AI safety and judgment, part 2: hallucinations, bias, copyright and policy

ChatGPT can be confidently wrong, quietly unfair, or borrow someone else's work. Know the four risks and the one-line disclosure habit.

Hallucinations: confident and wrong

A hallucination is an answer that sounds right and is not. ChatGPT may invent a statistic, a quote, a legal requirement, a product feature, a citation or a person. It does not warn you. The sentence reads exactly like a true one.

Where they happen most:

  • Specifics with no source: figures, dates, names, references, phone numbers.
  • Things that should exist: “the standard form”, “the usual clause”. It may describe one that does not exist.
  • Gaps in a document: when summarising, it fills what the document does not say with a plausible guess.

The defence is the two-minute check from lesson 6 plus one rule: a fact that matters needs a source you have opened. If ChatGPT cannot show where a claim came from, it is unverified.

The three hotspots and the defence, together:

Hallucination hotspots: specifics with no source, things that should exist, gaps in a document; the defence is a source you have opened yourself
Know where to look, then insist on a source you have opened.

Bias: quietly unfair

ChatGPT learned from human writing, which carries human assumptions. Unchecked, AI drafts can:

  • Assume a nurse is “she” and a mechanic is “he”.
  • Rank job applicants on names, schools or career gaps rather than skills.
  • Write more warmly to one customer than another based on a name or address.

Two habits. Never let the AI decide about a person: hiring, firing, pricing, credit, housing, discipline. Use it to draft and organise; a person decides. And when you ask for anything that compares people, ask ChatGPT to list the criteria it used, then check them.

Output you create at work is generally yours to use, with two cautions:

  • Do not ask it to reproduce someone else’s work. “Rewrite this competitor’s brochure so it’s ours” is a copyright problem whoever types it.
  • Images: no other companies’ logos, characters or products, no real people without consent (lesson 5).

If a long passage reads like a quote from a real book, article or song, assume it may be one and do not publish it.

Your company AI policy

Most small businesses now have a short AI policy. It usually says:

  1. Which tools and accounts are approved.
  2. What may be pasted (the three buckets from part 1).
  3. Which tasks need a human decision.
  4. When AI use must be disclosed.

If you have not read yours, ask for it today. If there is none, the rules in these two lessons are a safe default.

Disclosure: the one-line habit

Situation Disclose? How
Internal draft you edited and own Usually no It is your work
Customer-facing text you edited and checked Follow policy Often no, as long as a person reviewed it
Anything a reader assumes you wrote personally: condolences, references, apologies Yes, or write it yourself “I used an AI tool to help draft this”
Formal documents: legal, medical, financial Yes, and a qualified person reviews Note in the document or cover email
Images, audio or video that look real Yes, always Caption or label

When unsure, disclose. Nobody is upset to learn you were honest.

The same rules as a quick decision path:

Disclosure decision flow: disclose when a reader would assume a human wrote it, for formal advice, and for realistic images or audio; otherwise follow policy
Three yes-or-no questions cover nearly every case.

Try it now

Find your company AI policy, or ask who owns it. Compare it with lessons 7 and 8 and note one thing it adds and one thing it misses. If there is no policy, write three sentences that could start one and send them to your manager.

Remember: a fact that matters needs a source you opened, people decisions stay with people, do not borrow others’ work, and disclose when a reader would want to know.

Checking your progress…