Most people who go looking for a beginner generative AI course have already tried the do-it-yourself version. They opened ChatGPT, typed something vague, got a bland answer back, and quietly closed the tab. The distance between that experience and the one where the tool feels genuinely useful is usually six to ten hours of structured practice. Not a twelve-week bootcamp. Certainly not a £1,200 certificate with a PDF at the end.
So here’s what those hours should actually contain, how to tell a course that respects your time from one that’s recycling 2023 blog posts, and a realistic four-week route through the material if you have a job and maybe forty minutes a day.
What a beginner course should cover in the first ten hours
How the models work, at the level that actually matters
You do not need linear algebra. You do need three facts. A language model predicts the next chunk of text from patterns in its training data. It keeps no memory of you between sessions unless the product adds one. And it will invent plausible-sounding details when it doesn’t know something. Those three facts explain roughly 80% of the strange behaviour you’ll run into during your first month.
If you want that grounding before you spend money on anything, this plain-English starting point for AI beginners covers the vocabulary without the maths, which is the right order to do things in.
Prompting as a skill, not a magic phrase
Beginners write prompts the way they write search queries: three words, then Enter. A decent course drills the opposite habit — give the model context, state the format you want, and ask for a draft you intend to critique rather than publish.
A drill worth doing in week one: pick one task, say summarising a meeting transcript. Write five prompts for it, each slightly more specific than the last. Run them side by side. Most people see the output improve noticeably by version three, and that single exercise teaches more than an hour of lecture slides.
The limits, including where not to use it
Any course worth the name will spend real time on failure modes: fabricated citations, arithmetic that looks confident and is wrong, summaries that quietly drop the caveat. If a course only shows you wins, it’s marketing.
Three skills that matter more than knowing the tools
- Writing a brief instead of a sentence. The people who get good output are the ones who can describe a task precisely enough that a competent freelancer could do it cold. That’s a writing skill, and it transfers to email, specs, and client work.
- Judging output quickly. Generating ten options takes a minute. Knowing which two are usable takes taste. Practise on things you already understand — your own CV, a report you wrote last quarter — so you can spot the errors.
- Building one repeatable workflow. A single reliable process you run every week beats fifty clever prompts you tried once. Weekly newsletter draft, meeting notes, product descriptions: pick one and refine it for a month.
How to tell a decent course from a recycled one
Generative AI moves fast enough that a course dated before mid-2024 is closer to history than training. A few quick checks before you commit:
- Look for a last-updated date on the curriculum page, not just the landing page.
- Check whether the exercises produce something you can keep — a working document, a finished piece of writing, a small tool.
- See if the instructor names specific model versions and says what changed between them. Vagueness here usually means copy-paste.
- Be wary of anything selling a 500-prompt library as the main event. Prompts date quickly; the underlying method doesn’t.
There’s a surprisingly good set of free self-paced AI courses that are actually worth finishing if you’d rather test the waters before paying anyone.
What you can safely skip for now
A lot. Fine-tuning, vector databases, retrieval-augmented generation, tokenisation, agent frameworks, anything with the word “orchestration” in it. These are month-six topics, and studying them in week one mostly produces the feeling of being behind.
The same goes for tool tourism — bouncing between eight chatbots to see which one writes better haikus. Pick one model, get fluent, then branch out. There’s a useful breakdown of what to learn first in an AI course and what to skip that lays out the sequencing in more detail.
Free training that’s genuinely good, and where it hides
Google and Microsoft both publish real training, not brochures, and most of it costs nothing but an account. Google’s offerings lean toward practical use of its own models alongside general concepts. Microsoft’s path is broader and slightly more technical, with decent coverage of responsible AI and data handling.
Worth knowing: Google’s free AI courses and Microsoft’s free AI training both bury the useful material a few clicks deep, so the site navigation matters more than the marketing page suggests.
A four-week plan that fits around a real job
Week one: vocabulary and first contact. Forty minutes a day. Learn what a model, a prompt, a context window and a hallucination are. Spend two sessions just experimenting with one tool on tasks you already know how to do.
Week two: prompting and critique. Take three tasks from your actual work. Write increasingly precise prompts for each. Score the outputs honestly. Keep the prompt patterns that worked.
Week three: limits and safety. Deliberately try to break the tool. Ask it for facts it should know and check them. Ask for a citation and follow the link. This week builds the instinct that stops you pasting something wrong into a client document.
Week four: one workflow, end to end. Build the single process you’ll keep using. Document it in five steps. Run it three times and fix what breaks.
If you want to start today without creating a single account, there are free AI courses with no signup that will get you moving in the next five minutes rather than the next fortnight.
What “finished” should actually look like
Forget certificates. Four weeks in, you should be able to do four things without looking anything up. Write a prompt that reliably produces usable output for a task you do weekly. Explain to a colleague why the model made something up, and what to do about it. Decide in ten seconds whether a given task is a good fit for generative AI or a terrible one. And show someone the one workflow you built, step by step, in under five minutes.
If you can do all four, you’ve had a better return on ten hours than most people get from a paid course three times as long. If you can do two of them, you’re normal, and the fix is more practice on real tasks rather than more videos. Pick the workflow you’ll actually run next Monday and start there.