Core Content
Part 1 — What Generative AI Actually Is (And Isn't)
A large language model is a system trained on enormous amounts of text to do one thing: predict the next most likely word (technically, the next "token") given everything before it. Chatbots like ChatGPT, Claude, Gemini, and Microsoft Copilot are LLMs wrapped in a friendly chat interface. When you ask a question, the model is not "looking up" an answer. It is generating text that statistically resembles a good answer based on patterns it learned in training.
That single fact explains almost everything else in this module. The model is fluent, confident, and grammatical — because it learned from fluent, confident, grammatical text. Fluency is not accuracy. The model can produce a beautifully worded, completely false statement with exactly the same confidence as a true one, because to the model they are the same kind of object: likely text.
YOU TYPE THE MODEL DOES YOU GET BACK
┌───────────┐ ┌────────────────────────────────┐ ┌────────────┐
│ a question│ → │ predicts likely next words, one │ → │ fluent text│
│ or task │ │ at a time, from learned patterns │ │ (maybe true│
└───────────┘ │ — NOT a lookup, NOT a calculator │ │ maybe not)│
└────────────────────────────────┘ └────────────┘
It optimizes for "sounds right," NOT "is right."What it is good at: drafting, summarizing, rephrasing, brainstorming, explaining a concept, translating tone, generating first-draft code or text you will review.
What it is NOT: a search engine you can trust without checking, a database of facts, a calculator, a lawyer, a doctor, or a substitute for your own judgment. (Many tools now add web search or a calculator on top of the model — useful, but it does not change the rule below.)
Treat every AI tool as a brilliant, fast, eager intern who has read everything and remembers some of it wrong, never says "I don't know," and will confidently invent a citation rather than admit a gap. You would not forward an intern's first draft to a client unreviewed. Same rule here.
Do / Don't
| Do | Don't |
|---|---|
| Use AI to draft, then verify the facts yourself | Treat output as a verified source of truth |
| Ask it to explain its reasoning so you can check it | Assume confidence means correctness |
| Use it for tasks where you can judge the answer | Use it for things you cannot check and cannot afford to get wrong |
Part 2 — How AI Fails: Hallucination, Bias, Confident-But-Wrong
There are three failure modes everyone must recognize on sight.
1. Hallucination. The model produces information that is fabricated but plausible: a fake legal case, a non-existent statistic, a citation to a paper that was never written, a quote no one said, an API function that does not exist. It is not "lying" — it is doing exactly what it was built to do (produce likely text) in a situation where the likely text happens to be false.
Example: A staffer asks for "three peer-reviewed studies showing X." The model returns three studies with titles, authors, journals, and years — and two of them do not exist. Every detail looks real because the model has seen thousands of real citations and generated text in that shape.
2. Bias. The model learned from human-written text, which carries human biases. It can reproduce or amplify stereotypes about gender, race, age, and other attributes — in hiring language, in who it assumes holds which job, in tone toward different groups, in which examples it reaches for first. Bias is most dangerous in consequential decisions (hiring, lending, eligibility, content moderation), which is exactly where extra human review is mandatory.
3. Confident-but-wrong. This is the meta-failure that makes the other two dangerous. The model has no built-in signal for its own uncertainty in the way a careful human does. A wrong answer arrives in the same fluent, assured tone as a right one. There is no flashing red light. The burden of doubt is entirely on you.
┌───────────────────────────────────────────────────────┐
│ HALLUCINATION → invented facts, fake citations │
│ BIAS → reproduces/amplifies human stereotypes │
│ CONFIDENT-WRONG → same calm tone for true AND false │
└───────────────────────────────────────────────────────┘
The fix for all three is the same: a human verifies.The higher the stakes, the more verification you owe. A brainstorm of blog headlines needs almost none. A statistic going into investor materials, a clause in a contract, a dosage, a financial figure, or anything a regulator could read needs full, independent verification — every time.
Part 3 — The Article 4 Obligation (Your Legal Baseline)
This module is the company's Article 4 control. You should understand what the obligation is so you understand why we log it.
The rule, in one sentence: Anyone who provides or deploys AI systems in the EU must ensure staff and contractors operating AI on their behalf have a sufficient level of AI literacy.
Two things matter in that sentence:
- "Provider AND deployer." You do not have to build AI to be covered. If you use a third-party AI tool in your work, your company is a deployer, and the obligation applies. Almost every company is now a deployer.
- "Sufficient AI literacy." The European Commission and major law firms have been explicit: distributing usage instructions is not enough. Sufficient literacy means people understand, at minimum:
- what AI is and how it can fail,
- whether the company is building or merely using a given system,
- the specific risks of the systems they actually touch,
- and content tailored to their technical level and role — a non-technical sales rep and a platform engineer need different depth.
And it must be documented. Training that happened but cannot be evidenced does not help you in an audit.
ARTICLE 4 — minimum sufficient literacy
┌─────────────────────────────────────────────────┐
│ general understanding of AI ........... Part 1 + 2 │
│ build-vs-deploy posture awareness ..... Part 3 (this part)│
│ risk awareness for systems you use .... Part 2 + 4 │
│ tailored to staff technical level ..... role-based tracks │
│ DOCUMENTED COMPLETION ................. the log, signed │
└─────────────────────────────────────────────────┘
In force 2 Feb 2025 · Enforcement begins 2 Aug 2026"We told everyone to be careful with ChatGPT" is not Article 4 compliance. "Here is the dated, signed completion log showing every employee and contractor completed role-tailored AI literacy training, version 1.0" is.
Waiting for the August 2026 enforcement date to start. The obligation has been in force since February 2025; the log you build now is the record you will be asked for later. Procrastination here is the most common — and most avoidable — gap.
Part 4 — Acceptable Use & Data Classification (What Can and Can't Go In)
The single most common way staff create risk is by putting the wrong data into an AI tool. Once you paste something into a tool, you may have no control over whether it is stored, logged, reviewed by humans, or used to train a future model. The rule is simple: classify before you paste.
We use four data tiers. Learn them — this is the part you will use every single day.
| Tier | What it is | Examples | May it go into an AI tool? |
|---|---|---|---|
| 1 — Public | Already published, no harm if seen | Marketing copy, public docs, press releases | ✅ Any sanctioned tool |
| 2 — Internal | Not public, low sensitivity | Internal memos, draft blog posts, meeting notes (no names) | ✅ Sanctioned enterprise tools only |
| 3 — Confidential | Business-sensitive; harm if leaked | Source code, financials, roadmaps, contracts, strategy | ⚠️ Only sanctioned tools with verified training-opt-out + approval |
| 4 — Restricted / Regulated | Legally protected personal/regulated data | Customer records (PII), PHI/health data, payment data, secrets/credentials | ❌ Never — no AI tool without explicit, documented sign-off |
The default-safe rule for non-technical staff: If you are not sure which tier it is, treat it as Tier 4 and do not paste it. When in doubt, ask before, not after.
Acceptable-use essentials (the short version):
- Use only sanctioned tools from the approved list. Sanctioned tools have terms negotiated so your input is not used to train the vendor's models, and they sit inside the company's security perimeter.
- Never paste Tier 4 data (customer PII, PHI, payment data, credentials, secrets) into any AI tool.
- Never paste secrets — API keys, passwords, tokens, private keys. Treat these as Tier 4 instantly.
- AI output is a draft, not a deliverable. A human owns and signs off on anything that leaves the company.
- Disclose AI use where a policy, a client contract, or a grant requires it.
The most expensive paste is the one you didn't think about: a support agent dropping a full customer ticket — name, email, account details — into a free chatbot to "draft a nicer reply." That is a Tier 4 → free-tool transfer, and depending on the vendor it may now be training data. The fix is not "don't help customers faster." The fix is the sanctioned tool, where that transfer is contained.
THE 5-SECOND DATA CHECK (do this before every paste)
┌─────────────────────────────────────────┐
│ 1. What tier is this? (Public / Internal / Conf / Restr)│
│ 2. Is this a sanctioned tool? (yes / no) │
│ 3. Tier 3? approval + opt-out verified? (yes / no) │
│ 4. Tier 4? STOP. Do not paste. Ask. │
│ 5. Any secret/credential? STOP. Never. │
└─────────────────────────────────────────┘Part 5 — Shadow AI Safety + Prompting and Evaluating Output
Shadow AI is any AI tool used for work that the company has not sanctioned — the personal ChatGPT account, the browser extension that "summarizes anything," the free image generator, the random app a colleague recommended. People reach for shadow AI because it is convenient and the sanctioned path feels slower. The danger is invisible: you do not know the vendor's data practices, whether your input trains their model, or whether it meets the company's security and compliance bar.
The shadow-AI safety basics:
- Prefer the sanctioned tool every time, even when it is one extra click.
- If the sanctioned set is missing something you genuinely need, ask for it to be added — do not route around it quietly. (Module K covers this in depth; the company manages shadow AI by enablement, not bans.)
- Be alert to prompt injection: malicious instructions hidden inside content you feed the AI — a webpage, a document, an email — that try to make the AI do something it shouldn't. If an AI suddenly behaves oddly after you fed it external content, stop and report it.
Prompting basics — how to get a useful answer:
A good prompt gives the model role, task, context, and format.
- Weak: "Write about our product."
- Strong: "You are a B2B copywriter. Write a 120-word LinkedIn post announcing our new analytics dashboard to operations managers at mid-size SaaS companies. Plain, confident tone. No buzzwords."
Then iterate: tell it what to fix. "Shorter." "More concrete." "Remove the marketing tone." Treat it as a conversation, not a vending machine.
Critically evaluating output — the four checks before you trust it:
- Verify facts. Any name, number, date, quote, citation, or claim of fact — confirm it independently. Assume citations may be fabricated until checked.
- Check for the tells. Over-confidence on something niche, suspiciously round numbers, a citation you cannot find, reasoning that skips a step.
- Apply your own judgment. Does this match what you know about the situation? If it contradicts your domain knowledge, trust yourself and dig in.
- Own it. If you forward it, you are vouching for it. Your name, not the model's, is on the output.
The skill that separates a productive AI user from a dangerous one is not prompting — it is evaluating. Anyone can get an answer. The literate user knows which answers to throw away.
Part 6 — When to Escalate
Knowing when to stop and ask a human is a core literacy skill, not a sign of weakness. Escalate — to your manager, the AI/security point of contact, or legal/compliance — when any of these is true:
- High-stakes output. Anything going to a customer, regulator, board, court, investor, or the public, where an error has real consequences.
- Consequential decisions about people. Hiring, firing, lending, eligibility, discipline, or anything where bias would cause harm. A human must own these decisions.
- Regulated or restricted data is involved. Any time Tier 4 data is in the picture and someone is tempted to use AI on it.
- You can't verify it and you can't afford to be wrong. If you cannot check the answer and the cost of error is high, do not ship it on the model's word.
- The tool behaved strangely. Suspected prompt injection, a tool ignoring its instructions, output that seems manipulated, or a leak of information it shouldn't have.
- You're being asked to use a non-sanctioned tool for real work, or to paste data you're unsure about.
ESCALATION QUICK-PATH
high stakes? ─┐
about people? ├─→ YES to any → STOP → ask a human before proceeding
Tier 4 data? ─┤ (manager / AI-security contact / legal)
can't verify? ─┘The most damaging incidents almost never come from people who escalated and waited an hour. They come from people who didn't know there was a line to escalate at. If you finish this module knowing only the escalation triggers, it was worth your time.