AI for Research and Information Gathering

Teacher note

Accounts to pre-stage

• NotebookLM runs on the free Google account staged back in Module 4. Confirm that account still works for anyone who has not touched it since.
• Google Scholar and DOI lookup need no account, which is why the verification half is logistics-proof.
• Have your own notebook ready to project, and a second one you can break on purpose.

Fifteen minutes before

• Ask your own chatbot for a cited mini-essay and verify two of its citations, so you know tonight's failure mode before the room does.
• Open the notebook and click one inline citation through to its passage. That click is the demo.
• Open Spot the Troll and click past the intro screen.
• Draw the tally on the board: valid, suspicious, fabricated, with room for a second row after the re-run.
• Put the printed packets somewhere you can reach without announcing that the wifi died.

Wifi fallback

Printed citation packets, some real and some fabricated, verified against library terminals or phones. The verification race works fine on paper and the tally still happens, which is the part that teaches. The NotebookLM contrast is the block you lose, so move it to the next session's wander rather than cutting it.

Depends on earlier modules

• Module 1's predict-the-next-word demo is the mechanism behind everything tonight. Call back to it explicitly, the room will remember shouting words at a screen.
• Module 4 built the habit of checking a number against its source. Tonight is the same move pointed at a citation.
• The privacy conversation started here feeds the policy students write in Module 6.
• If Module 7 is looking heavy, this module's wander block is the natural place to pre-stage GitHub accounts.

Teacher note

Rates, not anecdotes. "AI sometimes makes things up" is a warning nobody acts on. A number the room produced with its own hands, then cut with a better prompt, is a finding they will repeat to other people. Get the tally on the board and leave it there.

Teacher note

The 15-minute mini-lesson is carved out of this module and teaches on its own. If you are ever asked to give a single taster session, that is the one to give, and it is already built.

Anticipated wrong turns

• Every citation checks out. Happens more often now, especially with search-connected tools. Do not treat it as a failed lab. Ask what the tool did differently, and have them check whether the source actually says what was claimed, which is where the real failures have moved.
• "Suspicious" becomes a dumping ground. Define it before they start: a real-looking record you could not find, versus one you positively confirmed does not exist.
• A real paper is filed as fabricated because Scholar's first page did not have it. Show one recovery live, searching the exact title in quotes, before you trust the tally.
• The constrained re-run produces four references instead of five, and someone calls that worse. It is better. Fewer and checkable beats five and confident.
• Somebody wants to argue that this proves the tools are worthless. The tally answers that on its own: the rate went down when the prompt changed, and it did not go to zero.

Discussion, with the answers you are steering toward

"What did a fabricated citation look like?"

Plausible authors, a real journal, a believable year, a DOI with the right shape. It looks exactly like a real one, which is the entire problem.

"Why does the constrained prompt help?"

It changes what a good answer looks like to the model. It does not give it access to a library, which is why the rate drops instead of hitting zero.

"Where would this have cost you something?"

Take real answers: a paper, a client memo, a court filing in the news. Then the practical rule, said once: never pass on a citation you have not opened.

Anticipated wrong turns

• Grounded gets read as "cannot be wrong". It can. It is faithful to its sources, and if a source is wrong the answer is wrong with a citation attached. Say this explicitly, it is the most common misread of the block.
• The break attempts succeed. Great. Show one to the room and look at what it actually did: usually it reached for general knowledge and said so, or it stretched a passage further than the passage goes.
• The shared notebook hits a limit with thirty simultaneous readers. Have a second copy, and be ready to run it as a projector demo with the room calling out questions.
• A pair spends the whole block curating sources and never asks a question. Time-box the curation to five minutes.

Discussion, with the answers you are steering toward

"When would you reach for the grounded one?"

Anything that has to be right and has a defined source set: a policy, a contract, a set of readings, your own company's documents.

"When is the open one better?"

Thinking out loud, drafting, brainstorming, anything where you are the source of truth and it is the assistant.

"What did it feel like when it refused?"

Mildly annoying, and that is worth naming. A tool that declines is doing its job, and the tools that never decline are the ones to distrust.

Anticipated wrong turns

• Students read the About page and stop. That is the site describing itself. Push them off the domain entirely.
• The race rewards speed over judgment. Score the reason, not the verdict. A right answer with no reason gets no credit.
• A site is genuinely ambiguous. Say so. "I could not tell in sixty seconds" is a legitimate and professional verdict.
• It gets political fast. Pick your example sites in advance, and pick boring ones. The technique is the lesson, not the topic.

Discussion, with the answers you are steering toward

"What did you check first?"

The best answers leave the page immediately. Fonts, layout and a professional design tell you nothing at all.

"How is this different from what you normally do?"

Most people read down. Fact-checkers read across. That is the whole finding, and it is one sentence long.

Spotlight sweeplab 1:20 – 1:30

Best catch of the night. Three or four students show their most convincing fabricated citation and how they caught it. The competitive framing works here: people want to show the good fake.

Prompts to hand the student showing

• "Read it out. Would you have believed that?"
• "What was the step that caught it?"
• "How long did it take you to be sure?"

Prompts for the room

• "Hands up if you would have accepted that in a paper."
• "What did the tally change for you?"

Wander block menulab 0:55 – 1:20

Twenty-five minutes, the longest in the course. This module produces genuine arguments, so let the discussion pods run and circulate through them.

• Lateral reading practice at your own pace, from the Civic Online Reasoning exercises.
• Privacy-policy skim of the tool you actually use: what is collected, who owns the input and output, does it train future models. Three questions, one document, fifteen minutes.
• Bad News or Cranky Uncle stations, both account-free browser play.
• Discussion pods on the tally. This is the pod that forms on its own, every time.
• Pre-stage a GitHub account if Module 7 looks tight from here. It fits naturally in this wander and it takes Module 7 from crowded to comfortable.

Synthesis, five minutes

Point at the two rows on the board and read them out. Then the worst hallucination contest, thirty seconds, whoever has the most confident wrong answer. Ending on a laugh is fine, the number stays on the board either way.

Exit ticket

One-minute paper: what will you check from now on that you did not check before? Specific answers only. "Be more careful" does not count.

Homework handoff

Ten factual questions to an assistant about a topic they know well, scored, with the most convincing wrong answer kept for next session. That kept answer is the warm-up for Module 6, so make it clear it gets used.

Preview Module 6 in one sentence: next time we stop asking whether the output is true and start asking who it lands on.