds knowledge and judgement while using AI to reduce friction, obtain feedback, organise information, and practise more deliberately.
ChatGPT, transcription tools, and Canvas Assistant can all contribute to either pattern. None of them determines the outcome. The decisive factor is whether the student remains responsible for understanding.
The final question is therefore not whether AI belongs in education. It already does. The question is whether students will use it to escape the work that develops competence, or to remove unnecessary friction so they can do more of that work.
Endnotes #
[1] Emma Whitford, “Brown Professor Suspects Majority of His Class Used AI to Cheat,” Inside Higher Ed, 8 July 2026.
[2] Evan F. Risko and Sam J. Gilbert, “Cognitive Off,” Trends in Cognitive Sciences, vol. 20, no. 9, 2016, pp. 676–688.
[3] Hamsa Bastani et al., “Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics,” Proceedings of the National Academy of Sciences, vol. 122, no. 26, 2025.
[4] Muhammad Abbas, Fawad Ali Jam, and Tariq Iqbal Khan, “Is It Harmful or Helpful? Examining the Causes and Consequences of Generative AI Usage Among University Students,” International Journal of Educational Technology in Higher Education, vol. 21, 2024.
[5] Jeffrey D. Karpicke and Janell R. Blunt, “Retrieval Practice Produces More Learning Than Elaborative Studying with Concept Mapping,” Science, vol. 331, no. 6018, 2011, pp. 772–775.
[6] Fengchun Miao and Wayne Holmes, Guidance for Generative AI in Education and Research, UNESCO, 2023.
Ultimate Teacher Planner #
The ultimate all-in-one education management system in Notion.
Learn More
Ultimate Teacher Planner #
The ultimate all-in-one education management system in Notion.
Learn More
2026 Notion4Teachers. All Rights Reserved.
2026 Notion4Teachers. All Rights Reserved.
2026 Notion4Teachers. All Rights Reserved.