First, the key terms.

Knowledge base
The facts and rules a system is given.
Inference rule
A condition that permits a specified conclusion or action.

The idea, explained

Before today’s generative models, many AI systems relied on human-authored facts and inference rules. That approach remains useful when policies must be explicit and the problem is well defined. A rule engine can explain which rule fired, but that does not prove that its rules are complete, fair or current. Machine learning instead estimates patterns from examples. Real applications can combine both: a model interprets a message while deterministic rules enforce account permissions.

STEP 1

Represent the policy

Write a rule in a form a machine can evaluate: if a borrower is under 12, require guardian approval. This is a policy choice encoded as logic. It is not something the computer discovered about what is fair.

STEP 2

Evaluate facts carefully

The rule needs an age and a definition of acceptable approval. Missing or contradictory information should produce a defined outcome, such as asking for clarification. Quietly treating missing age as adulthood would introduce a hidden rule.

STEP 3

Handle exceptions explicitly

A real policy may include authorized school accounts or legal exceptions. Adding exceptions without controlling priority can create conflicting rules. Review which policy is authoritative and preserve a record of the rule version used.

Follow a worked example.

A library routes requests with a rule: if a borrower is under 12, require a guardian’s approval. A language model may identify the request type, but it should not decide to ignore the approval rule because the request sounds persuasive.

Walk it through, one step at a time

  1. A borrower aged 11 requests an item. The age condition is true, so the system checks approval.
  2. Approval is absent. The correct output under this policy is “approval required,” rather than a guess about whether the request seems reasonable.
  3. A different borrower has no age recorded. The system cannot establish whether the rule applies; a separate missing-information policy is needed.

Where people get stuck

An explainable rule can still encode a mistaken assumption or an unfair policy.

Now try it yourself.

Pause before opening each answer. Explain your reasoning in your own words.

1. Check the core idea

Does replacing a learned model with explicit rules eliminate bias?

Reveal the answer & reasoning

No. Rules reflect their designers’ assumptions and can use inappropriate criteria. Both approaches need evaluation against real cases.

2. Apply what you learned

Why can a rule engine be easier to audit without necessarily being safer?

Reveal the answer & reasoning

You can inspect the rule that produced an output, but the rule may be outdated, incomplete or based on an inappropriate criterion. Auditability makes evaluation easier; it does not replace evaluation.