Verifying AI Output
Build the checking habit that makes AI safe to use at work, and learn which tasks are appropriate to delegate to it at all.
By the end of this lesson you can
- Explain what a hallucination is and why it occurs
- Apply a verification process proportionate to the stakes
- Identify the categories of output that always require checking
- Decide which tasks are appropriate for AI assistance
Lesson Notes
Read through the key concepts before you try the challenge.
The failure mode is confident invention
You prepare a policy summary at Lakeside Medical Associates.
You ask an AI tool to summarize the HIPAA requirements for records retention. The answer is well organized, appropriately hedged, and cites a specific section of the regulation with a specific retention period. The section number does not exist. Nothing in the response looked wrong.
Your task: Build a verification habit strong enough that plausible errors do not reach anyone.
A hallucination is output that is fluent, confident, and false. It happens because the model generates plausible continuations rather than retrieving verified facts — and a fabricated regulation citation is just as plausible-looking, statistically, as a real one. The model has no mechanism to tell the difference, which is why it presents both identically.
| Output type | Risk | How to verify |
|---|---|---|
| Citations and sources | Frequently fabricated, including plausible authors and dates | Locate every source. If you cannot find it, it does not exist |
| Statistics and figures | Invented or misremembered | Trace to the original publication |
| Legal or regulatory claims | Confidently wrong; consequences are serious | Check the actual regulation, or ask compliance |
| Clinical information | Potentially harmful | Never rely on it. Use authoritative clinical references |
| Arithmetic | Language models are unreliable at calculation | Recompute it yourself |
| Names, dates, specifics | Plausibly wrong | Check against the source record |
Matching verification to the stakes
Decide how much checking four different AI-assisted tasks need.
- 1
Rephrasing a paragraph you wrote — read it once.
You already know the content, so you can spot a change in meaning immediately. The risk is low and the check is nearly free. This is the category where AI is most straightforwardly useful.
- 2
Drafting a routine internal email — read it properly before sending.
Low stakes, but it goes to colleagues under your name. An ordinary proofread is proportionate — the same attention you would give anything you wrote yourself.
- 3
Summarizing a policy for staff — verify every specific against the source.
Staff will act on this. Every number, deadline, and requirement must be checked against the actual policy document, because a confident wrong retention period becomes practice behavior.
- 4
Anything clinical or regulatory — do not use AI output as the source at all.
The consequences of a confident error are serious enough that AI-generated content should not be the basis of the answer. Use it to help you phrase what an authoritative source already told you, never to establish the fact.
Result: Verification effort proportionate to consequence, rather than uniform or absent.
Ask what happens if this is wrong. The answer sets how hard you check — and sometimes says not to use the tool for it.
An AI summary of a regulation cites a specific section number and retention period. What should you do before circulating it to staff?
Challenge
Apply what you've learned in this lesson.
Find a hallucination yourself. It is more convincing than being told they happen.
- Ask an AI tool a detailed question in an area you know well and ask it to cite sources. Attempt to verify each source and record what you find.
- Ask it to perform a multi-step calculation. Check the arithmetic by hand.
- Build a verification checklist for your own work, with different levels for low, medium, and high stakes output.
- Write a short guidance note for colleagues explaining what AI is genuinely useful for at work and what it must never be trusted with.
Finished this lesson?
Progress is saved in this browser only. It is not a grade — official progress lives in Brightspace.