AI as Learning Companion
A free 90-minute course for university students, postgraduates, and researchers who want AI to make them smarter learners - not just faster typists.
🎓 Who is this course for?
This course is for anyone studying at university level and beyond: undergraduates, postgraduates, and PhD researchers. It assumes you already use AI tools occasionally and want to use them with more intention, depth, and confidence in your actual academic work.
The aim: smarter learning
This course is built around one idea - smarter learning. Not learning less by outsourcing it to AI, but learning better: deeper understanding with visible mechanics, faster orientation in new material, longer retention, better grades, and wider, more precisely sourced research.
By the end, you will be able to do five things you likely cannot do confidently today: see a concept from its big picture down to its working parts using AI-generated analogy and illustration; turn AI into an honest feedback and self-testing partner instead of a flattering one; use AI to bring genuine originality to coursework; ground your research in real, traceable sources using purpose-built tools; and run a ready-made AI learning pipeline for whatever task is in front of you.
What you will learn
- How AI actually generates an explanation - and where sycophancy and hallucination quietly creep in
- How to move from the big picture of a concept down to its mechanism, using analogy and AI-generated illustration
- How to turn AI into a rubric-aware feedback partner and an unseen-exam revision engine
- How to use AI to bring originality to coursework, without it doing the thinking for you
- How to ground research in real sources with NotebookLM, Consensus, and Elicit
- How to build a reusable AI learning pipeline for any subject or task
This course is the second in StudyAnalyst's AI literacy pathway. It builds directly on Critical AI Literacy: Using LLMs Responsibly - if you have not taken it yet, several ideas here (hallucination, sycophancy, the jagged frontier) are explained there in more depth. You do not need to have taken it first, but it is a useful companion.
🔒 A note on AI policy
This course shows you what is technically possible and pedagogically useful. It is your responsibility to check your own institution's policy on AI use in assessed work before applying anything here to a real assignment, dissertation, or thesis. When in doubt, declare your AI use transparently.
How Do You Learn With AI Right Now?
Five quick questions - no right or wrong answers. This helps you see your starting point before the lessons begin.
1. When you do not understand a concept, what do you currently do?
2. Have you ever asked AI to check or mark your work, and felt it was a bit too generous?
3. How do you currently use AI for finding academic sources or literature?
4. When a piece of coursework asks for an "original" contribution, what is your honest approach?
5. Do you have a repeatable way of using AI for a given type of task (revision, essays, data analysis) - or does it vary every time?
Remember: this is just a starting point. By the end of this course, you will have a method for all five areas.
The AI Learning Companion Framework
Study is not one task. The way AI helps you read is different from how it helps you revise, give feedback, or find sources. This framework gives you a map to navigate the difference.
Two frameworks, one pathway
Our first course, Critical AI Literacy, introduced a framework for how AI skills grow: from asking simple questions (Level 1) through using AI as a daily partner (Level 2) toward building workflows and systems (Level 3 and beyond). This course introduces a companion framework: how AI supports learning specifically. The two sit side by side.
The key insight from the learning research is that understanding cannot be automated. AI can speed up support. It cannot replace the process of building knowledge. Each of the six learning moves below is something you must still do - AI changes how fast and how well you can do it, not whether you need to do it at all.
Six learning moves
Different study tasks call for different kinds of AI help. The six moves below cover the main things you do as a student. The lessons in this course follow this sequence:
Read & Orient
Get the map of unfamiliar material before deep reading. Decode structure, key terms, and central claims.
Understand & Explain
Move from big picture to mechanism through analogy, examples, and illustration.
Practise, Get Feedback & Originate
Turn knowledge into use through self-testing, rubric-aware feedback, and original angles.
Research & Source
Ground claims in real, traceable sources. Use purpose-built tools rather than ungrounded chat AI.
Plan & Pipeline
Build a repeatable workflow for each recurring study task, with verification built in.
Know the Limits
Sycophancy, hallucination, and the citation trap appear across every learning move - covered first so you carry the habit throughout.
Four levels of AI support in learning
For each learning move, AI can help at different levels of depth and responsibility. Most learners start at Level 1. The practical target for this course is Level 2 across the main moves, with a clear view of what Level 3 involves.
Quick help
You ask for an answer, definition, summary, or example. Fast and useful for getting unstuck.
Main risk: accepting fluent answers without checking.
Study coach
You use AI to test, challenge, and improve your learning. AI compresses feedback loops.
Main risk: feedback that is too agreeable or too easy.
Learning workflow
You build repeatable steps for a whole study task - a reusable pipeline for reading, revision, or research.
Main risk: errors can compound across steps unnoticed.
Personalised learning system
You connect tools, prompts, and records into a system used over time across subjects.
Main risk: a system that looks impressive but weakens your own judgment.
The full framework at a glance
🔎 What to carry from here
As you work through the lessons, you can return to this framework any time. Each lesson focuses on one move - but the same sycophancy and hallucination risks from Lesson 1 run through all of them. The framework also gives you a language for self-assessing: after each lesson, you can ask yourself which level you are currently at for that move, and what it would take to reach Level 2.
How AI Actually Thinks
By the end of this lesson you will be able to:
- Explain why a fluent AI explanation is not automatically a correct one
- Recognise sycophancy and sloth - the two cognitive risks that matter most for learning with AI
- Understand why claiming AI output as your own means taking responsibility for verifying it
A companion, not an oracle
This course treats AI as a learning companion: something that sits alongside your own thinking, expands what you can do, and speeds up how you do it. A companion is useful precisely because it is present throughout the work - but a companion can also be wrong, agreeable, or out of its depth, and you are still the one responsible for the outcome.
If you have taken StudyAnalyst's Critical AI Literacy course, you have already met the basic mechanics: AI generates the most statistically plausible next word given everything in the conversation so far, rather than retrieving a verified fact. It is fluent because it is trained on enormous amounts of human language. It is sometimes wrong because fluency and accuracy are not the same thing. This course assumes that foundation and builds on it directly - if any of this is new to you, that course is the right place to start in parallel.
🧠 The jagged frontier, for academic work specifically
AI can summarise a dense journal article in seconds and stumble on a simple logic check a moment later. For your studies, this means AI is extremely strong at explaining, illustrating, drafting, and restructuring - and much weaker at being reliably correct about specific facts, citations, and edge cases. Every lesson in this course leans into the first set of strengths while building in a check for the second.
Choose the safer default
For each academic task, decide whether it mainly uses an AI strength or needs human judgement to lead. Even AI-strength tasks still need a final check.
The limitation that matters most for learning: sycophancy
Hallucination - AI generating a fluent but false statement - is the limitation most people have heard of. There is a second, quieter limitation that matters just as much when AI is helping you learn: sycophancy.
💡 What sycophancy means in practice
AI models are designed to be a helpful friend. They are tuned to be agreeable, encouraging, and supportive - and that is genuinely useful much of the time. The problem is that AI does not know where the line is between being helpfully positive and being unhelpfully agreeable. It crosses that line without noticing. The result is a built-in tendency to validate your framing, praise your reasoning, and soften disagreement - even when your reasoning has a real flaw in it.
For studying, this is dangerous in a specific way: the AI explanation that confirms what you already believed feels like understanding. It is not the same thing. If an explanation, a piece of feedback, or a mark feels a little too easy or a little too kind, that is exactly the moment to slow down and verify it - against your lecture notes, the rubric, or a primary source.
The second cognitive risk: sloth
Sycophancy is about what AI says to you. Sloth is about how you respond to it. AI's training pulls it toward the statistical average of everything it has been trained on — which means its first response to any question tends to be competent, smooth, and unremarkable. The learner who stops there, accepts the first output without pushing further, and moves on, has effectively outsourced their thinking to the average of the internet. Their understanding stays shallow; their work looks like everyone else's.
💡 The antidote is not less AI - it is more deliberate prompting
The way past sloth is to refuse to accept the first average response. Refine the context. Ask a follow-up that targets the gap in what AI gave you. Push back on vague explanations and ask for a concrete example. An AI that has been pressed three times, with increasingly precise context, will give you something substantially different from what it gave someone who stopped at the first reply.
"Use AI to stress-test ideas, not just to confirm them." If every AI response you receive feels satisfying and complete, that is itself a warning sign - either the questions are too easy, or you are not pushing hard enough.
Every technique in this course is an answer to sloth. The four-layer decoding in Lesson 2, the iterative analogy method in Lesson 3, the rubric-specific feedback prompts in Lesson 4, the source-grounding in Lesson 5 - each one is a way of pushing AI past its default average and getting output that is genuinely matched to your specific question, your specific level, and the specific gap you are trying to close.
Hallucination, secondary hallucination, and the citation trap - a quick recap
Hallucination: AI generates a fact, date, statistic, or detail that sounds plausible but is invented.
Secondary hallucination: AI gives you a source about a piece of knowledge, data, or a claim - and that source does exist - but the specific fact or information AI attributes to it is not actually in that source. The source is real; the attribution is invented. These are the most difficult hallucinations to catch, because the reference checks out on the surface. We covered this in Critical AI Literacy; it is important to maintain that definition consistently here.
The citation trap: a broader pattern that includes secondary hallucination - AI produces a reference that looks completely real (plausible author, plausible journal, plausible year) but either does not exist, or exists but does not say what is claimed. We will return to this directly in Lesson 5.
When you claim the output, you claim the responsibility
Before AI tools existed, the process of finding and reading sources built verification in automatically. You encountered facts embedded in their original context. You judged source quality as you went. You corrected for bias by reading multiple perspectives. AI hands you a finished-looking output that skips all of that process. Which means five responsibilities that used to happen naturally now have to happen consciously - and the moment you put your name on an AI-assisted piece of work, you are implicitly claiming you did them.
Five responsibilities that ownership implies
1. Fact-checking. AI generates plausible language, not verified truth. Every claim is a hypothesis until you confirm it against a primary source.
2. Knowing the edges. AI is weakest at the boundaries of established knowledge — recent findings, contested theories, emerging fields. Knowing where to probe is itself a high-order skill.
3. Evaluating sources. AI cannot tell you whether a source is authoritative, contextually appropriate, or methodologically sound for your discipline. That judgement is yours.
4. Correcting for bias. AI reflects patterns in its training data — often skewed toward majority voices, Western contexts, and older, well-represented perspectives. Your discipline trained you to notice this; apply it.
5. Catching hidden misinformation. AI can state something false with complete grammatical confidence and no hedging. Surface fluency is not evidence of accuracy. You are the last checkpoint before it becomes your claim.
"If you didn't verify it, you didn't write it."
Where this course is going: six moves toward smarter learning
Each remaining lesson gives you one concrete way to use AI that takes you from where you are now to a measurably different way of learning:
- Lesson 2 - Reading, Orientation & Decoding: use the four-layer framework to orient yourself in dense material before deep reading
- Lesson 3 - AI for Understanding: move from the big picture of a concept down to its working parts, with analogy and AI-generated illustration
- Lesson 4 - Practice, Feedback & Originality: turn AI into a rubric-aware feedback partner, an unseen-exam revision engine, and a source of genuine originality in coursework
- Lesson 5 - AI for Research: ground your work in real, traceable sources with NotebookLM, Consensus, and Elicit
- Lesson 6 - Build Your AI Learning Pipeline: build a reusable workflow for whatever recurring task is in front of you
Sycophancy, sloth, and hallucination will resurface across all six lessons - not as a side note, but as something you actively check for at the point where each is most likely to appear: settling for the first answer rather than iterating, in feedback that feels too kind, and in sources that feel too convenient.
✍ Activity
Open any AI tool you have access to. Describe a concept from your course honestly, including the bit you are shakiest on, and ask it to mark your understanding out of 10 with specific gaps. Read the response and ask yourself: did it actually find a gap, or did it mostly reassure me?
Reading, Orientation & Decoding
By the end of this lesson you will be able to:
- Orient yourself in a dense reading before attempting to read it in full
- Use the four-layer decoding framework to unpack complex text systematically
- Use AI to guide a second, deeper reading rather than replace the first one
The first job is orientation, not memorisation
When you face a dense reading - a journal article, a textbook chapter, a policy document - the instinct is often to start at the beginning and read through to the end. This works for novels. It is a slow and inefficient way to tackle academic reading, because you are encountering every piece of information cold, with no sense of what matters or why.
A better starting point is orientation first: get the map before you try to navigate the territory. AI is exceptionally useful for this. Ask it to identify the central claim, the key terms, the structure of the argument, and the two or three things you must verify yourself before relying on the text. This costs two minutes and makes the reading that follows far more directed.
Orientation prompt - use before reading
"Here is a reading extract [paste text or upload the document]. Give me: (1) the central claim in one sentence, (2) five key terms I should understand before reading in full, (3) three questions I should be able to answer by the end of the reading, and (4) two claims I must check in the original text rather than take from a summary."
This is Level 2 use if you then actually do the reading - using the AI output to guide it rather than replace it.
The four-layer decoding framework
Once you are reading, the question is what level of the text you are working at. Research on structured learning suggests four distinct layers, each building on the one below. AI can help you work through them explicitly, which is faster and more reliable than hoping the meaning will emerge on its own.
Look at one text through four lenses
Select each layer to see which part of this short example it helps you notice. This illustrates why the layers should be read in order.
💡 Start at the top, work down
Most students start at Layer 4 - they highlight everything and try to memorise details without a clear mental structure to attach them to. This is why revision feels slow and forgetting feels fast. Reversing the order - logic first, arbitrary details last - gives every detail a place to live in your understanding. AI can help you build the top layers quickly; the memorisation at Layer 4 is still yours to do.
Using AI across the four layers
You can ask AI to unpack a piece of text through all four layers in one prompt, and then ask it to turn the layered explanation into a visual concept diagram - a fast way to see the structure before committing it to notes.
Four-layer unpacking prompt
"I am a [your level] student studying [your subject]. Here is a text I need to understand [paste or upload the document or extract]. Unpack it through four layers: (1) the overall logic and main argument, (2) the key concepts and how they connect, (3) the most important supporting details and evidence, (4) any specific facts, names, or statistics I should note. Then, in a final step, suggest a simple concept diagram I could draw to represent the structure."
After receiving this, check the Layer 1 and 2 output against the actual text before trusting it - sycophancy means AI may simplify or subtly shift an argument to make it sound cleaner than it is.
✍ Activity
Take a reading that has been feeling dense or slow. Run the orientation prompt first - paste a passage or the abstract and ask for the central claim, key terms, reading questions, and things to verify. Then use the four-layer prompt on the section you find hardest. Compare what you understood before and after. Notice whether AI's Layer 1 summary matches what the author actually argues when you check it yourself.
AI for Understanding
By the end of this lesson you will be able to:
- Use the four-prompt workflow to unpack a concept from big picture down to working parts
- Turn a layered explanation into a concept diagram and know how to read it critically
- Apply a structured verification prompt to find oversimplifications before you rely on the output
Macro to micro: the analogy method
Most difficult concepts are difficult because you are missing one of two things: a big-picture handle on what the concept is for, or a clear view of how its parts actually fit together mechanically. AI is unusually good at supplying both, on demand, at exactly the level of detail you ask for.
The method is simple: ask for an analogy that captures the big picture first, then ask AI to zoom in, mapping each part of the analogy onto the real mechanism, one layer at a time, until you reach the level of detail you actually need. This connects naturally to the four-layer framework from Lesson 2: the analogy helps you build Layers 1 and 2 (the logic and the concepts), and mapping its parts takes you into Layer 3 (important details).
Worked example - zooming from macro to micro
Macro: "Explain monetary policy transmission the way you'd explain it to someone who has never studied economics, using one analogy." AI might respond: it's like adjusting the water pressure at the start of a long pipe network - turn the valve (interest rates) and pressure changes ripple through every branch (borrowing, spending, prices), but with delays and leaks along the way.
Micro: "Now map each part of that analogy onto the real mechanism: what is the valve, what are the pipes, and where are the leaks?" AI then connects the analogy's parts to the actual transmission channels - the policy rate, bank lending rates, household and business spending, and the lags and frictions that cause real-world effects to differ from the simple model.
The same method works for a STEM concept: "Explain how a neural network learns, using one analogy" followed by "Now map the parts of that analogy onto backpropagation, gradient descent, and weight updates."
A four-prompt workflow for genuine understanding
A single well-formed prompt is a starting point, not a finished process. The four-prompt sequence below is a complete understanding workflow: it takes you from a layered breakdown of a concept, through a visual representation, to a structured verification pass that actively looks for what AI got wrong or oversimplified. Each prompt builds on the previous one. Working through all four with one concept produces a different quality of understanding than asking a single question and accepting the answer.
Prompt 1 - Four-layer understanding
Build your context first, then request the layered breakdown. The more specific your context — who you are, what exactly you are studying, which aspect you are focusing on — the further past the average AI's response will be. If you have a relevant paragraph from a reading or your notes, paste it in: a grounded explanation from your actual material is more useful than a general one.
📄 Try it now - sample context text
No reading to hand? Use this passage. Paste it into the field above and set your topic to "AI skills and responsible learning" with aspect "the risk of over-reliance and what good use actually looks like."
"AI tools are becoming part of study, work and professional communication. Students and early-career professionals can use AI to understand difficult topics, summarise information, practise interview answers, improve writing, brainstorm ideas and receive quick feedback. However, using AI well is not the same as asking AI to do the work. Strong AI use requires clear instructions, subject knowledge, critical judgement and verification. There are also risks. AI can produce confident but incorrect answers, invent references, reinforce weak assumptions, or make learners over-dependent. The most useful approach is to treat AI as a learning companion rather than a replacement for thinking. AI can speed up support, feedback and drafting, but the human learner still needs to understand the task, check the output, adapt it to context and take responsibility for final decisions."
Prompt 2 - Turn the explanation into a concept diagram
Once you have the four-layer breakdown, ask AI to turn it into a visual. The key is being specific about what the diagram should show and what it should leave out — otherwise AI will give you a cluttered illustration rather than a genuine learning tool.
Copy-ready: Prompt 2
Now turn the four-layer explanation into a concept diagram. Requirements: show the big picture at the top; show the main concepts beneath it; show important details only where they directly support understanding; leave out the arbitrary details; use arrows to show cause, influence or dependency; use short labels. Also give a one-sentence explanation of how to read the diagram.
Use this in Claude, ChatGPT, or Gemini immediately after Prompt 1 — no need to re-explain the topic.
Prompt 3 - Structured verification
This is the prompt that does the most to beat sloth. Instead of asking "is this right?" — which invites a sycophantic "yes" — you ask AI to find specific weaknesses in its own output. A well-formed verification prompt forces AI to be a critic rather than a cheerleader.
Copy-ready: Prompt 3
Now review the explanation and diagram for responsible use. Please identify and briefly explain: (1) one part of the explanation that is strongest and most reliable; (2) one part that may be oversimplified or where the analogy breaks down; (3) one relationship in the diagram that should be checked against evidence before being relied on; (4) one question I should ask next to understand this topic more deeply.
This is a direct antidote to sycophancy and sloth from Lesson 1. AI, if asked gently, will reassure you. Asked specifically and structurally like this, it will find the seams.
Prompt 4 - Polished output
If you need a clean, shareable version of what you have produced — a study note, a Padlet contribution, a revision card — this final prompt wraps it up.
Copy-ready: Prompt 4
Give me a final clean version: a one-sentence summary of the big picture; the main concepts in a short numbered list; one key detail per concept that is necessary for real understanding; and one verification point I should check before using this in my work.
Optional — use this when you want a polished note you can actually keep, not just a conversation you scroll back through.
For more on how structured AI prompting supports genuine learning rather than surface summarising, see StudyAnalyst's How LLMs Can Support Your Learning Journey and Thinking with AI.
✍ Activity
Pick one concept from your current course that you understand "in theory" but could not explain clearly to someone else. If you do not have a reading to hand, use the sample context text above. Run it through all four prompts in sequence: four-layer breakdown, concept diagram, structured verification, final clean output. The question to ask yourself at the end: what did Prompt 3 reveal that you would not have caught if you had stopped at Prompt 1?
Practice, Feedback & Originality
By the end of this lesson you will be able to:
- Generate self-test questions pitched at unseen-exam difficulty from your own notes
- Use AI to check a draft against a marking rubric without being misled by sycophantic feedback
- Use AI to bring genuine originality to coursework, through six concrete techniques
Self-testing: building your own unseen exam
Retrieval practice - actively recalling information rather than re-reading it - is one of the most evidence-backed ways to improve retention. AI makes this fast to set up. Most general chat AI tools (ChatGPT, Claude, Gemini and others) allow you to paste or upload your lecture notes, a reading, or your own summary, and ask for questions pitched at unseen-exam difficulty.
How to give AI your material
Paste: Copy and paste text directly into the chat if your material is short enough.
Upload a file: Most major AI chat tools let you upload a PDF or Word document. This is faster for longer materials, and lets you target a specific section or the whole document. For example: "Using only Chapter 3 of the uploaded document, generate 8 exam-style questions."
Heavy materials (many readings): If you are working with a whole module's worth of documents, this is where a dedicated tool like NotebookLM starts to outperform general chat AI - we cover that in Lesson 5. For a single reading or set of notes, general chat AI works well.
Example prompt
"Here are my notes on [topic]. Generate 10 flashcard-style questions at the difficulty of an unseen final exam - include at least three that require applying the concept to a new scenario, not just recalling a definition. Give me the answer key separately so I can test myself first."
Run this a few days before a real exam, then again closer to the date with different scenarios, so you are testing understanding rather than memorising the same ten questions.
Rubric-matched feedback, without the flattery
Asking AI "is this good?" is close to the least useful question you can ask. Asking it to check your draft against the actual marking rubric, criterion by criterion, is far more useful - but it is exactly the situation where sycophancy from Lesson 1 shows up most. AI reviewing your own work has a built-in pull toward reassurance.
📄 Give AI your actual rubric
You can give AI your rubric in two ways: paste the text directly, or upload it as a file (PDF, Word document, or even a screenshot image if that is all you have). AI can read a rubric from an uploaded image well enough to apply it criterion by criterion.
Being specific about the rubric - rather than asking generally "is this good?" - is the single most important step for getting useful feedback, because vague feedback comes from vague rubrics, not just vague prompts.
🤔 Use the friend reframe from Critical AI Literacy
Critical AI Literacy's prompting lesson covers a specific fix for sycophancy in feedback: strip your own identity out of the context. Instead of "is my essay good?", try: "This essay was written by a friend who wants to do well against this rubric before submitting. Here is the rubric and the draft. What would lose marks against each criterion, specifically?"
This single reframe consistently produces more direct, criterion-by-criterion feedback than asking AI to review your own work.
Bringing originality, not outsourcing it
A common worry is that AI flattens coursework into generic, average answers. Used deliberately, it can do the opposite: AI is well suited to helping you generate a genuinely original angle, because it can rapidly try out reframings you would not have time to brainstorm alone. The six techniques below come from research on what makes academic and creative work feel original, and each pairs with a concrete AI prompt.
Six ways to make coursework original
a. New place or period: apply your argument or theory to a geography or era it is not usually discussed in. "This theory is usually applied to 1990s UK policy. Help me explore how it would apply to present-day Southeast Asia."
b. New angle on a familiar theme: approach a well-covered topic from an unusual entry point. "Most essays on this topic start from the economic angle. What would this look like starting from the psychological angle instead?"
c. Anchor a major event to a minor one: explain a big trend through a small, overlooked case. "Instead of explaining this trend through its most famous example, help me find a minor, lesser-known case that illustrates the same mechanism."
d. New disciplinary context: borrow a method or concept from another field. "What concept from ecology could usefully be applied to how information spreads in social networks?"
e. New theoretical context: re-read your data or case through a theory not standard in your field. "My field usually analyses this through Theory X. What would Theory Y, from a different field, notice that Theory X misses?"
f. Listening for absence: identify what nobody has addressed, and make the gap itself your contribution. "Looking at how this topic is usually discussed, what question does almost nobody ask, and why might that be?"
💡 Find the right technique for your topic
Not every technique fits every assignment. Ask AI to help you match one to your specific brief: "I am writing [describe your assignment]. Of the six originality techniques - new place/period, new angle, minor-to-major, new disciplinary context, new theoretical context, or listening for absence - which one or two would work best for this topic, and why?"
AI is good at pattern-matching your topic against these frames, and giving you a starting point that you then develop in your own direction.
AI for technical originality: analysis, models, and diagrams
Originality is not only about argument framing. AI can also help with the technical side of producing something that did not exist before your work: analysis you could not have done unassisted, a conceptual model adapted to your specific context, or a diagram inspired by one you saw in a paper.
Three concrete uses
Analysis and coding: "I have this dataset in CSV. Write and explain code that tests for [specific pattern], and talk me through what the output means." AI can help you run an analysis that would otherwise be out of reach, while you stay responsible for interpreting it correctly.
Adapting a standard model: "Here is the standard version of [a model from your field]. Help me adapt it to apply specifically to [your unusual context], and tell me where the standard version's assumptions stop holding."
Replicating or adapting a diagram: "I saw a figure in a paper showing [structure]. Help me build a similar diagram or a comparable analysis for my own dataset or context, in a way that is clearly my own work and properly attributed to the original."
💡 The verification habit applies here too
An AI-generated model, analysis, or diagram is a draft to interrogate, not a result to submit unchecked. Ask it directly where the adapted model's assumptions break down, and verify any numerical output yourself. The more impressive and "finished" an AI output looks, the more it deserves a second look, not less.
✍ Activity: three parts
Part A - Exam or standard learning: Take a real upcoming assessment or topic you are revising. Upload or paste your notes and generate five unseen-exam-style questions. Answer them without looking at your notes, then check.
Part B - Coursework or written work (including presentations): Pick a piece of upcoming coursework. Pick one of the six originality techniques above and use it to generate one genuinely different angle on your argument. If you are writing a presentation, use the same technique to find an unusual opening frame or an unexpected comparison that makes the talk more memorable.
Part C - Feedback against a rubric: If you have a rubric or marking criteria, use the friend-reframe prompt above and upload or paste your rubric. Run the check against your current draft, and note which criteria get the most critical feedback.
AI for Research
By the end of this lesson you will be able to:
- Explain why research-specific AI tools are safer than general chat AI for sourcing claims
- Use NotebookLM to ground answers in your own documents and study materials
- Use Consensus and Elicit to find and check real academic sources
Why general AI chat is the wrong tool for sourcing
Lesson 1 introduced the citation trap: AI inventing a reference, or attaching a real reference to a claim it never actually makes. A general-purpose chat model is especially prone to this for research, because it is generating its answer from compressed patterns in its training data rather than reading a specific document in front of it. A separate category of AI tool exists specifically to fix this, built around a method called RAG - Retrieval-Augmented Generation.
🧠 What RAG actually changes
Instead of generating an answer purely from training data, a RAG tool first retrieves the specific, real passages relevant to your question from a defined set of documents - your uploaded readings, or a database of real papers - and then generates its answer grounded in those retrieved passages, usually with a direct citation or quote pointing back to where it found it.
This does not make hallucination impossible. It makes it far easier to check, because the tool shows you exactly which passage it used, every time.
NotebookLM: grounding answers in your own documents
Google's NotebookLM is a free RAG tool built around your own materials: upload your lecture notes, readings, or a set of papers, and ask questions that are answered only from those documents, with every claim linked back to the specific source passage.
Practical details for getting started
Free with a Gmail account. You can add up to 50 documents per notebook, and you can create multiple notebooks - one per module is the recommended approach. A paid tier unlocks higher document limits if you need them.
More than just Q&A. NotebookLM has over 10 built-in study tools, including auto-generated audio overview "podcasts" and video lecture summaries of your uploaded material. These are genuinely useful for a first orientation pass before you go deeper. If you are not sure where to start, search for a short YouTube walkthrough - the interface is intuitive but worth five minutes of setup time.
Hygiene rule: one notebook per module or topic, never mix. If you upload materials from multiple courses into one notebook, NotebookLM answers across all of them, which is rarely what you want and can muddy your thinking. Keep notebooks clean and scoped.
Because NotebookLM is grounded in your actual readings, it cannot invent a reading that does not exist - it can only misread one that does, which is a far smaller and more checkable risk than general chat AI.
Consensus and Elicit: AI search across real academic literature
Consensus and Elicit are AI-powered search engines built specifically over academic paper databases, rather than general AI chat. Both retrieve real, existing papers first and summarise findings across them, rather than generating a reference from scratch.
What each is good for
Consensus is built for quick, evidence-based answers to yes/no or directional questions - ask "Does spaced repetition improve long-term retention?" and it surfaces real papers and a consensus meter showing how much agreement exists across the literature.
Elicit is built for systematic literature review work - extracting and tabulating findings, methods, and sample sizes across many papers at once, useful when you are building a literature review section rather than answering one quick question.
⚠ These tools reduce the risk - they do not remove it
RAG-based and literature-specific tools are substantially more reliable for sourcing than general chat AI, but "grounded in a real source" is not the same as "always correctly interpreted." Always click through to the original abstract or paper before citing a claim in your own work, exactly as you would check a fully manual citation. The citation trap from Lesson 1 still applies: a real, retrieved source can still be misquoted or have its finding overstated in the summary.
Building a good research stack
The combination across this course's tools gives you a research stack with each tool doing the job it is actually good at: NotebookLM for grounding answers in documents you already have, Consensus and Elicit for finding and checking the wider literature, and general AI chat reserved for explaining, illustrating, and brainstorming rather than sourcing facts. For more on building good research habits with AI, see StudyAnalyst's LLM-aware scholarly writing and Human Mind vs AI: the Jagged Frontier.
✅ A balanced view: reason for confidence, not paranoia
Throughout this course we have been honest about what can go wrong with AI - hallucination, sycophancy, citation traps. It is worth ending this lesson with an equally honest counterweight: the leading frontier models are making substantially fewer factual errors than they were two years ago, and RAG-based tools like NotebookLM significantly reduce the risk further by grounding answers in your own verified documents.
What this course has been building is not a blanket suspicion of AI, but an intuition: a feel for which tasks and which tools carry more or less risk, and a habit of checking proportional to the stakes. That is a much more sustainable and accurate stance than either trusting AI uncritically or refusing to use it at all. The goal is calibrated confidence - and that is what the pipeline in the final lesson is designed to make routine.
✍ Activity
Pick a real research question from your current coursework. Try it three ways: ask a general AI chat tool, ask Consensus or Elicit for relevant papers, and if you have your own readings, upload two or three into a NotebookLM notebook and ask the same question there. Compare how each answer is sourced, what risks each carries, and which one you would actually trust enough to cite.
Build Your AI Learning Pipeline
By the end of this lesson you will be able to:
- Explain why different study tasks need different AI workflows, not one generic habit
- Build a simple, reusable AI pipeline for a recurring study or coursework task
- Set up your own pipelines in advance, before you need them under time pressure
One AI habit does not fit every task
By this point in the course you have used AI for at least five different jobs: decoding a difficult text, explaining a concept, generating feedback against a rubric, building flashcards, and finding grounded sources. Treating all of these as "ask AI a question" misses what actually makes AI useful here - each job benefits from its own small, repeatable sequence of steps. That sequence is a pipeline: a workflow you have already worked out and tested, so that next time the task comes up, you are not improvising from scratch.
What a pipeline looks like in practice
A pipeline is just an ordered set of steps with a defined tool and prompt at each one, written down somewhere you will actually find it again - a notes app, a document, a saved prompt list. It does not need to be sophisticated to be valuable; it needs to be ready before the deadline pressure hits.
The anatomy of an AI learning pipeline
Every effective pipeline has the same three-part structure: define the task, run the AI steps, then verify. The verification step is the one that most commonly disappears under time pressure, which is exactly when you need it most.
What type of task? What materials do I have?
Right tool for each step; specific prompts ready
Sources, claims, sycophancy traps
Write the verification step into the pipeline as a named step, not an afterthought.
Three example pipelines, ready to adapt
Pipeline 1: New concept, first encounter
1. Orient: ask for a macro-level analogy (Lesson 3). 2. Map: ask AI to map the analogy parts onto the real mechanism. 3. Break: ask where the analogy breaks down. 4. Test: generate 5 self-test questions and answer them without notes. ☑ Verify: check the analogy's named weaknesses against your course materials.
Pipeline 2: Draft ready, before submission
1. Upload or paste the rubric (as a file or screenshot). 2. Use the friend reframe (Lesson 4) to get criterion-by-criterion feedback. 3. Revise the weakest sections. 4. Re-run the same check once more before submitting. ☑ Verify: treat "this is strong" as a sycophancy flag and push for specific weaknesses.
Pipeline 3: Exam in two weeks
1. Upload all readings for the module into a NotebookLM notebook (Lesson 5). 2. Generate a study guide and flashcards. 3. One week out, generate a fresh batch of unseen-exam-style questions with new scenarios. 4. Two days out, repeat with harder application questions, no notes. ☑ Verify: click through inline citations to check the source passage before trusting any key claim.
Build your own pipeline
The most valuable pipelines are the ones built around your own recurring tasks. Think about the study or coursework task you face most often this term, and write down the steps using the same structure as above: what you give AI at each step, which tool, and what you check before moving on.
Build a pipeline you can reuse
Choose a recurring task, add the materials you normally use, and name the evidence you will check. The builder will turn these choices into a three-stage pipeline.
💡 Build the verification check into the pipeline itself
The most common way a pipeline fails is that the verification step gets skipped under time pressure - precisely when sycophantic, hallucinated, or unchecked output is most likely to slip through. Write the check in as a named step with a specific action, not a vague intention to "double-check later."
✍ Final reflection
Write down one pipeline you will actually use this term, using the Task → AI Steps → Verify structure above. Then take two minutes: what is the single most important thing you have taken from this course, and what will you do differently the next time AI gives you an answer that feels a little too easy or too confident?
🏠 Where to go next
This course has focused on smarter learning with AI as your companion. If you have not yet completed StudyAnalyst's Critical AI Literacy course, it covers the underlying mechanics of hallucination, the jagged frontier, and responsible AI use policy in more depth, and pairs directly with everything covered here. For more on building durable habits around AI tools, see What High-Performing AI Users Do Differently and Why Learning Cannot Be Automated.
Test Your Understanding
12 questions. You need 8 out of 12 to complete the course. Answer options are randomised. Read each question carefully before answering.
Course Complete 🎉
You have completed AI as Learning Companion: Smarter Learning in Higher Education. Enter your name below to generate your certificate.
✓ Uses analogy and AI illustration to move from big-picture to mechanism
✓ Builds rubric-matched feedback and unseen-exam self-testing routines
✓ Applies AI-assisted originality techniques to coursework
✓ Grounds research and sourcing using NotebookLM, Consensus and Elicit
✓ Designs reusable AI learning pipelines and verifies AI output by habit
📩 Stay updated on new courses
We'll let you know when new courses are available. No spam — just course updates.
🚀 What next?
You now have five concrete AI-supported moves for smarter learning: understanding via analogy and illustration, rubric-matched feedback and self-testing, AI-assisted originality, grounded research, and reusable pipelines. To go deeper on the mechanics underneath all of this — hallucination, sycophancy, the jagged frontier — explore RAIS, a free diagnostic and personalised learning tool at rais.studyanalyst.com, which maps your AI literacy across five domains and gives you a personalised learning path.
If you have not yet taken it, Critical AI Literacy: Using LLMs Responsibly is the companion course to this one and covers the underlying limitations in more depth. A third course on AI for Everyday Productivity is coming soon for professionals. Use the form above to be notified when it launches.