Why Learning Is a Special Case

Most lessons in this course concern tasks in which the product matters: an evidence scan, a briefing note, a cleaned dataset. Learning differs because the product is a change in the learner. When a student works through a practice problem, the written solution has little value of its own. Its value lies in the knowledge and skill that producing it builds. An assistant that produces the solution therefore completes the visible part of the task and can leave the useful part undone. The same tool can help an analyst who needs a correct answer quickly and hinder a student who needs to become able to produce that answer, even when both are working on the same problem.

This section reviews the strongest evidence available in 2026 on what AI does to learning. Two randomized studies published in 2025 point in opposite directions, and reading them together yields a principle that organizes the rest of the lesson.

Box 4.1 introduces Jordan, a fictional co-op student whose studying during the Cedar Valley pilot serves as the running case for the lesson.

Judging whether an AI tutor helps requires a point of comparison. The next part sets out the benchmark that one-to-one tutoring provides.

Tutoring as the Benchmark

One-to-one tutoring has long served as the reference point for effective teaching. Bloom (1984) reported that students tutored individually performed about two standard deviations above students taught in a conventional class, and he described the search for group methods that could match tutoring as the "2 sigma problem". An effect size expressed in standard deviations compares the average scores of two groups in units of the spread of scores: an effect of 1.0 means that the average student in one group scored one standard deviation above the average student in the other.

Later syntheses found smaller effects. VanLehn (2011) reviewed experiments that compared human tutoring with no tutoring and found an average effect of about 0.79 standard deviations, and computer tutoring systems that guided students step by step through problems came close, at about 0.76. Tutoring helps substantially, though less than the two-sigma figure suggested, and it is expensive. Large language models raised the prospect of a patient tutor available to any student at any hour for little or no cost. The two studies below show that whether this prospect is realized depends on how the tool is designed and used.

Evidence That an AI Tutor Can Strengthen Learning

Kestin and colleagues (2025) ran a randomized trial in Physical Sciences 2, an introductory physics course for life sciences students at Harvard University, in the fall of 2023. The comparison condition was demanding. The course already taught through active learning, in which students work on problems in class, discuss them with peers and receive feedback from the instructor, an approach with strong evidence of benefit over lecturing.

The trial used a crossover design, in which each participant experiences both conditions in turn. During two consecutive weeks, on the topics of surface tension and fluid flow, each student attended the active-learning class for one topic and worked through the same material at home with an AI tutor for the other, with the order randomized. Of the 233 students enrolled, 194 were included in the analysis. Students took a short test before and after each lesson.

Students learned more with the AI tutor. The median post-test score was 4.5 after the AI lesson and 3.5 after the in-class lesson, from a median of 2.75 before instruction, so the median gain with the tutor was more than double the gain in class. A regression estimate put the effect at 0.63 standard deviations, which the authors regard as an underestimate because many students reached the top of the test's scale. Students also spent less time: the median time with the tutor was 49 minutes, compared with about 60 minutes of learning time in class (the class periods ran 75 minutes), and time spent was not related to the post-test score. Students reported higher engagement and motivation with the tutor.

The design of the tutor explains much of the result. It ran on a commercial language model, but its system prompt was written around established teaching principles: it asked students to engage actively, managed the amount of new information presented at once, encouraged a view of ability as something that grows with effort, and guided students through each part of a multi-part problem in sequence. The prompts also contained pre-written, step-by-step solutions, so that the tutor's feedback did not depend on the model working out the physics itself. The tutor gave hints and feedback as students worked; it was built to help students reach the answers themselves.

The study has limits that a reader should keep in view. It covered two introductory topics in one course at one institution, the tutor was built by the study team, and the authors do not claim that AI tutoring would outperform active learning for tasks that call for complex synthesis or higher-order critical thinking.

The Harvard trial shows that a tutor built around teaching principles can strengthen learning. The next part turns to a study in which a general assistant used during practice weakened it.

Evidence That a General Assistant Can Weaken Learning

Bastani and colleagues (2025) ran a randomized experiment with nearly 1,000 students in grades 9, 10 and 11, in about fifty classes at a large high school in Turkey, during the fall of 2023. Over four 90-minute sessions, students reviewed mathematics topics that made up about 15 percent of the semester's curriculum, worked practice problems, and then sat an examination on the same material without any assistant. Students were assigned to one of three conditions.

• In the control condition, students practised with their textbook and class notes.
• In the GPT Base condition, students practised with a chat interface similar to a standard general assistant, whose prompt contained only the current problem and a request to help the student solve it.
• In the GPT Tutor condition, students used the same interface, but its prompt also contained the correct solution and the common mistakes that teachers had identified, and it instructed the model to give incremental hints without revealing the full solution.

In short, unrestricted access to a general assistant raised scores on practice problems and lowered scores on a later examination taken without it, while a version designed to give hints in place of answers largely avoided the harm. During practice, scores were 48 percent higher than the control group's with GPT Base and 127 percent higher with GPT Tutor. On the examination, students who had practised with GPT Base scored 17 percent lower than the control group, a statistically significant difference. Students who had practised with GPT Tutor scored about the same as the control group; the authors found no harm and no gain.

Figure 4.1 shows these practice and examination results for both AI conditions relative to the control group.

The authors also examined what students typed. Students in the GPT Base condition most often simply asked for the answer, and "what is the answer?" was the most common first message. The authors concluded that many students were copying answers, so that the assistant did the work that practice is meant to make students do. The students did not notice the cost: those who had used GPT Base did not perceive that they had performed worse or learned less, and those who had used GPT Tutor believed that they had performed significantly better than the control group although their examination scores were the same. Section 4 returns to this gap between the feeling of learning and learning itself.

The next part sets the two studies side by side to identify what separated the tools that helped or did no harm from the one that harmed learning.

Reading the Two Studies Together

The two trials differ in setting, participants, design and comparison condition, and they also differ in what the AI tools did with answers. Table 4.1 sets out these features and the main learning result of each trial as the basis for comparing them.

Table 4.1. Setting, design, comparison condition and main learning result of the randomized studies by Kestin and colleagues (2025) and Bastani and colleagues (2025).

The tools that helped or did no harm shared three features: they held correct solutions supplied by teachers, they responded to the student's attempt with hints and questions, and they withheld complete answers. The tool that harmed learning gave students what they asked for, and what they asked for was the answer. The underlying technology was similar in all three AI conditions. What differed was who did the thinking.

The studies also differ in ways that limit direct comparison. Kestin's tutor improved learning beyond a strong class, while Bastani's GPT Tutor only matched practice with a textbook. The Harvard tutor was embedded in a carefully designed lesson with structured problems and video introductions, and it served university students who had chosen their course. A tutoring design is therefore necessary for benefit in these studies, and it may not be sufficient: the surrounding lesson, the student's purpose and the quality of the problems matter too.

The comparison reduces to a single question about who does the thinking. The next part states it as a principle and connects it to earlier research on memory and learning.

The Principle: Who Does the Thinking

The evidence supports a working principle for the rest of this lesson: AI supports learning when it prompts the student to think, and it weakens learning when it does the thinking. The principle agrees with decades of research on memory and learning that predates language models.

The generation effect is the finding that people remember information better when they produce it themselves than when they read it. Slamecka and Graf (1978) showed that participants who generated a word from a cue remembered it better than participants who read the same word. The testing effect, which Section 2 develops, is the related finding that retrieving information from memory strengthens it more than restudying it does. Bjork (1994) coined the term desirable difficulties for conditions of practice, such as retrieving, spacing and varying problems, that make learning feel slower and harder in the moment and yet produce better long-term retention and transfer. An assistant that removes every difficulty from practice removes the desirable ones along with the rest, and the practice then feels productive while producing little learning.

The principle does not say that handing work to AI is always harmful. A public health analyst who asks an assistant to format a reference list has handed over a task they already know how to do, so nothing is lost. The question to ask is whether the part of the task being handed over is the part one is trying to learn. A student learning epidemiology can reasonably ask an assistant to tidy the formatting of a table; a student learning to calculate an odds ratio gains nothing if the assistant calculates it for them before they have tried.

The principle applies to individual learning tasks. The next part widens the view to cognitive offloading, the use of tools and actions to reduce the mental demands of a task, and to survey evidence on its association with critical thinking.

Cognitive Offloading

Risko and Gilbert (2016) define cognitive offloading as the use of physical action to alter the information-processing requirements of a task so as to reduce cognitive demand. Writing a shopping list, setting a phone reminder and tilting one's head to read rotated text are all forms of offloading. Offloading is ordinary and often sensible, since it frees attention for other work. Language models extend the range of what can be offloaded from memory and calculation to explanation, argument and judgement.

Gerlich (2025) surveyed 666 participants in the United Kingdom and interviewed 50 of them. Participants who reported more frequent use of AI tools scored lower on a measure of critical thinking (a correlation of −0.68), reported more cognitive offloading, and the association between AI use and critical thinking was partly explained by offloading in a mediation analysis. Younger participants reported heavier use and more offloading, and participants with more education scored higher on critical thinking regardless of their use of AI.

These findings are consistent with the experiments above, but a cross-sectional survey cannot establish their direction. A cross-sectional survey measures everyone at one point in time, so it cannot show whether heavy AI use reduced critical thinking, whether people less inclined to think critically chose to use AI more, or whether a third factor influenced both. The measures were also largely self-reported. The survey is best read as a reason for attention and as a source of hypotheses, while the randomized studies carry the weight of the causal argument. Lesson 7 returns to related survey evidence on confidence in AI and critical thinking among knowledge workers.

Try the principle on everyday requests

Interactive 4.1 lists eight requests that students commonly make of an assistant. For each one, decide whether the request prompts the student to think or asks the assistant to do the thinking that the student is trying to learn.

Interactive 4.1. Eight requests that students commonly make of an assistant, each to be sorted by whether it prompts the student to think or asks the assistant to do the thinking.

The flip cards that follow recap the four sources discussed in this section: Kestin and colleagues (2025), Bastani and colleagues (2025), Gerlich (2025) and Risko and Gilbert (2016).

The accordion that follows takes up three questions that the evidence raises: whether students should avoid general assistants, whether the results apply to university health sciences students, and why the GPT Tutor students did not do better on the examination.

Box 4.2 applies the evidence in this section to a practice session that a student completes quickly and without a single wrong answer.

The evidence in this section establishes what a learning-oriented use of AI looks like: the assistant asks, hints, checks and gives feedback, and the student attempts, retrieves and explains. Section 2 turns this principle into tutoring prompts grounded in learning science.

What a Good Tutor Does

A skilled human tutor does more than explain. The tutor asks questions that reveal what the student already understands, adjusts explanations to that level, shows worked examples and then fades them, makes the student recall and apply ideas, and returns to earlier material before it is forgotten. Each of these moves has a research base, and each can be requested from a language model. The default behaviour of a general assistant is different: it has been trained to be helpful by answering, so it tends to supply the explanation and the solution at once. Tutoring behaviour has to be asked for.

Table 4.2 summarizes five techniques, the evidence behind them and how to request them. Dunlosky and colleagues (2013) reviewed ten common study techniques and rated two, practice testing and distributed practice, as having high utility; they rated elaborative interrogation (asking why a fact is true), self-explanation and interleaved practice as having moderate utility; and they rated summarization, highlighting, the keyword mnemonic, imagery for text and rereading as having low utility. Several of the low-utility techniques are exactly what a student does when an assistant summarizes a reading and the student rereads the summary.

Table 4.2. Five tutoring techniques grounded in learning science, with the research evidence for each and how to request it from an assistant.

Retrieval practice in more detail

Retrieval practice is the deliberate recall of information from memory, for example by answering questions or writing down what one remembers without looking. Roediger and Karpicke (2006) asked students to read short prose passages and then either reread them or take a recall test without feedback. On a test five minutes later, rereading looked better: in their first experiment, students recalled 81 percent after restudying and 75 percent after testing. Two days later the order reversed, at 68 percent for students who had been tested and 54 percent for those who had restudied, and after one week the figures were 56 and 42 percent. In a second experiment, students who studied a passage repeatedly predicted that they would remember it better than students who had been tested repeatedly, although they recalled less a week later. Rereading feels effective because the material is familiar; retrieval feels harder and works better. An assistant that quizzes the student turns this finding into a routine.

Because a general assistant's default is to answer, each of these techniques has to be asked for. The next part sets out prompts that set up a tutoring conversation built on them.

Setting Up a Tutoring Conversation

A tutoring prompt applies the parts of an effective request from Lesson 3 (role, task, context, format and constraints) to the goal of learning. The prompts in Box 4.3, Box 4.4, Box 4.5 and Box 4.6 are written in full so that they can be copied, adapted and saved in a prompt library. Words in square brackets are replaced with the student's own details.

Box 4.3 sets up a Socratic tutor that asks one question at a time and withholds answers until the student has made an attempt.

Each instruction has a purpose. Asking one question at a time keeps the student generating answers. Withholding solutions until an attempt reproduces the design of the tutors that helped or did no harm in Section 1. The hint ladder (a hint for the next step only) keeps most of the reasoning with the student. The instruction not to praise wrong answers counters sycophancy, the tendency of assistants to agree with their users, which Lesson 3 introduced. Ending with the student's own explanation builds in a test of understanding, and the request to flag uncertainty gives the student a route to check the tutor against course materials.

Box 4.4 asks for a single concept to be explained at three levels of background, which applies the second technique in Table 4.2.

The tabs show a three-level explanation of confounding of the kind this prompt produces, written for this lesson and checked against HSCI 230 Lesson 11, which teaches confounding in full.

Box 4.5 turns a passage from the student's own study materials into practice questions for retrieval practice, with an answer key that can be checked against the passage.

The answer key is the weak point of any generated quiz. A wrong key teaches the wrong answer with the authority of a marked test, so the quotation requirement exists to make the key checkable. The student answers every question first, then compares each keyed answer with its quotation and with the passage. Any key whose quotation does not support it, or whose quotation does not appear in the passage, is treated as wrong until resolved.

Box 4.6 applies spacing and interleaving by mixing questions on topics studied at different times, with more questions on the older topics.

Figure 4.2 returns to Prompt 1 and shows the loop it sets up, in which each attempt by the student determines how much help the assistant gives.

Prompts 1 to 4 place tutoring behaviour in a request that the student writes and can save. The next part describes the learning modes that several assistants added in 2025, which apply similar behaviour without a custom prompt.

Learning Modes Built into Assistants

In 2025, several assistants added modes that apply tutoring behaviour without a custom prompt. Anthropic introduced a learning mode for Claude in April 2025 for its education offering and made it available to all Claude.ai users in August 2025; it guides students toward their own answers with Socratic questions. OpenAI launched study mode in ChatGPT on 29 July 2025; it uses Socratic questions, hints, scaffolded explanations and knowledge checks, and it was built from custom system instructions written with teachers, scientists and experts in pedagogy. Google added Guided Learning to Gemini in August 2025; it breaks problems into steps, adapts explanations, and uses images, video and quizzes, and it draws on Google's education-focused LearnLM models. All three were available in 2026, and all can be switched on and off within a conversation.

These modes encode the same principle as Prompt 1, and they have limits. OpenAI stated that study mode relies on system instructions and that its behaviour may therefore be inconsistent across conversations, and its help pages note that it may still give direct answers at times. A learning mode responds to the student's requests, so a student who presses for the answer will usually get it. A saved custom prompt states the student's own level, goals and rules and works across tools, which suits a topic the student will return to over several sessions; a learning mode is convenient for a quick session.

The flip cards that follow recall four instructions from the tutoring prompts in this section.

Worked Example 4.1 follows Jordan through a tutoring session on confidence intervals with a general assistant and a practice quiz generated with Prompt 3, and records the checks Jordan applied to each output.

Box 4.7 asks for a weak request to be rewritten as a tutoring prompt using the principles in this section.

Tutoring prompts make an assistant ask, hint and check, and Worked Example 4.1 shows that the student still has to check the tutor. Section 3 applies the same approach to a different task: reading research articles, textbook chapters and statistical output that are difficult on first contact.

Why Research Reading Is Hard, and Where AI Helps

Health sciences students read material that was not written for them. A research article assumes familiarity with a field's vocabulary, methods and statistical conventions; a textbook chapter introduces many new terms at once; statistical output prints numbers without saying which ones matter. Older articles add a further difficulty, because conventions of writing and analysis have changed. Students respond to this difficulty in predictable ways, by rereading, highlighting or skipping to the abstract, and Section 2 noted that these are among the least effective study techniques (Dunlosky and colleagues, 2013).

An assistant can lower the barrier at each stage of reading. Before reading, it can produce a guide that orients the reader to the question, design and structure of the text. During reading, it can define terms in context and explain a table line by line. After reading, it can produce a summary for the student to check, or questions for the student to answer. The principle from Section 1 still applies: the student does the reading and the interpreting, and the assistant prepares the student and checks what was understood. A summary read in place of the article gives the student the shape of the content without the understanding that comes from working through it.

Figure 4.3 sets out these three phases and the parts that the assistant and the student play in each.

The worked example in this section uses a paper that HSCI 230 Lesson 4 presents as a landmark case-control study and walks through scene by scene in an interactive story: Doll and Hill (1950), "Smoking and carcinoma of the lung: preliminary report", published in the British Medical Journal. The paper's full text is freely available online. It is a good test of reading with AI because it is short and important, it reports its results as percentages and probabilities in the conventions of 1950, and it uses vocabulary that a modern student may not recognize.

The next three parts follow the phases in Figure 4.3 in order, beginning with a reading guide prepared before reading.

Before Reading: A Reading Guide

A reading guide is a short orientation prepared before reading: what question the text answers, how it is organized, which parts deserve close attention, which terms will be unfamiliar and which questions to keep in mind. Guides of this kind have long been used by teachers; an assistant can produce one for any text in a minute. Box 4.8 asks for a guide that withholds the findings, so that the student discovers them by reading.

Attaching the article matters. An assistant asked about a famous paper without the text will answer from patterns learned in training, and Lesson 2 explained why such answers can be fluent and wrong. With the text attached, the assistant works from the source, and its claims can be traced to it.

Worked Example 4.2 shows the guide that Prompt 5 produced for Doll and Hill (1950) and the checks Jordan made before reading on.

A guide of this kind orients the reader without supplying the findings. The next part turns to the help an assistant can give while the reading is under way.

During Reading: Terms and Tables

During reading, an assistant can define terms as they are used in the text and explain a table line by line. The two subsections below take these tasks in turn, with a table from Doll and Hill (1950) as the example for the second.

Defining terms in context

A dictionary definition and a definition as used in a particular text can differ. "Control" means something different in a case-control study, a randomized trial and a quality-control process, and older papers use words such as "almoner" that a modern dictionary may define without explaining their role in a hospital study. A useful request names the text: "Define these five terms as they are used in the attached article, and tell me if any is used differently from how a current epidemiology textbook would use it." Hovering over a term in a series lesson shows a short definition, and Lewis can explain a course term from the course text (Section 4); an assistant extends the same help to readings from outside the course.

Explaining a table and recomputing its numbers

Tables carry most of the evidence in a quantitative paper and are often the hardest part to read. Table 4.3 reproduces the counts from the paper's comparison of non-smokers and smokers among patients with lung carcinoma and their controls (Doll and Hill, 1950). A smoker was defined as a person who had smoked as much as one cigarette a day for as long as one year.

Table 4.3. Non-smokers and smokers among patients with lung carcinoma and their controls, by sex, in Doll and Hill (1950).

Jordan asked: "Explain this table row by row, say what comparison it is designed to show, and calculate the odds ratio for smoking separately for men and women, showing each step." The assistant explained that each row counts non-smokers and smokers in one group, that the comparison of interest is the proportion of non-smokers among patients against that among their controls, and that non-smokers were rarer among patients of both sexes. It then calculated the odds ratios. For men, the odds of having smoked among patients divided by the odds among controls is (647 × 27) / (2 × 622) = 17,469 / 1,244 = 14.0. For women, it is (41 × 32) / (19 × 28) = 1,312 / 532 = 2.5.

Jordan recomputed both values by hand with the cross-product formula taught in HSCI 230 Lesson 4 and obtained the same results. The explanation was accurate, but it left out two points that matter for interpretation. First, these are crude odds ratios calculated by the reader; the paper itself does not report odds ratios. Second, the estimate for men rests on only two non-smoking patients, so it is imprecise. Jordan asked for a 95 percent confidence interval by the Woolf method, then checked it: the standard error of the log odds ratio is √(1/647 + 1/2 + 1/622 + 1/27) = 0.735, which gives an interval from about 3.3 to 59.3 for men. The same calculation gives 1.2 to 5.2 for women, an interval that excludes 1.0 in agreement with the probability the paper reports for women (between 0.01 and 0.02). Recomputing the numbers caught nothing wrong in this case; asking what the explanation had left out produced the most useful part of the exchange.

The same requests apply to statistical output from a student's own analyses. Box 4.9 asks for a line-by-line explanation of output that separates the numbers that answer the research question from those that are diagnostic or secondary.

Statistical output usually contains no personal information, but the data behind it may be confidential, and the classification rules in Lesson 3 apply before anything is pasted. HSCI 410 teaches the regression models whose output students most often need explained, and Lesson 8 of this course returns to output in the context of AI-assisted analysis in R.

Terms, tables and output are explained while the reading is under way. The next part turns to the last phase, checking a summary of the article against the original.

After Reading: Checking a Summary Against the Original

Summaries are where AI reading support most often goes wrong, because a summary compresses many claims into a few sentences and each compression is a chance to distort. The check is simple and slow: break the summary into separate claims, find each claim in the source, and classify it. Jordan asked for a 100-word summary of Doll and Hill (1950) after reading the paper, with the aim of checking it. The summary in Box 4.10 is a representative example prepared for this lesson, with errors of the kind that summaries commonly contain.

Table 4.4 breaks the summary into seven claims, sets each beside what the paper says and gives a verdict on each.

Table 4.4. Claim-by-claim check of the AI-generated summary of Doll and Hill (1950) against the paper, with a verdict for each claim.

Four of the seven claims survived the check unchanged, which is typical: most of a summary is usually right, and the errors are hard to spot without the source. The two wrong claims are both plausible simplifications. "Healthy adults from the community" is what many readers assume a control group to be, and "every man" rounds 647 of 649 to all of them, erasing the two non-smokers that make the men's estimate imprecise. The overstated conclusion converts the authors' measured wording into a stronger causal claim; the distinction matters because the paper presented its results as preliminary, and the causal case was built over the following years by further studies, including the cohort of British doctors that HSCI 230 Lesson 5 describes. A summary that passes this check can be kept in study notes, with the corrections marked.

Box 4.11 sets out a prompt that asks an assistant to carry out the same claim-by-claim check.

An assistant can run this check as well as a student can, and asking it to do so is a good use of AI. Its verdicts are themselves claims, however, and the quotations it gives must be found in the source by the student. A student who writes the summary first, from memory, and then asks for the claim check gains twice: the summary is retrieval practice, and the check is feedback on it.

Reading guides, explanations of tables and summary checks all concern research articles. The next part adapts the approach to textbook chapters.

Textbook Chapters

Textbook chapters call for a different approach from research articles because their aim is to teach a set of connected ideas. Three requests work well. Asking an assistant to turn the chapter's headings into questions before reading gives the reading a purpose. Asking for a concept map, a list of the chapter's main ideas with the relationships between them, after reading shows how the ideas connect, and comparing it with a map drawn by the student shows what the student missed. Asking to be quizzed on the chapter, with Prompt 3 from Section 2, turns the reading into retrieval practice. Each request keeps the student reading the chapter itself.

These requests can be made of any general assistant. The next part describes notebook tools, which answer only from documents the user uploads.

Notebook Tools

A notebook tool is an AI application that answers questions and produces study materials only from documents the user uploads, citing the passages it used. A widely used example in 2026 is Google's Gemini Notebook, which was called NotebookLM until July 2026. It accepts PDFs, websites, videos and documents; answers questions with inline citations that link to the source passages; and produces study guides with key terms and practice questions, flashcards, quizzes, mind maps, short video overviews and Audio Overviews, in which two AI voices discuss the uploaded material in the style of a podcast. General assistants offer related features through project folders that hold uploaded files (Lesson 3).

Grounding answers in uploaded documents, a form of the retrieval-augmented generation described in Lesson 2, reduces fabrication and makes checking fast, because each claim links to the passage it came from. It does not remove error: the tool can still misread a passage, combine two passages into a claim that neither makes, or omit an important qualification. Its outputs also differ in learning value. Flashcards and quizzes generated from the readings support retrieval practice. Audio and video overviews are pleasant ways to preview or revisit material, but listening to a summary is closer to rereading than to retrieval, and it works best when followed by questions the student answers unaided.

The accordion that follows answers three practical questions: what to upload to a notebook tool or assistant, what to do when an assistant cannot read a file properly, and how this section relates to Lesson 5.

Box 4.12 returns to the counts in Table 4.3 and applies the claim check to a single sentence of the kind an assistant might write about them.

Reading with AI follows the same pattern as tutoring with AI: the assistant prepares, explains and checks, and the student reads, calculates and verifies. Section 4 brings tutoring and reading together with the series' own resources into a personal learning system, and considers when studying without AI serves learning better.

The Series' Own Study Tools

A personal learning system combines the tools a student has into a routine that follows the principles of Sections 1 to 3: prepare, attempt, retrieve, get feedback and return later. Every lesson in this series already provides several tools designed around those principles. Knowing how each one works, and what it cannot do, lets a student use it for the job it does best.

Lewis, the study assistant

Lewis is the pop-up study assistant that opens from the "Get Help" button in the bottom-right corner of each lesson page. Lesson 2 used him as a worked example of retrieval-augmented generation, in which a system first retrieves relevant text and then passes it to a language model to answer from. When a student asks a question, Lewis ranks passages from the text of the series' lessons by keyword relevance, favouring the lesson and section being read. He passes the best few passages, up to four, to a small open-weight language model that runs inside the student's own browser, with instructions to answer only from those passages and to cite them. Each citation appears as a button that opens the cited section and highlights the passage. When no passage matches the question well, Lewis replies that the course materials do not cover it, so he declines questions outside the course in place of guessing.

Three features make Lewis a study tool as well as a question-answering one. Quiz me asks one question drawn from the section being read or from a topic the student names, and does not give the answer with the question. When the student replies, Lewis first marks the answer as correct, partly correct or not yet correct and says what is right and what is missing, and only then gives a model answer; the student can ask for a hint, which keeps the question open. Summarize this section summarizes the narration of the section being read. Explain a term starts a definition question. When a student pastes a question from a knowledge check, Lewis responds with a hint in place of the answer.

Lewis also has practical features and limits that are worth knowing. No server is involved: the model runs on the student's device, and questions and answers stay in the student's browser. The first use requires a one-time download of the model, between about half a gigabyte and about two and a half gigabytes depending on the model size chosen, and the start screen states the size before anything is downloaded. Browsers that lack the graphics support the model needs receive a search-only mode that lists the matching passages with links. A copy of each conversation is saved as a PDF to the student's Downloads folder when the chat is closed. Because the model is small, it can misstate a passage or combine two passages into a claim that neither makes, and it can draw on courses the student has not taken, which it labels by course and lesson. The citation buttons exist so that every answer can be checked against the text it came from.

Narrated slide decks

The welcome page and each section of a lesson open with a short narrated slide deck that walks through the section's main ideas. The narration can be played and paused, a transcript can be opened, and autoplay and automatic advance to the next slide can be switched on or off. A deck works well as a preview before reading, because it gives the reader a structure on which to hang the detail, and as a quick review a few days later. Watching a deck is closer to rereading than to retrieval, so its value for review rises when the student pauses before each slide and tries to state what it will say.

Office Hours podcasts

Each lesson has a companion episode of Office Hours, a two-voice podcast meant for listening after the lesson. In each episode of about twenty minutes, Sarah and Kiffer add perspective and critique, work through the questions that students tend to find most difficult, and add further worked examples, including some that are harder than those in the lesson. Each episode page has the audio with a visual companion and captions, a summary, a list of chapters and a downloadable transcript. The podcast suits a walk or a commute, and its harder examples are best treated as problems: pausing at the point where a question is posed and answering it before the hosts do turns listening into practice.

Knowledge checks, reflections and the glossary

Each section ends with a reflection, which reveals a model answer once the student has saved their own, and a four-question knowledge check with explanations; each lesson ends with a final assessment. These are retrieval opportunities, and they lose most of their value if they are attempted with Lewis, the reading or an assistant open. Glossary terms show a short definition when the pointer rests on them, and the glossary page collects all the terms, people and frameworks of the lesson.

Table 4.5 summarizes the best use of each resource in study and what to watch for when using it.

Table 4.5. The series' study resources and a general assistant, with the best use of each in study and what to watch for.

Each resource is most useful at a particular point in a sequence of study. The next part assembles them into a weekly routine.

A Weekly Study Routine

A routine matters more than any single tool, because the strongest techniques in Section 2, spacing and retrieval, depend on returning to material at intervals. Figure 4.4 and Table 4.6 show one way to study a lesson over about a week. The timings are suggestions to be adapted; the sequence of preview, reading, unaided retrieval, tutoring on the hardest idea, and spaced review is the part that carries the evidence.

Table 4.6 lists the same routine day by day, with the role of AI and the unaided step at each stage.

Table 4.6. A weekly routine for studying one lesson, giving the activity, the role of AI and the unaided step on each study day.

The routine uses AI at every stage and keeps one unaided step in each. The unaided steps are where learning is measured and much of it is produced, and the AI steps supply orientation, feedback and an endless supply of questions.

The unaided steps raise a broader question about the occasions on which an assistant is best set aside, which the next part addresses.

When to Study Without AI

The advice in this part is about how learning works, and it applies to studying in any subject. There are occasions on which an assistant, even one set up as a tutor, gets in the way of what the student is trying to build. Recognizing them is part of using AI well.

The first occasion is building first fluency with a new method. The first few times a student calculates an odds ratio from a 2 × 2 table, sets up a confidence interval or writes a line of R, the struggle with each step is how the method enters memory. Once the steps are fluent, an assistant can take over routine parts with little loss, which is why Lesson 8 can teach AI-assisted analysis in R to students who have met the basics. Working the first examples by hand, and only then checking them with an assistant, gives the student both the skill and the speed.

The second occasion is retrieval practice itself. A self-test, a knowledge check or a blank-page summary works because the student pulls the material from memory. Any tool open during the attempt, Lewis included, turns retrieval into lookup. The assistant's place is after the attempt, to mark it and explain what was missed.

The third occasion is preparing for invigilated examinations in courses that use them. Practice works best under conditions like those in which the knowledge will be used, and Bastani and colleagues (2025) showed that performance with an assistant can greatly overstate performance without one. Some practice under examination conditions, timed and closed-book, gives the student an accurate reading of where they stand while there is still time to act on it.

The fourth occasion is working out what one thinks. A first attempt at an argument, an interpretation of a finding or a reflection is the point at which a student discovers what they understand and believe. An assistant's draft arrives fluent and complete and tends to replace the student's own thinking. Lessons 6 and 7 develop ways to use AI for writing and for testing ideas once the student has a position of their own.

The accordion that follows describes signs that an assistant is doing too much of the thinking and takes up two common questions about studying without AI.

The four occasions mark the points at which the student's own attempt matters most. The next part considers how a student can judge whether that attempt reflects real understanding.

Judging Your Own Understanding

Students judge their learning mainly by how material feels, and the feeling is an unreliable guide. Rozenblit and Keil (2002) asked people to rate how well they understood how everyday devices work, then to write a detailed step-by-step explanation, and then to rate their understanding again. Ratings dropped after the attempt to explain, and the authors named the gap the illusion of explanatory depth: people believe they understand causal mechanisms in more detail than they do, until they try to explain them.

The same pattern appears in classrooms. Deslauriers and colleagues (2019) randomly assigned students in introductory physics at Harvard to active-learning classes or to polished lectures by highly rated instructors on the same content. Students in the active classes learned more, as measured by a test, but they felt they had learned less, apparently because they mistook the effort of active learning for a sign of poor learning. Roediger and Karpicke (2006) found that students who restudied predicted better recall than students who had been tested, and recalled less. Bastani and colleagues (2025) found that students who had practised with a general assistant did not perceive that they had learned less.

A clear AI explanation creates exactly the conditions for this illusion. It is fluent, well organized and pitched at the reader's level, and reading it produces a sense of understanding that has not been tested. The sense is real, and it is a feeling about the explanation; whether the reader could produce the explanation is a separate question that only a test can answer.

Tests that do not depend on feeling

Four tests give a more accurate reading. The first is the blank-page explanation: close every source and write an explanation of the idea, as if to a classmate who missed the class. The second is application to a new case: solve a problem with different numbers or a different context from any example seen. The third is prediction: before checking an answer or turning to the next slide, state what it will be. The fourth is calibration: before a knowledge check, predict the score, and after it, compare the prediction with the result. A student whose predictions are regularly higher than their scores has learned something about their own judgement.

An assistant is a useful examiner once the unaided attempt exists. Box 4.13 asks it to check an explanation without replacing it.

The assistant's feedback in turn needs checking against course materials, as Worked Example 4.1 in Section 2 showed. The value of the prompt lies in its order: the student's own attempt comes first, the gaps are presented as questions for the student to answer, and the model answer comes last.

With tests of understanding in place, the last part considers how the system of tools, prompts and routines is kept in use and improved.

Keeping the System Going

A learning system improves with use. A practical next step is to try one tutoring session on a topic from another course, once with a built-in learning mode and once with Prompt 1, and to note which produced the better explanation at the end. Prompts 1 to 8 in this lesson can be saved to the prompt library introduced in Lesson 3 and refined as they are used. Lesson 10 brings the prompt library, the verification routine and the AI use log together into a personal workflow, and Lesson 5 applies the reading techniques of Section 3 to searching for and reading evidence.

Box 4.14 brings the lesson together, and Table 4.7 within it sets out how AI performs at six study tasks and the check that applies to each.

The reflection and knowledge check that follow are retrieval opportunities of the kind this lesson recommends, and they are best attempted before Lewis, the reading or any assistant is opened.

Bringing It All Together

This lesson began with two randomized studies that point in opposite directions. An AI tutor built on teaching principles helped Harvard physics students learn more in less time than an active-learning class (Kestin and colleagues, 2025), while unrestricted access to a general assistant raised high school students' practice scores and lowered their scores on a later examination taken without it, and a hint-giving version largely avoided the harm (Bastani and colleagues, 2025). Read together, the studies support a single working principle: AI supports learning when it prompts the student to think and weakens learning when it does the thinking. The generation effect, the testing effect and research on desirable difficulties explain why.

Sections 2 and 3 turned the principle into practice. Tutoring prompts make an assistant ask, hint and check before it explains; explanations at three levels meet the student where they are; practice questions are useful once their answer keys have been checked against the source. For reading, an assistant prepares a guide that withholds the findings, defines terms in context and explains tables, and its summaries are checked claim by claim against the original, as the Doll and Hill (1950) example showed.

Section 4 combined these techniques with the series' own resources, Lewis, the narrated decks, the Office Hours podcasts and the knowledge checks, in a routine that uses AI at every stage and keeps an unaided step in each. Because a clear explanation produces a feeling of understanding that a test may not confirm, the lesson ends with ways to judge understanding directly: explaining from memory, applying ideas to new cases, predicting and calibrating.

The final reflection asks you to plan how you would learn one concept with AI, using the principles and practices from all four sections.

Final Knowledge Assessment

Terms, tools and people introduced in this lesson on using AI to learn, study and read.