AI outputs are sources: same as a news article, same as a scientific paper, same as a tweet. OPVL applies to AI exactly the way it applies to any other source; see the OPVL Framework for the four questions and how they read when the source is AI.
This page covers what's specific to AI: the failure modes you have to watch for, the documentation template you fill out every time you use AI in your work, and three worked examples.
Critical AI failure modes to watch for
- Hallucinated citations. AI will confidently invent paper titles, authors, dates, and DOIs that don't exist. Always check that any citation an AI gives you actually exists.
- Confident wrongness. AI rarely says "I don't know." When it doesn't know, it usually makes something up that sounds plausible. Suspicion is appropriate when an AI is highly specific about something it has no reason to know.
- Training data cutoff. If you're asking about something that happened recently, the AI may not know. Always check the training cutoff date and the date of the event.
- Sycophancy. AI tends to agree with the framing of your question. If you ask "why is X true," it will tell you reasons X is true even if X is false. Try asking the opposite to test.
- Domain weakness. Some AI is great at writing but terrible at math, or great at general science but bad at specific regional geology. Test it on questions you can already check before trusting it with work that counts.
The AI Documentation Template
Every time you submit work where AI was used, you fill out this template. It goes at the END of your deliverable (last page, or a final slide, or a footer paragraph: depending on the format).
| Field | What you write |
|---|---|
| Tool(s) used | Name(s) and version(s) of the AI tool(s). Examples: "Source Evaluator (Flint)"; "Skeptic (BoodleBox)"; "Plate Tectonics Tutor (BoodleBox)". Name the partner and the platform. List all if you used more than one. |
| Prompts (representative) | 1–3 of the prompts you used. Doesn't need to be every prompt: just the ones that shaped the work most. Quote them. |
| What the AI contributed | Specific tasks: brainstorming, outlining, drafting a paragraph, giving feedback on writing, translating, generating examples, summarizing a source, explaining a concept you didn't understand. Be specific. |
| What YOU contributed | Decisions, evidence you chose, your interpretation, your voice, your edits, your conclusions. This should be the bulk of the work. Then name one part you deliberately kept for yourself, and say why that part. |
| OPVL of the AI's contribution | Quick OPVL on the AI as a source for this specific work. Highlight especially the Limitations: what was it wrong about, what did you have to fix, what couldn't it do? |
| One thing the AI got wrong | Required. Name at least one specific thing the AI got wrong, missed, refused, or couldn't do well. If you literally couldn't find one, that's a sign you weren't checking carefully enough. |
| Was it worth it? | One line. Running an AI costs energy and water. Was what you got back worth that cost, for this task? A reasoned no is as good an answer as a reasoned yes. |
Why "one thing it got wrong" is required
If you submit AI documentation that says the AI was perfect and contributed everything well, one of two things is true. Either (a) you weren't paying enough attention to notice its mistakes, or (b) you let the AI do too much of the work without checking. Both are problems. Requiring a specific failure point forces you into the critical-evaluation mode that makes AI use valuable instead of dangerous.
Worked examples
Strong example
Tool(s) used: the Skeptic and the Plate Tectonics Tutor, both on BoodleBox.
Prompts (representative): "Can you explain how megathrust earthquakes generate tsunamis in plain language?" / "Give me feedback on this draft paragraph: is my mechanism explanation clear?" / [Plate Tectonics Tutor] "What's the difference between liquefaction and lateral spreading?"
What the AI contributed: The Tutor explained the megathrust→tsunami mechanism three different ways until one clicked. The Skeptic gave feedback on the draft hazard profile and flagged that I'd confused magnitude with intensity. The Plate Tectonics Tutor walked me through liquefaction vs. lateral spreading using a Tokyo-specific example.
What I contributed: Topic choice (Tokyo), source selection (USGS, Smithsonian, BBC archive on Tōhoku 2011), the full OPVL on each source, my own writing in the final hazard profile, and the synthesis connecting historical evidence to Tokyo's current risk.
One thing the AI got wrong: The Tutor initially said the Anchorage 1964 earthquake was M9.0; the actual figure is M9.2 (per USGS). Small but important: corrected before final draft.
Why this is strong: specific tools, specific prompts, clear division of labour, specific identified weakness. A reader could reconstruct what role the AI played.
Weak example
Tool(s) used: the class platform.
Prompts: I asked it to help me with my hazard profile.
What the AI contributed: It helped me a lot.
What I contributed: I wrote it.
One thing the AI got wrong: I don't remember.
Why this is weak: No specificity. The reader can't tell what role the AI played. The "one thing it got wrong" answer is the giveaway: if you can't name it, you probably weren't checking. This earns lower C-strand scores and triggers a conversation about how you actually used the tool.
The violation
No AI Documentation Template submitted with the deliverable. When asked, the student admits to having used an AI to draft most of the work.
Why this is the violation: The student avoided documentation precisely because the AI did the work. The hiding is the problem. Heavy AI use disclosed honestly might earn a low T/T score; hidden AI use is an integrity issue. Always file the documentation: your worst-case grade with honest docs is still better than the consequences of hiding.
Next
Class Expectations covers the rest in plain language: what AI use is expected of you, the disclosure principle, and what to do if you think a use crossed the line.