AI Literacy

How generative AI actually works (the 30-second version)

A large language model is trained on huge amounts of text and learns the statistical patterns of language. When you ask it something, it predicts the most likely next words, chunk by chunk, to produce a plausible answer. It is not looking anything up and has no understanding or intent — it is an extremely sophisticated pattern-completer. This one fact explains almost everything else: why it's fluent, why it sometimes invents things, and why it can't be trusted blindly.

💡 Tip: The test rewards understanding that AI generates plausible text, not verified truth. Any answer treating AI as an all-knowing oracle is wrong.

Hallucinations: confident, fluent, sometimes wrong

Because the model optimises for plausible-sounding output, it can state false information with total confidence — invented statistics, fake quotes, and especially fabricated sources and citations. This is 'hallucination'. It is not a rare glitch; it is a direct consequence of how the technology works. The fluency is what makes it dangerous: a wrong answer looks identical to a right one.

💡 Tip: When AI gives you a fact, citation, or figure you'll rely on, verify it independently. 'It sounded confident' is never enough.

Bias and stale knowledge

AI reflects the data it was trained on, so it can reproduce and amplify societal biases (in hiring suggestions, image generation, language), and it only 'knows' up to its training cut-off unless connected to live search. Treating its output as neutral or automatically current is a competence error.

💡 Tip: Watch for systematically skewed output, and check dates for anything time-sensitive.

Prompting well: context beats cleverness

The biggest lever on output quality is the input. Give the model the goal, the audience, the constraints, the format you want, and ideally an example of 'good'. Then iterate — treat it as a conversation, correcting and refining rather than accepting the first draft. Vague prompts get generic answers because the model has nothing specific to predict toward.

💡 Tip: Giving a clear example of the format you want ('few-shot prompting') is one of the most effective techniques.

Privacy, confidentiality, and AI

Anything you type into a public AI tool may be stored and used to improve future models. That makes pasting confidential, personal, or client data into a public chatbot a real data-protection risk — potentially a GDPR breach at work. The competent approach: remove or anonymise sensitive details, or use an approved, private tool that contractually doesn't train on your data.

💡 Tip: Before pasting anything into a public AI tool, ask: would I be fine with this being stored on someone else's servers? If not, anonymise it or use an approved tool.

The EU AI Act and Article 4, in plain terms

The AI Act is the EU's risk-based law for AI. Article 4 requires providers and deployers of AI to ensure staff (and others using AI on their behalf) have a sufficient level of AI literacy. Crucially it does NOT mandate a specific course, format, or certificate — the level should be proportionate to the role and the AI's risk. There's no obligation to formally 'measure' literacy, but organisations are expected to take real, recorded measures. Being able to show your people understand AI's risks and proper use is exactly the evidence regulators view favourably.

💡 Tip: The test rewards knowing Article 4 requires literacy proportionate to use — not a one-size certificate, and not nothing.

1. An AI chatbot gives you a confident answer to a factual question you'll put in a report. What should you do?

2. Why can a language model invent fake citations?

3. Which is safest when using a public AI chatbot at work?

4. Does the EU AI Act (Article 4) require a specific AI-literacy certificate?