Roughly nine out of ten people who enrol in an online course never finish it. That number gets thrown around a lot, and the exact figure shifts depending on who’s counting, but the pattern holds: the drop-off in self-paced technical courses is brutal. And AI is now the category where it happens most, because the gap between what a landing page promises and what week four actually asks of you is enormous.
The good news is that this is a fixable problem. Judging an AI course well takes about half an hour, and it has almost nothing to do with how slick the sales page looks.
“Online AI course” covers three very different things
The label has stopped being useful. Right now it gets applied to at least three separate products.
ML fundamentals courses teach you gradient descent, regularisation, evaluation metrics, and enough linear algebra to read a paper without flinching. These take months, not weekends.
Applied tool courses teach you to build with existing models: prompting well, wiring up an API, putting a retrieval layer over your own documents, shipping something small. These take days and are usually the fastest route to something usable at work.
Domain courses sit in the middle. AI for clinicians, AI for accountants, AI for teachers. Quality varies wildly because subject-matter experts often make poor course designers, and course designers often know nothing about the domain.
Before you compare anything, decide which of those three you’re actually shopping for. Most confusion about whether a course is “good” is really confusion about category.
Free is a legitimate first filter
You can get a real grounding in AI for nothing. Coursera lets you audit most specialisations. fast.ai’s Practical Deep Learning course is free and remains one of the best-taught introductions anywhere. Kaggle Learn has short, sharp modules on pandas and model building. Hugging Face’s own course covers transformers better than plenty of paid programmes.
Free content gets you to the useful stage of knowing what you don’t know. That’s worth more than it sounds, because it turns a vague ambition into a specific question. “I want to learn AI” becomes “I need to understand why my validation loss diverges while training loss keeps dropping.” The second one is a course requirement you can actually shop for.
A word of caution: free courses are also the easiest to abandon, because nothing but your own discipline is keeping you there. If you’ve started four and finished none, the problem isn’t the price.
The price tag tells you almost nothing about quality
A £15 Udemy course on building LLM applications can be genuinely excellent, updated monthly by someone who ships for a living. A £2,500 bootcamp can be a set of recorded lectures from 2022 with a Slack channel nobody answers. Both outcomes are common.
What you’re actually paying for, when a premium course is worth it, is usually one of four things: structured feedback on your work, a cohort with deadlines, a credential someone recognises, or a syllabus that’s been curated so you don’t waste time on dead ends. If a course offers none of those and costs four figures, you’re paying for production values.
There’s a longer breakdown of how to read pricing in how to tell which AI courses are worth paying for, but the short version is that you should be able to name the specific thing your money buys before you hand it over.
Three questions before you pay
- Who marks my work, and how long does feedback take? “Peer review” with no rubric is close to worthless.
- When was the material last updated? In AI, anything untouched for eighteen months is a history lesson.
- What will I have built by the end, and will I own the code?
The real reason people quit is scheduling, not difficulty
Course listings routinely underestimate their own workload. A module advertised as “four weeks, 3 hours per week” often needs eight hours a week if you’re writing the code yourself rather than watching someone else write it.
Assume double the stated time and check whether that’s still viable in your actual week. An honest eight-hour-per-week commitment you can keep beats an optimistic three hours you’ll abandon in a fortnight. If your evenings are already contested, which AI classes deserve your evenings is a useful filter for cutting the list down.
One practical trick: pick a course with a fixed weekly live session, even if attendance is optional. The diary entry does more for completion rates than any amount of motivation.
Does it end in something you can point at?
The strongest predictor of whether a course sticks is whether it forces you to produce artefacts. Not certificates. Artefacts. A working script, a small app, a notebook that solves a problem you actually had.
This is why the best courses have a project spine running through them from week two rather than a capstone tacked on at the end. You learn retrieval-augmented generation by building search over your own notes. You learn classification by tagging your own email archive. The subject matter is trivial; the mechanics are what transfer.
If you’re going to build something anyway, this step-by-step guide to a first real AI project maps out the sequence, including the parts most courses skip, like setting up an evaluation set before you start tuning anything.
When the expensive, slow option is the right one
There’s a category where the free route genuinely doesn’t compete: accredited programmes. If you want the credential on a CV, access to research groups, or the kind of mathematical depth that gets you into a research engineering role, nothing on a marketplace will do it.
That’s a different purchase with a different evaluation method. It involves comparing module lists, faculty, cost per credit, and whether the online version carries the same accreditation as the residential one. A breakdown like this look at University of Michigan’s AI offerings, from EECS 492 to the online data science route shows how much variation there is even inside a single strong department.
Test-drive any course in half an hour
Don’t rely on the preview video. Do this instead:
- Jump to week four. Read the actual lecture notes or transcript. If week four is where the real content lives, that’s what you’re buying. Week one is always a welcome video.
- Find the last update date. Not the copyright line. Look for a changelog or recently added lectures.
- Check whether the instructor has shipped anything. A GitHub account, a paper, a product. Teaching credentials matter less than evidence they’ve built with the tools.
- Read the Q&A or forum. Are questions answered, and by whom? A dead forum means you’re on your own when you hit the hard part.
- Download a sample assignment. If there’s no rubric, there’s no feedback, and if there’s no feedback you’re paying for video.
The first week decides everything
Once you’ve enrolled, the window is short. Complete the setup and the first graded exercise within seventy-two hours of signing up, before the novelty wears off and before your calendar refills. Install everything, get the toy example running, break it, fix it. That single session does more for your odds than any study plan you draw up afterwards.
Then book the next four sessions in your calendar as recurring blocks with a specific task attached to each one. Not “study AI” but “finish the attention exercise and write up what confused me.” Vague blocks get skipped; specific ones get done.
One last thing worth doing: write down, in two sentences, what you expect to be able to build by the end. Re-read it at week three. If the answer has changed completely, the course is probably teaching you something more interesting than what you signed up for, and that’s usually a good sign.