AI Ethics & Limitations

The most important habit in this whole course is a two-second pause before you paste anything into an AI tool: what kind of information is this, and may I share it? Most organisations think in data classification levels:

AI feels like a private conversation, but it is not. With a free tool, your text may leave your control - treat the chat box like a postcard, not a sealed envelope. A useful filter: would you be happy if this appeared in tomorrow's newspaper under your name?

A hallucination is information that sounds authoritative but is made up - a fake statistic, a non-existent regulation, a quote nobody said. The danger is the tone: AI states wrong answers as calmly as right ones, so you supply the doubt.

AI also absorbs human bias from its training text. In office work this is subtle: a job description leaning toward one gender, a "pros and cons" list that stacks one side. Ask for the other side directly:

> Review this candidate shortlist for any language or assumptions
> that could disadvantage a group by gender, age, ethnicity, or
> background, and suggest neutral wording.

Verification sounds heavy, but for most tasks it takes under a minute - a habit you run before output leaves your hands. Use this five-point check; depth scales with the stakes.

For a dense document, make the AI do the first pass, then confirm the flags:

> Here is the draft variance-analysis summary. List every factual
> claim and figure as a checklist for me to tick once I have
> confirmed each against the source.

Beyond accuracy, there are rules about ownership and honesty. They vary by employer and country, but the principles below hold widely.

Intellectual property works two ways. Do not feed in material you are not licensed to use, such as a competitor's paywalled report. And because the legal status of AI output is still settling, do not assume you own or can freely publish it.

Disclosure matters more as AI shapes more of your work. Light help with phrasing rarely needs a note, but when AI drafts paid client work, informs a decision, or feeds a regulated submission, disclosure is often expected - check your contract and policy. A one-line footnote that an AI assistant helped draft the text, which you reviewed, is usually enough.

Not every task should be handed to AI, even when it could be. If a task scores high on several of these, slow down or keep it human:

A safe pattern: use AI for the draft and structure, then human judgement for the decision and sign-off. For borderline tasks, narrow the AI's role:

> Do not make a recommendation. Lay out the three options for this
> vendor decision with the trade-offs of each, so I can decide.

Now turn everything above into one page you will actually follow - pre-made decisions so you don't re-think the rules each time. Answer these five prompts:

1. What I will share - the data classifications I will put into my approved tool.
2. What I will never share - the hard "no" list for my role.
3. How I verify - the checks I run, and which tasks get the full review.
4. How I disclose - when I tell others that AI helped.
5. Where I stop - the tasks I keep fully human.

Use AI to draft it, with you making the calls:

> I am a [your role] who handles [type of work and data]. Draft a
> one-page personal AI usage policy covering what I will and won't
> share, how I verify, and when I disclose AI help. Ask me three
> clarifying questions first. Do not invent my company's rules.

Edit the result so it sounds like you, then keep it next to your Lesson 9 templates. Revisit it every few months, because the tools and the rules keep changing.