Two years ago a friend of mine paid £480 for a six-week ‘AI foundations’ course. Three weeks in, he recognised the slides. They were lifted almost line for line from Andrew Ng’s Machine Learning course on Coursera, which he could have audited for nothing. He finished it anyway, mostly because he had already spent the money.
That story is the whole case for starting with free AI courses online. The material at the top of the free tier is now as good as, and sometimes identical to, what people pay hundreds for. The catch is that free means nobody is minding you, and nobody chases you when you quietly stop in week two.
The free layer of AI education is bigger than most people realise
Coursera lets you audit most of its AI specialisations at no cost. edX does the same for MIT and Harvard computer science material. Stanford publishes its CS229 lectures and notes. fast.ai’s Practical Deep Learning for Coders is free in full, videos and notebooks both. Kaggle’s micro-courses take an afternoon each. Google, Microsoft and Hugging Face all run their own learning paths and charge nothing for them.
Almost none of it gets advertised, because there is no sales team behind a free course. You have to go looking, and a lot of it sits three clicks deep inside a corporate learning portal. This map of free AI classes and where the good ones hide is a decent place to start if you would rather not spend an evening on the hunt yourself.
Three kinds of ‘free’, and they behave very differently
Free forever
Kaggle, fast.ai, Hugging Face, Stanford’s CS229 page, most university open courseware. No account, no trial clock, no upsell at module six. You can disappear for a year and come back to the same material. Download the notebooks while they are up, because universities do retire pages.
Free to learn, paid for proof
This is Coursera’s audit mode and edX’s audit track. You get the videos and, on Coursera, the readings. Quizzes and graded assignments usually sit behind the wall, and a certificate costs somewhere between $49 and $99. If the credential is the part you actually want, read what an AI course certificate really buys you before you spend anything. Sometimes the honest answer is ‘not much’. Occasionally it is ‘an interview’ at a company that uses the same platform internally.
Free sample, paid course
Two or three modules unlocked, then a paywall at the halfway mark. Nothing dishonest about it, but worth knowing before you commit six hours. Check the syllabus page: if modules four through nine show no visible content without enrolling, assume the wall is coming.
A workable order for the first few months
Plenty of people start with transformers and end up confused and slightly angry. This order is slower and far less painful.
- Python basics (two weeks). Variables, loops, functions, and reading a traceback without panicking. freeCodeCamp’s Python course is more than enough.
- pandas and NumPy (one week). Load a messy CSV and answer one real question about it. Kaggle’s micro-courses are built for exactly this.
- Classical machine learning (three weeks). Regression, decision trees, and the evaluation metrics people actually argue about. Andrew Ng’s specialisation or fast.ai’s opening lessons.
- Deep learning (three to four weeks). Neural nets, then a CNN or two. fast.ai again, or DeepLearning.AI’s short courses.
- LLMs and applied AI (ongoing). Prompting, retrieval-augmented generation, evaluation. Google’s and Hugging Face’s free paths cover this well.
That is roughly ten weeks at eight to ten hours a week. If you want the week-by-week version, with the bits you can safely skip, there is a detailed plan in how to learn artificial intelligence in 10 weeks without drowning in courses.
Red flags that a free course will waste your evening
- It is a YouTube playlist wearing a course costume. Fine as a supplement, useless as a spine.
- Every example uses Titanic or Iris. You will learn nothing about the messiness of real data.
- No code you can download and break. If the notebooks are not available, expect to watch rather than do.
- Last updated three years ago. Anything pre-ChatGPT that claims to teach LLMs should be treated as history.
- The instructor reads bullet points off the slide. You can tell within four minutes. Close the tab.
Finishing is the hard part, not finding
Completion rates for free online courses sit somewhere near 10%. The material is rarely the bottleneck; the missing deadline is. Two things fix it. Put a fixed hour in your calendar and treat it like a dentist appointment. Then pick courses with a submission attached, whether that is a Kaggle leaderboard, a fast.ai notebook, or a Hugging Face exercise. Nothing motivates quite like a scoreboard other people can see.
One course at a time, as well. The folder of fourteen bookmarked courses is where enthusiasm goes to die.
Where free stops being enough
Free courses hand you the vocabulary and working code. They rarely give you feedback on your own code, a cohort to argue with, or a brief that resembles an actual job. Once you can train a model and explain what it does, the return on yet another free tutorial drops off sharply.
That is the point to consider paying, and to be picky about it. How to tell AI training courses worth your money and this guide to choosing AI training for professionals that pays you back both cover the questions to ask first: who teaches it, what you build, and whether past students got anything out of it beyond a PDF.
What you should be able to show after 40 hours
Forty hours of free AI courses, done properly, is enough for a small portfolio. Three things are realistic.
- A notebook that takes a genuinely messy dataset and produces one clear finding, with the cleaning steps left visible rather than hidden.
- Something small and specific that works: sorting support tickets into categories, pulling dates out of rambling emails, summarising a long thread.
- A short write-up of what did not work. This is the piece hiring managers actually read, because anyone can run a tutorial and almost nobody documents the dead ends.
Post them somewhere public (GitHub, a Hugging Face Space, a plain page) and link them in one line on your CV. The courses got you the skills; the artefacts are what prove it. Then choose the next course based on what you could not build, not on which syllabus looked most impressive. Run that loop three or four times and you have most of what learning AI actually consists of.