Responsible Use of Generative AI at Work · Unit 1 of 5 · about 10 minutes
Unit 1: What generative AI actually does, and where it fails
A generative AI tool does not look up answers. It generates plausible text, one word at a time, based on patterns learned from enormous amounts of training data. Almost every workplace AI incident traces back to someone forgetting this single fact.
Prediction, not retrieval
When you ask a colleague for last quarter's revenue figure, they either know it, look it up, or tell you they don't know. When you ask a generative AI tool, it produces the kind of text that usually follows a question like yours. Often that text is correct, because the pattern it learned was built from correct examples. Sometimes it is wrong in ways that look exactly like being right.
Some tools now add live web search on top of the model. That changes where facts can come from, not who is accountable. Cited sources still need opening and checking, because the model still writes the connecting text and can misread what it retrieved.
This is why the same tool that drafts an excellent meeting summary can also invent a court case, a statistic, or a colleague's job title without any signal that it has done so. The fluency never drops. Confidence is a property of the writing style, not of the underlying knowledge.
What these tools are genuinely good at
- Transforming text you provide: summarising, rewriting, translating, changing tone, extracting action points. When the source material is in the prompt, the tool is working from your facts, not its memory.
- First drafts: emails, outlines, job descriptions, meeting agendas. A draft you will revise is low-risk; a final version nobody checks is not.
- Explaining and brainstorming: unpacking a concept, generating options, challenging your thinking. The cost of an imperfect answer here is low.
- Routine formats: tables, checklists, templates, simple code snippets that will be tested before use.
The four risk classes
Every workplace generative AI problem falls into one of four classes. The rest of this course takes them one at a time.
- Leakage: confidential information leaves your organisation through a prompt. (Unit 2)
- Fabrication: the tool invents facts, sources or figures, and they are published or acted upon. (Unit 3)
- Misattribution: AI-produced work is passed off in ways that mislead, breach licences or damage trust. (Unit 4)
- Policy breach: use of a tool that your organisation has not approved, or use outside the rules that apply to your role. (Unit 5)
Scenario: the confident intern
A useful mental model: treat a generative AI tool like a brilliant, endlessly willing intern who joined yesterday, has read most of the internet, never says "I don't know", and signs nothing. You would happily ask that intern for a first draft. You would never let them email a client unsupervised, quote figures without a source, or sit in on a confidential board discussion.
Key takeaways
- Generative AI predicts plausible text; it does not retrieve verified facts.
- It is strongest when working on material you supply, and weakest when asked to produce facts from memory.
- The four risk classes are leakage, fabrication, misattribution and policy breach. All four are manageable with simple habits.
Knowledge check
Q1. A generative AI tool answers your question fluently and confidently. What does that confidence tell you about the accuracy of the answer?
Fluent, confident prose is what these systems produce whether they are right or wrong. Confidence in style carries no information about accuracy.
Q2. Which task carries the LOWEST inherent risk?
When the source material is in the prompt, the tool transforms your facts rather than generating facts from memory. The other three ask it to produce facts, which is where fabrication lives.