Fabricated citations and sources
Only cite sources you can verify exist; if unsure, say so instead of citing.
A free reference from the course
Hallucination guardrails
A hallucination happens when an AI system sacrifices factual accuracy, evidence, logic or context in order to produce a fluent, plausible-sounding answer. It is not lying and not malfunctioning. Producing something that reads well is what the system is built to do, and when the facts run out it keeps going anyway.
You cannot switch this off. You can make it far less likely by saying the right thing in your prompt. Below are twenty-four failure modes and the specific sentence that heads each one off.
Start here
If you only ever add one thing to your prompts, add this. Four lines that cover the most common failures, suitable for almost any task.
Paste it at the end of a prompt, or save it once in your AI tool's custom instructions so it applies every time.
Only cite sources you can verify exist; if unsure, say so instead of citing. If information needed to complete this task is missing or unclear, ask a clarifying question instead of inventing details. If you're uncertain about any part of your answer, state your confidence level explicitly. After drafting your answer, review it once for errors or unsupported claims before finalizing.
The full reference
Six categories. Pick the one that matches the work you're doing, then copy the line that fits.
The most damaging failure for anyone writing a grant application, a sermon, a board memo or a proposal. AI will produce a citation that looks perfect — real-sounding author, plausible journal, correct formatting — for a paper that does not exist. Lawyers have been sanctioned for filing briefs built on invented cases.
Invented book, case or law names
If you can't recall exact titles or case names, describe the concept instead of inventing a reference.
Citation details pulled from memory
Cite only sources retrieved in this conversation; do not generate citation details from memory.
Language models predict text, they do not calculate. Arithmetic that looks confident can be quietly wrong, which matters when the number is going into a budget, a grant report or an invoice.
Silent arithmetic errors
Show your work step by step and re-check each arithmetic step before presenting the final answer.
Mental-math mistakes on complex calculations
If a calculation involves more than two steps, perform it using a tool (code execution) rather than mental math.
False precision in numeric answers
State your confidence level in the final numeric answer, and flag any assumptions or estimations used to reach it.
Asked to produce a complete document, AI will produce a complete document — inventing whatever it needs to reach the end. The invented parts read exactly like the real ones. Telling it that asking a question is an acceptable outcome changes the behavior entirely.
Invented details to finish a narrative
If information needed to complete this task is missing or unclear, ask a clarifying question instead of inventing details to fill the gap.
Fabrication blended in as fact
Clearly distinguish between facts you are certain of and details you are inferring or extrapolating, using phrases like 'it's likely that...' for the latter.
Plausible-sounding placeholders
If you don't have enough information to finish a section, state what's missing rather than generating a plausible-sounding placeholder.
A wrong answer delivered in a steady, competent tone is harder to catch than an obviously shaky one. These lines slow the model down and get its uncertainty out into the open where you can see it.
Skipped or shortcut reasoning
Think through this step by step before giving your final answer.
Overconfident weak answers
If you're uncertain about any part of your answer, state your confidence level explicitly.
Blending outside assumptions with your facts
Base your answer only on the information provided; do not use outside assumptions unless you flag them as such.
Errors missed in the first draft
After drafting your answer, review it once for errors or unsupported claims before finalizing.
Misreading an ambiguous request
If the request is ambiguous, state your interpretation before answering.
AI is trained to be helpful, which shades into telling you what you want to hear: a complete-looking answer instead of an accurate partial one, a confident position instead of an honest disagreement. These lines give it permission to be less satisfying and more correct.
Speculation presented as completeness
Prioritize accuracy over completeness — a partial but correct answer is better than a full but speculative one.
Vague, unverifiable claims
Use precise, falsifiable language rather than vague qualifiers when stating facts.
One-sided answers on contested topics
When multiple valid perspectives exist, present them rather than picking one as definitive.
Unverified claims presented as fact
If a claim can't be verified with available tools or context, label it as unverified rather than omitting the caveat.
Contradictions across a long conversation
Keep your answer consistent with earlier parts of this conversation; flag it if something changes.
Uploading a file feels like the AI has read all of it. Often it has not. It may have sampled part of a spreadsheet, stopped partway through a PDF, or skipped a sheet it could not parse — and then answered as though it had the whole thing. These five lines force it to tell you what it actually saw.
Reading only part of the file
Before answering, confirm the full scope of the file — total rows, sheets, or pages — and state that count back to me.
Extrapolating from a partial sample
Process every row/sheet/page in the file; do not summarize or extrapolate from a subset unless I explicitly ask for a sample.
Silent partial answers on oversized files
If the file is too large to process in full, tell me that explicitly and propose how to split the task, rather than silently giving a partial answer.
Unreadable or malformed sections missed
Flag any rows, sheets, or pages you were unable to read or parse, and specify which ones.
Unverifiable figures
When citing a figure from the file, state its exact location (sheet name + cell, or page number) so I can verify it.
Better prompts reduce the risk. They don't remove it.
Every line on this page shifts the odds in your favor. None of them turns the output into something you can send unread. Hallucination is a property of how these systems work, not a bug awaiting a fix, so the last check is always a person.
Before anything goes out under your organization's name, run the five questions from the course.
- Pause. Does this answer matter enough to verify?
- Identify the claims. What facts, numbers, names, dates or quotes am I being asked to believe?
- Check the source. Where did this come from, and is it reliable?
- Compare against reality. Does it match trusted documents or what I already know?
- Decide how to use it. Use as-is, revise, verify further, or discard?
This is one page of Module 3
The full course covers where hallucinations come from, how to spot the red flags in a draft, what never to paste into a public chatbot, and how to make AI a genuine thought partner rather than a fast way to produce work you then have to check.