Rules or patterns
The ordinary explanation of the difference is a calculator against a chatbot: one follows fixed instructions, the other learns from examples. It is correct, and it is not usable, because almost nothing you actually encounter is a calculator.
A more useful version asks one question. Was the behaviour written down by a person, or inferred from examples?
A digital form that rejects an empty field was written down. Somebody typed the rule. A filter that decides this message is unwanted was inferred: nobody wrote a rule for the message you received this morning, because that message did not exist when the system was built. It was shown large numbers of messages already labelled useful or unwanted, it found patterns that separate them, and it applied those patterns to something new.
The consequences are the reason the distinction is worth learning.
A rule-based system gives the same answer to the same input, and when it is wrong you can find the rule and change it. A pattern-based system produces a likely answer rather than a determined one, can give different answers to the same question, and when it is wrong there is often no single line to correct. What went wrong is distributed across the examples it learned from.
That second property is where the fairness problem lives, and it is not mysterious. If the examples carry a history, the patterns carry the history forward. There is no line of code expressing a preference. There is a system doing exactly what it was built to do, on the material it was given.
Bayesian spam filtering was published in 1998 at an artificial intelligence conference, by artificial intelligence researchers, described by its authors as probabilistic learning. On a corpus of 1,789 messages split into 1,538 for training and 251 for testing, it reached 97.1 per cent precision on unwanted mail with words alone, and 100 per cent precision with additional features (Sahami, Dumais, Heckerman and Horvitz, AAAI 1998).
Do not repeat the 100 per cent figure as a fact about spam filters. It is an artefact of a corpus of 1,789 messages, and the honest reading of that number is that the test set was small.
The reason the paper matters is definitional. Anyone who tells you that spam filters are ordinary software and artificial intelligence is something else is drawing a line that the field did not draw, in 1998 or since.
There is a third category the two-box version hides, and it is where most working systems actually sit.
A great many services are a rule-based system with a learned component inside it, or a learned component wrapped in rules. A bank's fraud process may score a transaction with a model and then apply a written threshold, above which a human being is called. A recruitment system may rank applications with a model and then apply a written rule that nobody below position fifty is shown. A navigation application predicts traffic and then applies a written rule about toll roads.
This matters for a practical reason. When something goes wrong in one of these systems, the answer to "was it the model or the rule" determines who can fix it and how fast. A wrong threshold is an afternoon's work by someone with the authority to change it. A model producing systematically different scores for one group of people is a project, and frequently a project nobody has been asked to start.
So the question to carry out of this lesson is not which box a service goes in. It is: which part of this decision was written by a person, and which part was learned, and who can change each one. That question survives contact with real systems. The two-box version does not.
A payroll system calculates tax using published rate bands, then flags any payment more than three times an employee's usual amount for review, then uses a separate component to predict which flagged payments are likely to be genuine errors. Which description is accurate?
A colleague argues that a face recognition system on an office door is not really artificial intelligence, because it just compares a photograph to a stored photograph. What is the strongest response?
Notes are kept with your account, alongside your progress and your gate claims. The lesson itself is readable without one.
This lesson has a tool
Open it and get your draft reviewed. Drag-and-drop tools need a wider screen; the review works anywhere.