Does conditioning always make an event more likely?

No. Evidence can raise, lower, or leave unchanged its probability.

Is conditioning the same as changing the world?

No. It updates an assessment given information. An intervention changes something and generally requires a causal model to analyze its effects.

Can the simple ratio use an event of probability zero?

No. More advanced formulations handle appropriate continuous conditioning cases; direct division by zero is undefined.

AI Glossary / Graphs and symbolic AI / Beginner

Conditioning

Definition

Conditioning is the operation of evaluating a probability distribution given specified information. It focuses the calculation on possibilities compatible with that information and renormalizes their probability. In AI, conditioning is fundamental to probabilistic inference: a system updates its assessment of unknown variables after observing evidence, using the relationships represented in its model.

What it means

The expression P(A given B), written P(A | B), asks about event A under the information that B holds. For an event B with positive probability, it equals P(A and B) divided by P(B).

The denominator matters because the reference group has changed. A probability among all possible cases and a probability among cases satisfying B answer different questions.

This entry uses the probability-theory meaning found in AI reasoning. It does not describe behavioral conditioning in psychology or the numerical conditioning of a mathematical problem. Those other uses of the same word refer to different concepts and should be named explicitly when intended.

How it works

In a finite probability model, retain the probability mass of outcomes consistent with the evidence and exclude incompatible outcomes. Divide the retained masses by their total so the resulting probabilities sum to one.

Bayes' rule provides an equivalent way to compute posterior probabilities from prior probabilities and evidence likelihoods. With uncertain observations, the evidence may be a sensor report rather than the underlying event. The model should condition on what was actually observed.

Conditioning can increase, decrease, or leave unchanged a probability. The direction depends on the modeled relationship between the query and the evidence. It is not automatically a confidence-increasing operation.

For continuous variables, conditioning on an exact value requires the appropriate density-based or more general formulation. The elementary event-ratio formula cannot be applied by dividing by the zero probability of an exact point in a continuous distribution.

Three practical examples

1. A forecasting model updates its distribution after receiving a new measurement, using the measurement model to interpret the observation.

2. A document analyst compares the frequency of late processing among tagged records with the frequency among all records, keeping the different denominators visible.

3. A probabilistic assistant revises beliefs about several possible explanations after observing a signal, rather than treating the most visually salient explanation as certain.

Think of it this way

Think of conditioning as changing the group within which you ask a probability question. Instead of asking about every item in a catalog, you ask about the items that match a stated filter.

The analogy is exact only for suitable finite examples with the correct weights. In a general probability model, outcomes need not be equally likely, and noisy evidence requires a model rather than a simple unweighted count.

Worked example

Consider a hypothetical archive of 100 records, sampled uniformly for this example. There are 30 tagged records, 20 records processed late, and 12 that are both tagged and late.

Before learning whether a sampled record is tagged, the probability it was processed late is 20/100 = 0.20. After learning that it is tagged, the relevant group contains 30 records, of which 12 were late. The conditional probability is therefore 12/30 = 0.40.

Reversing the question changes the denominator. Among the 20 late records, 12 are tagged, so P(tagged given late) = 12/20 = 0.60. The two conditional probabilities are not interchangeable.

For an untagged record, 8 of the remaining 70 were late, giving 8/70, approximately 0.114. Weighting the two conditional rates by their group proportions recovers the overall rate: 0.30 times 0.40 plus 0.70 times 8/70 = 0.20.

These are properties of the constructed archive distribution. They do not establish that adding a tag causes a delay.

Illustrative case study

Fictional case: Birch Records prepares a dashboard using the archive above. One panel says 40 percent of tagged records were late, while a colleague reports that 60 percent of late records were tagged. The team initially treats the figures as contradictory.

A reviewer writes each question with its conditioning information and denominator. Both numbers are correct because they summarize different groups. The dashboard labels are revised to identify the reference population explicitly.

The example illustrates clearer probabilistic reporting. It does not claim real operational results or infer the effect of changing the tagging process.

Common misconceptions and mistakes

Do not reverse P(A | B) into P(B | A). Their values can differ substantially, as the archive example shows.

Do not confuse conditioning with intervention. Observing that a record has a tag differs from changing its tag through an action; causal conclusions require additional assumptions or evidence.

Do not count the same observation twice as though it were independent evidence. Repeating the identical known fact does not create another independent signal. Also check whether evidence has positive probability under the model before using the elementary ratio calculation.

Business use

For business teams, conditioning clarifies statements such as given the available information and among customers with this characteristic. It encourages precise denominators and makes differences between population-level and subgroup-level claims easier to assess.

The calculation is only as relevant as the underlying probability model and evidence definitions. A historical archive may not represent future cases, and a noisy measurement should not be treated as perfect knowledge. Reporting should state what was observed, which population is modeled, and what question the resulting probability answers.

Check your understanding

Twelve of 30 tagged records are late, while 12 of 20 late records are tagged. Which ratio answers the probability of lateness given a tag?

Answer: 12/30. The conditioning information selects tagged records as the reference group.

Related terms

Learn next

Practice reversing conditional questions while keeping their denominators separate. Then study Bayes' rule and the law of total probability, followed by noisy observation models that distinguish a reported signal from the event it is intended to measure.

Sources

Poole and Mackworth, Semantics of Conditional Probability: https://artint.info/html1e/ArtInt_143.html ; Pishro-Nik, Introduction to Probability, conditioning continuous variables: https://www.probabilitycourse.com/chapter5/5_2_3_conditioning_independence.php . Relevant primary textbook passages were consulted. The archive counts and fictional dashboard case are original; scope is probabilistic conditioning.

Previous: Conditional random field Next: Conditioning case