AI Ethics & Responsible Use: A Guide for Every Employee

AI predicts plausible text; it doesn't "know." The phone-autocomplete analogy, and the gap between "sounds right" and "is right."

Where the patterns come from; why an AI can only be as good, fair, and current as its data. Out-of-date, gaps, and bias.

The honest capability map; the "brilliant assistant who sometimes makes things up" image you must always double-check.

Sentience, objectivity, confidence = correctness, "just a search engine," prompts stay private, and "the AI is responsible."

The three mechanisms: biased history, data gaps, and biased proxies. "Common ≠ fair," and no villain is required.

Hiring (the tool that learned to prefer men), health-care (cost-as-proxy), and facial analysis (data gaps) — three cases, three mechanisms.

A five-question checklist anyone can run, ending on the master question: "if I were on the receiving end, would this feel fair?"

The response ladder: don't pass it on → verify → document → escalate. Noticing → telling is the step most people skip.

Stored, reviewed, reused; your prompt can become training data, and you can't un-train a model.

The five buckets (PII, company secrets, credentials, others' data, regulated data) and the two-second "newspaper test."

Redact, use fake data, describe-don't-dump, separate sensitive from safe, match the tool. "The AI needs the problem, not your secrets."

Plain-language GDPR; the organization stays legally responsible even when an employee pastes data into an unapproved tool.

"Which AI do I open?" reframed as a risk decision. Free public vs enterprise-approved — same screen, different contract.

Defines shadow AI; the scale (80%+ of workers); the confidence trap; why it's riskier than shadow IT — data flows out.

The Samsung 2023 leaks and the corporate-ban wave (JPMorgan, Apple, Amazon); scaled down to the learner's own data.

What approval buys: a data-processing contract, a no-training guarantee, deletion/audit control, standards alignment.

Locating the policy; the request path; reporting an over-share without fear (when, what data, which tool).

A daily habit: right tool? right data? right account? Pulls the section into one reusable routine.

Fabricated facts, fake quotes and citations; the danger isn't the lying, it's the confidence — fiction wears the costume of fact.

Scale checking to the stakes; "what happens if this is wrong?"; confirm facts outside the AI, not by asking it to check itself.

Two real cases (the sanctioned lawyer; the airline held liable for its chatbot). "The AI did it" is never a defence.

When AI use should be disclosed: work meant to show your ability, decisions about people, or AI presented as human/fact.

Overdependence as a real risk (skill atrophy); AI amplifies what you bring, so keep bringing something. Stay the thinker in the loop.

Keep your voice and ownership, verify facts before sending, guard the data in your prompt. The lazy-vs-skilled before/after.

Decision support vs decision making; the work-vs-personal split; "the bigger the decision, the smaller AI's vote."

The highest-stakes use: a human owns the decision, hunt for bias, guard the data, be transparent. Trigger: "this affects a real person."

Plain-language, not legal advice: a prompt alone doesn't make you the author; real human creative input is what creates ownership; watch for using others' IP.

Four habits (recognise limits · keep critical thinking · balance automation · use varied strategies) and the accurate/fair/safe/honest pocket check.

The EU AI Act's risk tiers made simple: minimal (use freely) → limited (disclose) → high (hiring, credit, biometrics) → unacceptable (banned).

The principles don't go out of date, only the specifics do; light habits — point your judgment at new tools, follow workplace updates, ask when unsure, share what you learn.

Recap of the five principles with their practical meaning; the accurate/fair/safe/honest pocket version; a personal responsible-use commitment.