Why AI states false things confidently — and a fast workflow to verify facts, quotes and citations before they reach your work.
Module · Research + Writing
Lesson 6 of 9 available lessons
⬡ What you'll build
A language model predicts plausible-sounding text. Plausible is not the same as true. It will invent a statistic, misattribute a quote, or cite a paper that does not exist — and do it in the same confident tone it uses for correct answers. For a student, an unverified AI claim in an essay is a real risk: to your grade and to your credibility.
This lesson makes verification a fast, automatic habit.
A model generates the next most likely word given everything so far. It has no built-in fact-checker and no concept of "I don't know" unless prompted toward it. When it lacks the real answer, it produces the most plausible one — which can be fabricated detail that reads exactly like fact.
Key implications:
Treat these as unverified until you confirm them from a primary source:
What's lower-risk: explanations of well-known general concepts, brainstorming, rephrasing your own text, structure and outlining. Even then — if a specific fact rides along, check it.
| If the AI output contains… | Risk | Do this before using it |
|---|---|---|
| A statistic, number, or date | High | Find the primary source that reports it; don't trust the figure alone |
| A quote or attribution | High | Confirm the person actually said it, in the original |
| A citation / reference | Highest | Resolve it — exact-title search + DOI check + does it support the claim |
| A named fact (law, definition, event) | Medium | Confirm against an authoritative source |
| A general explanation / your own text reworded | Low | Use it — but check any specific fact riding along |
The simple gate: specific claim → verify from a primary source → if it won't verify, drop it.
Flag the claims
Find a primary source
Verify citations exist and say what's claimed
Cross-check uncertain facts
If you can't verify it, don't use it
This is the highest-risk failure for students. A model can produce a reference that is perfectly formatted and completely invented, or a real author paired with a title they never wrote.
Checks that catch most fakes:
Run every AI-suggested reference through this sequence — stop and discard the moment one step fails:
Exact-title search
Resolve the DOI
Match the metadata
Confirm it supports your claim
Never paste an AI-generated reference list into an essay without resolving every entry to a real, checked source.
These lower the rate of confident fabrication but do not replace verification:
Answer using only what you are confident is accurate. For any claim you are unsure about, say "unverified" instead of guessing. Do not invent citations — if you don't have a real source, say so. List which parts of your answer I should independently verify.
Even with this, you still verify the facts that matter. The prompt makes the model flag its own uncertainty; it doesn't make it reliable.
A student asks for sources on a psychology topic and gets five clean, formatted references. Three are real. One pairs a real author with a paper they never wrote. One doesn't exist at all. All five look identical in confidence and formatting. Only checking each title and DOI reveals which two are fabricated. Submitting the list unchecked would mean citing sources that don't exist — a serious integrity problem the student didn't intend.
AI-Assisted Research Workflow for Students
Where verification fits in a full literature-discovery-to-synthesis pipeline.
Maintaining Citation Integrity with AI
Citing AI use correctly and keeping your reference list honest.
Related reading
AI-Assisted Research Workflow for Students
Where citation verification fits the full pipeline
Writing with AI as a Partner, Not a Ghost
Prompting Fundamentals for Students
Prompts that make the model flag its own uncertainty
Maintaining Citation Integrity with AI
The citation-specific deep dive