Chatting with notes: how to check your understanding

Before you start

This lesson is designed for secondary students in Years 7–12. It will take approximately 25 minutes to complete. Before you begin, you should have a basic understanding of the difference between passive recognition (such as re-reading or highlighting) and active production (such as retrieval practice and unassisted timed sits). You will only need a notebook, a pen, and a timer for this lesson.

Related workflows: sit a new mock or section drill, see focus areas on your improvement plan, pick your next practice test type.

What you'll learn

• Identify the cognitive risks of using AI "chat with notes" tools for exam preparation.
• Explain the difference between "borrowed fluency" and genuine long-term memory encoding.
• Apply the 40-Minute Active Recall Block to convert passive AI summaries into active retrieval cues.
• Design a safeguarded AI study prompt that promotes desirable difficulties instead of cognitive offloading.

Core concepts

The Illusion of Competence and Borrowed Fluency

When you upload your class PDFs to an AI tool and start chatting with your notes, the experience feels incredibly organized and efficient. The AI instantly summarizes chapters, finds specific quotes, and answers your questions with perfect clarity. However, this ease creates a psychological trap known as the illusion of competence. Because the AI is highly fluent, your brain mistakes the clarity of the AI's response for your own understanding. This is called borrowed fluency. In reality, because your brain did not have to struggle to find, organize, or retrieve the information, very little of it is encoded into your long-term memory.

The Scientific Evidence Against Unrestricted AI Chat

Recent learning-science research warns us about the dangers of relying on unrestricted AI search and chat tools during study. A landmark study by Bastani et al. (PNAS 2025) found that students who had unrestricted access to standard GPT-4 during practice scored 17% lower on unassisted math exams compared to students who studied with no AI at all. While the AI boosted their practice performance by 48% during study, it acted as a cognitive crutch. When the crutch was removed in the exam, the students failed. Furthermore, researcher Barcaui (2025) demonstrated an 11% retention deficit (57.5% vs 68.5%) on a surprise test 45 days after studying with unrestricted AI compared to traditional active methods. This occurs because of cognitive offloading—the act of outsourcing our thinking to external tools, which reduces our brain's deep processing. Gerlich (2025) also identified a strong negative correlation (r = -0.68) between high AI dependency and critical thinking benchmarks, proving that the more we let AI think for us, the less analytical we become.

Desirable Difficulties and Cognitive Load

To truly learn, your brain must experience what cognitive scientists call desirable difficulties. This concept, aligned with the NSW Education Standards Authority (NESA) Cognitive Load Theory guidelines, explains that long-term retention requires effortful processing. When you struggle to recall a definition or solve a math problem without help, you strengthen the neural pathways associated with that knowledge. Before you sit down to sit a new mock or section drill, you must ensure your study methods build independent recall rather than relying on a digital assistant.

The Verification Gap

Another major risk of the "chat with notes" method is the verification gap. If you rely on an AI to summarize your knowledge base, you will struggle to spot subtle AI hallucinations or factual errors. Because you offloaded the initial reading and struggle phase, you lack the deep domain knowledge required to verify if the AI's output is actually correct. To target your weak spots safely, you can identify focus areas on your personal improvement plan and use AI as a feedback tool rather than an answer generator.

Four Rules for Learning Better with AI

If you want to use AI without damaging your exam performance, follow these four rules: First, never let AI write your first draft or solve a problem first. Second, use AI for feedback, not answers, by prompting it to act as a Socratic tutor. Third, close the loop with unassisted sits where you work entirely offline. Fourth, convert AI summaries into active retrieval cues rather than just reading them. If you are preparing for high-stakes entrance exams, you should select your next practice test type early to align your retrieval practice with these rules.

Worked examples

Example 1

A student uses an AI "second brain" to chat with 50 pages of biology notes. They ask the AI to "summarize the process of cellular respiration" and read the beautiful bulleted list. The student feels highly prepared but fails a surprise closed-book vocabulary test on the Krebs cycle the next day. Explain why this study session failed and how to fix it.

1. Analyze the failure: The student engaged in passive recognition. Reading an AI-generated summary did not require the brain to retrieve information from memory. The AI did the cognitive work of organizing the steps of cellular respiration.
2. Identify the cognitive offloading: By asking the AI to find and summarize the steps, the student bypassed the "desirable difficulty" of active recall, leading to borrowed fluency.
3. Redesign the study loop: The student should first close the AI tool. They should write down everything they can remember about cellular respiration on a blank sheet of paper. Then, they can use the AI to compare their handwritten attempt against the source material, identifying missing steps.

Example 2

A student preparing for the HAST exam uses AI to write practice essays. They feed the prompt into ChatGPT, read the generated essay, and try to memorize its structure. During the actual 25-minute unassisted writing test, they freeze and cannot complete the essay. Explain why this occurred and how to apply the Formative Writing Loop.

1. Analyze the issue: The student bypassed the planning and drafting phase. Writing an essay requires active retrieval of vocabulary, structural planning, and real-time decision-making. Memorizing a pre-written AI essay does not build these active skills.
2. Apply the Formative Writing Loop: On Monday, the student must write a 30-minute draft completely independently, without any AI assistance.
3. Use AI for feedback: On Monday night, the student can input their independent draft into the AI and ask for feedback against specific rubric criteria. On Tuesday, they spend 20 minutes rewriting only the lowest-scoring sections of their essay.

In tests and exams

• Closed-book multiple-choice questions: These questions require precise definition recall. If you have only practiced by chatting with your notes, you will experience the illusion of competence and fail to distinguish between similar-looking terms under exam pressure.
• Extended response and essay writing: Examiners reward independent structural cohesion, logical flow, and unique voice. Relying on AI-generated outlines leaves you unable to generate complex arguments independently within strict time limits.
• Multi-step mathematical problem-solving: Exams require you to justify intermediate steps without external prompts. Unrestricted AI practice prevents you from developing the resilience needed to navigate difficult, multi-step problems when you get stuck.

Practice

1. Warm-up: Review this study diary entry: "Spent 90 minutes chatting with my history PDF. I asked the AI to find the main causes of World War I and copy-pasted its summary into my study guide." Identify the primary active-learning risk in this session and suggest one immediate change.
2. Warm-up: A student has a list of 20 French vocabulary words. Instead of using flashcards, they ask an AI to write a story using those words so they can read it. Explain why this is a passive study method.
3. Standard: Write a Socratic prompt that you can use with an AI tool to study quadratic equations without triggering the 17% exam score drop identified by Bastani et al. (2025).
4. Standard: Suppose a student currently scores 75% on their unassisted physics practice tests. If they switch to a study method that relies entirely on unrestricted AI chat for homework assistance, calculate their predicted unassisted exam score based on the PNAS 2025 study's findings. Show your working.
5. Challenge: Design a 40-minute active recall study block for a complex science topic of your choice. Your plan must include the timeline segments, the specific active retrieval tasks, and how you will safely use AI at the very end of the block.

Answers

1. Warm-up Answer: The primary risk is cognitive offloading and borrowed fluency. The AI did the active work of extracting and summarizing the causes of World War I. To fix this, the student should close the AI, write down the causes of the war from memory, and then use the AI only to check for missing details.
2. Warm-up Answer: Reading a story is a recognition-based task. The brain does not have to actively retrieve the translation or meaning of the French words from memory. To make it active, the student must write their own sentences using the words first, then ask the AI to correct their grammar.
3. Standard Answer: A safe Socratic prompt would be: "Act as a high school math tutor. Do not give me the answers to my quadratic equation problems. Instead, look at my working, point out any errors in my steps, and ask me guiding questions to help me find the correct path myself."
4. Standard Answer: The Bastani et al. (2025) study showed a 17% decline in unassisted exam performance for students using unrestricted AI. To calculate the predicted score: 75% * (1 - 0.17) = 75% * 0.83 = 62.25%. The student's predicted exam score would drop to approximately 62%.
5. Challenge Answer: A highly effective 40-minute block is structured as follows: 10 minutes mapping the topic structure from memory on a blank page (Outliner phase); 20 minutes writing down explanations for each node without looking at notes (Active Retrieval); 10 minutes using AI to compare the student's written explanations against the textbook syllabus, highlighting gaps and generating one targeted correction.

Common mistakes

• Mistake: Chatting with a PDF to find answers instead of reading the text.
  Fix: Use an outline to map the unit structure first, then close the document and retrieve the details from memory.
• Mistake: Assuming a clean AI-generated summary means you understand the topic.
  Fix: Promote key concepts into active recall flashcards and rate your unassisted flips to measure actual retention.
• Mistake: Practicing only with open-book AI hints or continuous chat guidance.
  Fix: Run timed, unassisted mocks to prove that you can transfer your knowledge under real exam conditions.

Quick recap

• Chatting with notes creates "borrowed fluency" and an illusion of competence without building long-term memory.
• Unrestricted AI use during study can lead to a 17% drop in unassisted exam scores due to cognitive offloading.
• True learning requires desirable difficulties, meaning you must actively retrieve information rather than passively reading summaries.
• Always attempt problems and write drafts independently before using AI tools for feedback.
• Use AI as a Socratic tutor that provides hints and identifies errors, rather than an answer machine.

To ensure you don't fall into the AI chat trap, keep a log of the questions you missed during your unassisted practice sessions. Maintain a dedicated mistake book for spaced retry and build the independent recall skills needed to excel under real exam conditions.

Related workflow

Open your Mistake Book for spaced retry