The data science education market has exploded in the last decade. MOOCs, bootcamps, university electives, corporate training programmes — there’s no shortage of options. And yet, a consistent complaint surfaces from employers, managers, and even graduates themselves: “I learned a lot, but I’m still not sure how to actually use any of it.”

That gap — between completing a course and being able to do something with what you learned — is the central failure of most data science education. And it’s almost always the course’s fault, not the student’s.

The three failure modes I see most often

1. Teaching tools, not thinking

The most common mistake: courses that spend 80% of their time on syntax and libraries, and 20% (or zero) on why you’d reach for them. A student can learn to write a random forest model in Python in an afternoon. What takes far longer is understanding when a random forest is the right tool, how to evaluate whether it worked, and how to explain the results to someone who doesn’t know what a random forest is.

A course that teaches you to call sklearn.ensemble.RandomForestClassifier() has taught you very little. A course that teaches you to think about model selection has taught you something real.

2. Using toy datasets

Titanic. Iris. MNIST. These are fine for illustrating a point, but they have almost nothing in common with real data. Real data is messy. It has missing values, inconsistent formats, ambiguous categories, and domain-specific quirks that no tutorial will prepare you for.

Good courses use messy, real-world datasets — or at least realistic simulations of them. If every exercise in a course produces clean, beautifully distributed results, you’re not learning data science. You’re learning to solve crossword puzzles.

3. No emphasis on communication

Data science is a team sport. Analyses that can’t be communicated clearly have limited value. And yet most courses spend exactly zero time on how to explain a model’s output to a business stakeholder, how to write a recommendation based on data, or how to handle a room full of sceptics.

The best data scientists I know are also clear, direct writers and presenters. This is not a coincidence.

How to spot a good course

Ask these questions before enrolling:

  • What proportion of the course is projects versus lectures? The answer should be “more projects than you’d expect.”
  • Are the datasets real or toy? If they can’t answer this, the datasets are probably toy.
  • Is communication or presentation anywhere in the syllabus? If not, consider that a flag.
  • Who teaches it, and do they have industry experience? Teaching and doing are different skills, but instructors with no practical experience tend to produce a particular kind of abstract, jargon-heavy teaching that doesn’t transfer.
  • What do graduates say they were able to do differently afterwards? Not “what they learned” — what they were able to do.

The best courses are the ones that make you uncomfortable at first, because the problems are hard and the data is messy and you have to figure things out without a clean answer key. That discomfort is the feeling of actually learning.