AI_Assisted_Literature_Review_Student_Guide_V12.docx

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 1
Literature Research with AI
From Survey to Synthesis
Scientific Inquiry with AI
Student Guide
Detailed Instructions & Course Material
Why This Lesson
Every research project begins inside a conversation that other researchers started before you. A literature
review helps you learn what the field already knows, what it still argues about, and where a new question
might fit.
AI can help with the slow parts: finding papers, comparing them, mapping citations, and drafting first-pass
syntheses. But AI can also sound confident while being wrong. It can misremember a finding, overstate a
claim, or cite something that does not support the sentence
Learning Goals
Upon completion of this investigation, you should be able to:
Phenomena:
● Distinguish manual retrieval tools from generative AI, and understand why the two warrant different kinds
of trust.
● Read a field rather than a paper: identify what a research community treats as settled, what it treats as
contested, and the signals that distinguish the two.
Experimental Practices:
● Formulate a focused, researchable question by hand, then use AI as a critic to refine it — and
understand why question quality determines search quality.
● Execute the five-step AI-assisted research workflow — claim survey, citation mapping, corpus reasoning,
grounded synthesis, and orchestrated deep research — using structured, reusable prompts.
● Design the dimensions of a synthesis matrix appropriate to your own research question, and use it to
evaluate and compare sources.
● Use a validated corpus (your verified papers plus your deep research report) in a grounded-Q&A tool
such as NotebookLM to deepen your own understanding.
Habits of Mind:
● Verify AI-generated claims against primary sources, and articulate why this is the same discipline
scientists have always applied to any intermediary source.
● Identify a gap, contested point, or possible extension that no source you found has directly addressed —
the first step of original inquiry.
● Recognize when a fully automated “deep research” tool is the right way to get oriented in an unfamiliar
topic, and when its output still needs further investigation before you can rely on it.
What You Will Need
● A general-purpose AI model with web search, file upload, and a deep research mode — Claude,
ChatGPT, Gemini, or Perplexity. Check your deep research usage limits before class; you start one run
during the session.
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 2
● Google Scholar — free. The only search tool you use for the manual baseline in Part 1.
● ResearchRabbit — free. Used for citation mapping in Stage 4.
● NotebookLM — free with a Google account. Used in Stage 9.
● Your institutional library proxy login — the route to full-text papers behind paywalls.
● Full-text PDFs of the two or three papers most central to your question, obtained in Part 1.
● Your two written pause notes from the demonstration video, on paper. Class opens by revisiting them.
● Optional: Semantic Scholar — free. Its open-access PDF links are often the fastest route to a full text.
The Core Idea
● AI output sits on a spectrum of trust. A search tool gives you real papers you can open. A language
model gives you a written summary that may or may not be accurate. Do not automatically trust or
distrust either one. Instead, ask: What kind of output is this, and how should I check it?
● The five steps are ordered by how much work you hand to the AI. As the AI does more searching,
organizing, and synthesizing, your job shifts to directing and checking its work. The most important
decisions still stay with you: what question to ask, what claims to trust, and what gap to pursue.
● A matrix cell is the natural unit of verification: Each cell should contain one small claim from
one source, so you can verify it quickly. Blank cells are also useful. If no paper in your set
measured something, that blank may point to a research gap.
● The driving question the demonstration runs on: “Our example asks: When does AI help
undergraduate STEM students learn, and when does it just help them get answers without learning? You
will run the same process on your own research question.
Part 1 — Pre-Class Preparation ~2 hours
A. Manual Literature Work and the Demonstration Video (~85 minutes)
Do this half of the prelab in the order given. The manual work comes first on purpose: you cannot judge
what an AI-assisted search is worth if you have never felt what the search costs by hand. Everything you
build here — your question, your friction notes, your PDFs — gets used in class.
1. Define your research question (10 minutes)
Pick a topic you genuinely care about, or adopt and adapt one from the starter list below. Then write a
focused, researchable question by hand — no AI yet. A strong question names a system or population, a
mechanism or variable, and a comparison.
Weak: “How does caffeine affect the body?”
Stronger: “Does caffeine improve reaction time on a simple visual stimulus task in habitual coffee drinkers
compared with non-drinkers, and if so, is the effect dose-dependent?”
Starter topics (build a question around one of these if you are stuck — adopting one is not a lesser path):
● Does spacing your studying out over time actually help you remember?
● Does listening to music while you study help you learn?
● Does caffeine boost performance?
● Does using fast chargers shorten a rechargeable battery's life?
● Does the color of an object influence human response?
● Is genetic modification of food, animals, or humans beneficial or harmful?
Why it matters: Question quality determines search quality. A vague question returns a vague literature —
and a long automated run on a weak question is the most expensive mistake available in this lesson.
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 3
2. Manual search — Google Scholar only (20 minutes)
Keyword search, no AI. Find the five papers that seem most relevant. Read titles and abstracts.
What to look for: The friction. Take notes as you go: what was hard to find, what turned out to be wasted
effort, whether you could realistically have reviewed every plausible source, and whether you suspect
relevant work exists under search terms you could not guess.
Why it matters: The friction is the point. These notes are the reference experience against which every
AI-assisted result in this lesson gets judged, and class opens with them.
3. Deep manual read (15 minutes)
Pick one paper and read it without AI. Write a three-sentence summary in your own words, plus one
question the paper alone cannot answer.
4. Obtain the PDFs (10 minutes)
Get the full text of the two or three papers most central to your question. These seed the synthesis matrix
you build in class; you will add to them with whatever the AI-assisted search surfaces.
Routes to a paywalled paper, in rough order of preference: your institutional library proxy; the publisher
page reached through the library's link resolver; preprint servers (arXiv, bioRxiv, medRxiv, PsyArXiv,
ChemRxiv); Google Scholar's “All versions” link, which often surfaces an author-hosted PDF; Semantic
Scholar's open-access links; the author's own webpage; and, as a last resort, a polite email to the
corresponding author — which works more often than you would expect, and is itself a small lesson in
scholarship as conversation.
Judgment call: If a paper resists access after a genuine attempt, document what you tried and move on. An
abstract is an acceptable fallback — marked as such in your matrix.
5. Watch the demonstration video (~30 minutes, including pauses)
The video runs the complete five-step workflow on a shared demonstration question about AI in education. It
contains two writing pauses, and both are mandatory:
● Pause 1 (before Step 1 is revealed):
“If you had a human research assistant surveying this literature, what exact instructions would you
give them? Include what kinds of papers to find, what details to extract, and how they should handle
uncertainty.” Write it before you resume.
● Pause 2 (before Step 4 is run). “Write one claim you would bet is settled and one claim you would
bet is contested. What signal would tell you the difference?” Write both before you resume.
Bring both written notes to class. The tutor session asks about them, and class opens by revisiting them.
Watching the video without doing the pauses defeats its purpose — the gap between what you wrote and
what the video reveals is where the learning happens.
B. Socratic AI Tutoring Session (~15 minutes)
Watch the video and complete both pause prompts first — this session will ask about them. Then upload the
Prelab Tutor Document to your AI model and follow its lead.
The tutor will ask questions instead of giving a lecture. It will help you think through four key ideas: how
much to trust search results versus AI summaries, what each step hands over to AI, why verification
matters, and what decisions still stay with you.
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 4
Answer honestly. There are no wrong answers here. Saying “I’m not sure, but here’s my thinking” is better
than guessing. At the end, you will get an engagement score as a personal target, not a grade, plus a short
roadmap for what to focus on next.
Share the chat log with your instructor when you finish. Only completion is recorded for the course.
Note: The tutor session is also an example of the lesson’s main question: when does AI help students
learn instead of just giving them answers?
C. Additional Pre-Class Preparation (~15 minutes)
Written summary (~10 minutes). After the video and the tutor session, write a brief summary of the key
terms and core concepts in your own words: the five steps and what each one delegates, and the difference
between a search result and a generated summary. Consolidating this before class is what makes the
in-class work fast.
Tool access verification (~5 minutes). Confirm all of the following before you walk in:
● Your AI platform works and its deep research mode is available to you.
● ResearchRabbit opens (you use it in Stage 4).
● Your Google account opens NotebookLM (you use it in Stage 9).
● Your library proxy login works — you already used it to get your PDFs.
D. Review of In-Class Investigation (~10 minutes)
Skim the stages below. Do not try to memorize them — get the shape of the block, and note two things as
you read: where you expect to make a decision, and where you expect to want AI help. Class moves quickly,
and the stages you have previewed are the ones you will actually run rather than read.
Part 2 — In-Class Investigation ~2 hours
The Investigation, Stage by Stage
You will run the full workflow on your own research question while the instructor guides the class
through each stage. This guide gives you the prompts, steps, and decisions you need to make.
Remember two things: start your deep research run early so it can work in the background, and
expect to leave with a possible research gap, not a final conclusion.
Stage 1 — Refine your question with AI (~5 minutes)
Run the refinement prompt on the question you wrote by hand in the prelab. Read the critique and the
sharpened versions it proposes.
Prompt 0 — Question refinement (run before everything else)
Here is my research question: [question]. Critique it as a literature-review question: does it name a specific system or
population, a mechanism or variable, and a comparison? Is it answerable from existing literature, or is it really
several questions bundled together? Propose 2–3 sharpened versions with different scopes, and for each, say what
kind of literature it would point toward. Do not choose one for me — give me the trade-offs and I'll decide.
Decide: Choose the version of the question you want to use: your original, one AI suggested, or a mix of
both. The AI gives feedback, but it does not choose for you. Your research question should still reflect what
you care about and what you can realistically investigate.
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 5
Why it matters: You wrote your question by hand first. Now use AI to review it. The AI checks whether your
question has a clear population or system, variable or mechanism, and comparison. It can suggest
improvements, but you still decide what the final question becomes.
Stage 2 — Launch deep research (~2 minutes)
Open your AI's deep research mode, paste your refined question, and start the run. It works in the
background for the rest of the block. You come back to it in Stage 8.
Why it matters: This is the one step you cannot make up later. If you forget it, you will have nothing to
compare at the end of class.
Judgment call: Launch it after the refinement pass, not before. A long autonomous run on a weak question
is the most expensive mistake available in this block.
Stage 3 — Step 1: Claim survey (~10 minutes)
Fill in the claim-survey prompt with your own details in the brackets, then run it with web search on. Skim the
results, choose the two or three most relevant papers, and check which ones you can access as full-text
PDFs. These are possible sources for your matrix.
Prompt 1 — Claim survey
I'm researching [question]. Search the literature and return 8–10 relevant papers. For each, give me: full citation with
DOI or stable link, population or scope, methodology in one sentence, key finding in one sentence, and whether the
finding supports, contradicts, or qualifies [hypothesis]. Include at least two papers from fields adjacent to [primary
field] that study the same phenomenon under different terminology, and tell me what terms those fields use for it. If
you cannot verify a detail from the paper or abstract, mark it as 'not confirmed' rather than inferring it.
Why it matters: This step shows the difference between manual search and AI-assisted search. AI can
return a useful paper list quickly, but its summaries still need to be checked against the original sources.
Checkpoint: If more than half the results seem off-target, your question's key terms are ambiguous. Revise
the question and rerun — refinement is iterative, not one-and-done.
Do now: Jot one line toward the comparison table in Stage 8: what did each approach surface, and what
would you trust each one for?
Stage 4 — Step 2: Citation mapping (~5 minutes)
Open ResearchRabbit and paste one starred paper from Stage 3. Explore the visualization for a few
minutes: the cluster around your paper, the earlier work it builds on, and the later work that builds on it.
What to look for: You are not just collecting more papers. You are looking at how the field is connected.
Older, highly connected papers may be foundational. Clusters may show different methods or schools of
thought. Newer papers near the edges may show where the field is still developing.
Why it matters: This is a map of a research conversation, not a search. The goal in this stage is
awareness, not mastery: know that this deeper mode of exploring a literature exists, and know what kind of
question it answers.
Do now: Note one or two papers the map surfaces that Stage 3 did not.
If the tool is unavailable, a general model can approximate this step — but be honest with yourself about the
difference: the specialized tool reads a structured citation graph that the general model does not have direct
access to.
Prompt 2 — Citation mapping (general-model approximation)
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 6
Take this paper [citation]. Identify the foundational papers it builds on (the work it's responding to) and the recent
papers that have built on it. For each, give me the citation, its relationship to the seed paper, and a one-sentence
summary. If you are inferring the relationship rather than confirming it directly, say so.
Stage 5 — Step 3: The synthesis matrix (~15 minutes)
Build your source set from your prelab PDFs plus the best papers from Stages 3 and 4. Use three to five
sources for your first matrix. Use full-text PDFs when possible, and mark any abstract-only sources clearly.
Before running the AI prompt, decide which columns your matrix should include.
Decide — this is the stage's real work: The columns of a comparison table are an analytical decision, and
they depend on your question. A core set applies almost everywhere: method or study design, system or
population, independent variables, dependent variables, key finding, stated limitations. Then you add the
one or two your particular question demands (ex. dose levels for a caffeine question, type of exercise for a
mental-health question). The columns are the thinking. The AI populates the cells.
Prompt 3 — Corpus reasoning (synthesis matrix)
I've uploaded [N] papers on [topic]. Build a comparison table with one row per paper and these columns:
method/study design, [system or population], independent variables, dependent variables, key finding, and stated
limitations, plus these question-specific columns: [your additions]. Add any further columns you think my question
requires and tell me why. Then, below the table: where do the papers agree, where do they disagree, and what
different predictions would their findings make in a new experiment? Ground every cell in the uploaded papers, and
leave a cell blank rather than inferring a value the paper doesn't state. Where a source is an abstract only, mark its
row as abstract-based.
What to look for: Blank cells. The prompt tells the model to leave a cell empty rather than infer, and empty
cells are where gaps live: “nobody in my set measured X” is a proto-gap statement.
Pause before you read the AI's synthesis: Take two minutes and write your own read of the matrix first —
where the papers agree, where they disagree, which blank cells look like gaps — in your own words, before
the AI offers any synthesis at all. You compare the two in the next stage.
Stage 6 — Step 4: Grounded synthesis and gap analysis (~8 minutes)
Run the orientation prompt with web search on. Include the papers you reviewed in Stages 3–5 so the AI
can compare its synthesis with your source set. Read the response carefully: look for what seems
well-supported, what seems uncertain, and what you may need to check next.
Deep Research does a larger version of this same process in the background.
Prompt 4 — Grounded synthesis (directed, with confrontation)
Give me a 200-word orientation to the current state of research on [question]. Separate what is well-established,
what is contested, and what appears to be emerging. I have already reviewed these papers: [citations]. Note whether
your synthesis is consistent with them, and flag any claim in your synthesis that those papers would dispute. Cite
primary sources rather than secondary summaries, and flag any places where the evidence base is thin.
What to look for: Two things. First: does the settled/contested split match the bets you wrote at the video's
second pause? Score yourself honestly. Second: what kinds of sources did it cite — primary research,
review articles, popular press? A synthesis is only as good as the sources it drew from.
Why the confrontation clause matters: An open synthesis built from sources the AI chose is being
checked against a closed corpus built from sources you chose. The two modes triangulate. If they agree,
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 7
you have independent corroboration. If they disagree, either your curation missed something or the model's
search pulled weak sources — and working out which is exactly the judgment this lesson exists to build.
Compare: Set the AI's synthesis beside the read you wrote at the end of Stage 5. Where did you agree?
What did it see that you did not — and what did you see that it did not?
Stage 7 — The spot-check (~8 minutes)
Pick one specific, checkable claim the AI made about one paper — a number, a method, a population, a
finding — and go to the primary source. Read the relevant section. Record three lines: the claim as the AI
stated it; what the source actually says; and your verdict (confirmed / overstated / wrong / couldn't access).
Why it matters: This is not framed as defense against AI failure. It is the standard discipline of working with
primary sources — the same one scientists have always applied to a review article, a textbook, or a
colleague's summary. Any summary is a pointer toward the primary source, not a substitute for it. If the AI
turns out to be right every time you check, good: the habit is the point, and the checking is what makes the
claim genuinely yours to build on.
Judgment call: “Couldn't access” is an acceptable verdict, provided you document what you tried. The
access struggle is data too.
Worth knowing: Finding an AI error and documenting it earns credit. It never costs it.
Stage 8 — Step 5: Deep research and the comparison table (~7 minutes)
Your deep research run has finished by now, or will shortly. Take a first look at the report with two questions
in hand: can you see the steps inside it — where is it surveying, where is it synthesizing? And would you put
it in a paper without verifying the sources yourself? Then open the comparison table and start filling it in. You
complete it at home; reading the report closely is homework too.
What it helped me
find/do
What I verified What I would trust it
for
Manual search (prelab)
Claim survey (Stage 3)
Citation map (Stage 4)
Synthesis matrix (Stage
5)
Grounded synthesis
(Stage 6)
Spot-check (Stage 7)
Deep research report
Judgment call: Stage 6 and deep research are not competing versions of the same move. Stage 6 costs
seconds and is the everyday probe; deep research costs many minutes and a limited budget of runs, and is
the commissioned survey for a serious project.
Stage 9 — Learning from a validated corpus (~20 minutes)
This stage sits outside the five-step workflow. The five steps help you find what the field knows; NotebookLM
helps you learn from the sources you collected. After a short instructor demo, build a notebook using your
verified papers and your Deep Research report if it is finished. Explore the notebook, and click at least one
inline citation to check the exact source passage.
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 8
Why it matters: NotebookLM is only as trustworthy as the sources you upload. Since you spent class
building and checking your source set, this is one of the safest ways to use AI for learning. You can also use
this approach in your other classes.
What to watch: NotebookLM only knows the sources you upload. That makes it more grounded, but not
perfect. It may still miss context or combine ideas incorrectly, so always click citations to check the original
source.
Share: If you find a feature worth showing, show it. The class shares discoveries at the end of the stage.
Reflect: Where Does It Break Down?
Close the block by writing one-line answers to three prompts. Some will be shared aloud.
● What is the consensus claim in your area — what does the field treat as basically settled?
● What is contested or unresolved — where are researchers still arguing? And how could you tell?
● What is a question you could investigate that no paper you have found has directly answered?
You are leaving with a possible research gap, not a final answer. In the postlab, you will test that gap
against your matrix. If you do not have one yet, that is okay. Before you leave, ask: when would I use Deep
Research, and when would I do the steps myself?
Part 3 — After Class [~2 hours]
A. Finish the In-Class Work and Generate Your Report
Complete anything the class did not finish; your instructor will say what at the end of the session. Two things
carry over by design: read the deep research report closely and complete the step-comparison table.
Then have your AI produce a draft report of your analysis. Write the prompt yourself: specify what you want
in it, what evidence it should draw on, and what it should not invent. Then review it for accuracy. You are the
one who signs it, and any claim you cannot trace to a source you have actually read does not go in.
B. Literature Map and Gap Statement (~90 minutes)
This is the main deliverable: a one-page Literature Map and Gap Statement, built on the synthesis matrix
you began in class and supported by your NotebookLM notebook. Five parts.
● The question and its evolution. Three stages — the prelab version you wrote by hand, the AI-critiqued
version after Stage 1, and the final version after seeing the literature — with one line on what changed at
each stage and why.
● Synthesis matrix: Expand your matrix to 5–10 papers. Use full-text PDFs when possible, and update
any abstract-only rows. Verify the cells you rely on, leave unknown cells blank, and attach the matrix to
your Literature Map.
● A small map of the field. 5–10 papers with one-sentence characterizations you have personally verified
against the primary source (not against an AI summary).
● Gap statement: Use your matrix to test your possible research gap. Look for blank cells, disagreements
between papers, or missing populations/conditions. A real gap should matter and should point to
evidence that would help resolve it. If your first gap does not hold up, revise it and explain what
changed.
● A reflection, written after you have built and used the notebook. Which steps were most useful for
which sub-tasks? What did the notebook add that reading alone did not — including any features your
classmates showed you? Did the deep research report find the gap you ended up identifying, miss it, or
suggest something you had not considered? What surprised you? And how did going to primary sources
change your understanding compared with relying on AI summaries?
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1

Scientific Inquiry with AI • AI-Assisted Literature Review • Student
Guide
Page 9
Stretch goal, if you finish early — the Swanson move. Pick a field adjacent to yours, ask what that field
knows about your phenomenon, and look for a connection neither field has drawn.
Prompt 5 — Cross-field connection (stretch)
I'm studying [phenomenon] within [field A]. What does [field B] know about this phenomenon or its underlying
mechanisms? What terminology does that field use, what are 2–3 key papers, and is there a finding in field B that
researchers in field A appear not to have engaged with? Distinguish clearly between connections you can support
with specific sources and analogies you are speculating about.
C. Retrieval Practice / Course Review (~30 minutes)
Spend about thirty minutes on integrated course review using the course's AI-enhanced retrieval tool. This is
the standard routine in every lesson: it spaces your practice out over time and interleaves this lesson with
earlier ones, which is what makes the material stick rather than fade.
Everything You Turn In
● Your hand-written research question, and your manual search notes — including the friction.
● Your three-sentence deep-read summary, and the one question that paper alone could not answer.
● The full-text PDFs of your two or three central papers (or documented access attempts).
● Your tutor session chat log, shared with your instructor.
● Your two written pause answers from the demonstration video.
● The step-comparison table, completed.
● The spot-check record: claim, source, verdict.
● The Literature Map and Gap Statement — all five parts, with the extended synthesis matrix attached.
● Your NotebookLM notebook of validated sources (or a link to it).
Companion documents: Lesson Plan • Instructor Guide • Prelab Tutor
Document
V1