Lesson 8 of 8 · 7 min read
Every lesson so far has been about getting more out of AI. This one is about the edges — the places where using it well means using it less, or differently, or not at all.
This isn’t the disclaimer section. It’s the part that separates someone who uses AI from someone you’d trust to use AI at work. The failures below are all real, all common, and all avoidable by people who know they exist.
Bias, and where it comes from
From Lesson 2: the model learned patterns from an enormous amount of human writing. That’s the whole mechanism — and it’s also the problem.
Human writing contains human assumptions. About who’s a nurse and who’s a surgeon. About which names sound professional. About whose English is “correct.” The model didn’t decide any of this. It absorbed the statistical shape of what we’ve written, and some of that shape is our prejudice.
Nobody programmed it in. There’s no line of code to delete. It’s diffused through the patterns, which is exactly what makes it hard to remove.
Where you’ll actually meet it:
- Screening CVs or applications — the highest-risk use in this lesson, and increasingly a legal one
- Writing about groups of people
- Anything where “typical” quietly becomes “correct”
The developers work hard on this and it’s much better than it was. It is not gone, and “much better” is not a defence if it goes wrong on your watch.
What not to paste
The most practical thing in this lesson.
When you paste something into a public AI tool, you’re sending it to a company’s servers. What happens next depends on the tool, the tier, and the settings — and most people have never checked any of them.
Don’t paste:
- Personal data about customers, staff, or patients — names, contact details, anything identifying
- Financial or commercially sensitive material
- Anything under NDA
- Passwords or credentials, ever
- Anything you’d struggle to explain in a headline
In the UK, this isn’t just etiquette. Customer personal data pasted into a tool your organisation hasn’t approved is a potential UK GDPR problem, and the liability is your employer’s. That’s not a hypothetical — it’s why companies have AI policies, and why “I didn’t know” hasn’t worked as a defence.
Two practical rules:
Use the approved tool. Lesson 3 said start with whatever your workplace provides. This is why. An enterprise account has terms your organisation agreed to. Your personal account doesn’t.
Anonymise. Most of the time you don’t need the real names. “Customer A” works exactly as well, and the output is identical.
If you’re not sure what your company allows — ask. That question takes a minute and has ended careers when unasked.
Hallucination is your problem, not the model’s
Lesson 2 explained why it invents things. This is about what that means for you.
There have now been enough professional disasters to see the pattern clearly: lawyers filing briefs citing cases that never existed, reports built on fabricated statistics, articles quoting people who never said it. In every one, the AI produced something plausible and a person passed it on without checking.
Note where the failure was. The model did what it does. The human did the thing that was actually negligent.
This is the part that doesn’t feel intuitive: you are fully responsible for output you didn’t write. “The AI said so” carries the same weight as “I found it on a forum.” Nobody will accept it, and they’ll be right not to.
So the duty is simple. Anything that leaves your hands with your name on it — names, numbers, dates, quotes, citations, legal or medical claims — you verify. Not because AI is bad. Because you’re the one accountable, and that was always true.
Being honest about using it
The disclosure question comes up constantly and has no universal answer. But there’s a workable line.
Nobody expects you to declare a tool. You don’t tell people you used spellcheck. Using AI to tidy your own email is not a confession.
It matters when someone is relying on the work being yours. Academic submissions. Anything where your judgement is what’s being paid for. Client work where the terms say otherwise.
The test: would the person be annoyed to find out? If yes, tell them beforehand. That instinct is more reliable than any policy, and it’s the same test you’d apply to any other shortcut.
Meanwhile: check your organisation’s policy, because it exists and probably says something specific. And know that AI output can closely reproduce existing work without either of you realising — if you’re publishing, that’s your risk to manage.
Where a human has to stay
Some decisions shouldn’t be delegated regardless of how good the output looks.
Anything about a person’s life. Hiring, firing, promotion, credit, medical, legal. Not because AI is always wrong, but because these need someone accountable who can explain the reasoning to the person affected. “The system suggested it” is not an explanation you can give someone whose application you rejected.
Anything irreversible. If you can’t take it back, a person reads it first.
Anything where you can’t check the reasoning. From Lesson 6 — if you can’t audit how it got there, you can’t sign it off.
Notice these aren’t about capability. They’d hold even if the model were flawless. Some things need a human answerable for them, and that’s a fact about accountability rather than about technology.
What you’re actually responsible for
Here’s the honest summary of the course.
AI is a fast, tireless, occasionally confidently wrong assistant. It’s brilliant at language, unreliable about facts, incapable of judgement, and it does not know the difference between right and wrong answers.
You do. That’s your job in the loop, and it doesn’t go away as the models improve.
Which brings back the thing from Lesson 1: the people who get the most out of AI are the ones who know the most about their own work. They’re the ones who can spot the wrong answer. Everything in this course has been in service of that — not making you dependent on the tool, but making you the person who can tell when it’s wrong.
Key takeaways
- Bias comes from the training data, not a decision — it’s diffused through the patterns and can’t simply be removed
- Never paste personal, confidential, or commercially sensitive data into an unapproved tool; anonymise by default
- You’re fully accountable for output you didn’t write — “the AI said so” is not a defence
- Disclose when someone is relying on the work being yours; the test is whether they’d be annoyed to find out
- Decisions about people, and anything irreversible, need a human answerable for them
- The model has no judgement. You’re the judgement.
You’ve finished
Eight lessons. You now know what AI is, how it learns, which tool to open, how to prompt it, how to iterate, where it breaks on data, what’s worth automating, and where the edges are.
That’s genuinely more than most people using AI at work every day.
Where to go next: the natural step is a structured certification that covers this ground formally and gives you something to show for it. AI+ Foundation is free, covers AI fundamentals, applications, and ethics, and needs no technical background — it’s the closest match to what you’ve just read.
If you’d rather see the full range first, browse all certifications.
Or just go and use it. Open the tool, pick something from Lesson 7’s list, and write a proper prompt. That’s the only way any of this becomes a skill.