The Limits of Learning

Lessons 1 through 7 showed you what machine learning can do. This lesson is equally important: what it cannot do, where it fails, and why those failures matter. A student who only knows the power of a tool without knowing its failure modes is only half-educated. Machine learning is extraordinary — and limited in specific, predictable ways.

Limit 1: ML Needs Data — Lots of Good Data

Machine learning systems do not reason from first principles. They generalize from examples. If you do not have examples, you cannot train a model. If your examples are sparse, the model will be unreliable. This creates real problems. Rare diseases have few documented cases, making ML-based diagnosis difficult. Languages spoken by small populations have little digitized text, making translation models poor. Events that have never happened before — a new type of cyberattack, a financial crisis with no historical parallel — have no training data at all. The requirement for large, high-quality, representative datasets is the most fundamental constraint in machine learning.

Limit 2: ML Can Be Wrong — Confidently A machine learning model assigns probabilities, but those probabilities can be wildly miscalibrated. A model might classify a photo with 99% confidence as a cat when a human would instantly see it is a blurry sock. This is called overconfidence. Adversarial examples make this vivid: a researcher can add tiny, imperceptible changes to an image — differences invisible to a human eye — that cause an image classifier to completely change its prediction. The pixel values have been manipulated to exploit weaknesses in how the model represents patterns. For high-stakes uses — medical diagnosis, autonomous driving, criminal sentencing — a confident wrong answer is dangerous. This is why human oversight remains essential even for very accurate models.

Limit 3: ML Has No Understanding A language model that writes flawless-sounding paragraphs about photosynthesis does not understand photosynthesis. It has learned statistical patterns — which words tend to follow which other words across billions of documents. When you ask it a question it produces text that matches the pattern of good answers. It is not consulting a mental model of biology. This matters because it means ML systems cannot detect their own errors through reasoning. They cannot ask themselves Does this make sense? and catch a logical contradiction the way a thoughtful human can. They can produce fluent nonsense with no awareness that it is nonsense.