Judgment: Limits, Hallucinations & Safe Use

By the end you will know when to trust AI and how to use it safely at work.

See it first

Ask AI for five sources on a niche topic. It will give you five, beautifully formatted, with authors and dates. Some may not exist. The model did not "lie." Recall from the refresher: it generates the most likely-looking text, and a plausible-looking citation is easy to generate whether or not it is real.

Why hallucinations happen

A hallucination is a confident, false statement. It is not a malfunction; it is the normal output of a system whose only job is to produce likely text, with no built-in check on truth. Knowing this changes how you work: you stop expecting the model to police itself, and you take that job yourself.

Trust, but verify

The durable habit: use AI's speed, then verify anything that matters before you rely on it.

Safe to trust directly

Drafts you will read and edit anyway
Brainstorms and option generation
Rewrites and tone changes of your own text
Explanations you can sanity-check

Verify before relying

Facts, statistics, dates, names
Quotes and citations (check they exist)
Legal, medical, financial specifics
Anything that will carry your name or a decision

A simple verification workflow: ask for the source, then check the source yourself; cross-check a key fact against a second tool or a quick search; and for anything high-stakes, treat AI output as a first draft a human expert must sign off.

What not to delegate

Some things AI should help with but never fully own:

Final decisions with real consequences (hiring, firing, money, safety)
Judgment calls that require accountability, someone has to own the outcome
Anything requiring guaranteed accuracy that you have not verified
Relationships and sensitive communication where authenticity matters

AI drafts; you decide. That division of labor keeps you safe.

Data privacy: what is safe to paste

This is where leaders most often slip. When you paste text into a consumer AI tool, it may be stored or used to improve models. So:

For sensitive work, use your organization's approved enterprise AI (which typically does not train on your inputs), or redact specifics before asking. When in doubt, ask your IT or security team, and assume the default consumer tier is not private.

What you can now do

Explain why AI hallucinates and why it cannot police its own truth
Apply "trust but verify": know what to accept directly and what to check
Name what should never be fully delegated to AI
Apply the paste-safe rule to protect confidential and regulated information

Check your understanding

1 / 3

verify-citations

An AI gives you three citations for a claim. What should you do before using them?

What's next

You can use AI safely and judge its output. Now turn that into daily leverage: AI at Work: Your Productivity Toolkit.