Is a degree of belief always Bayesian probability?

That is the meaning used here and in the supplied AIMA source; other uncertainty frameworks use related terminology with different formal rules.

Can beliefs change without the event changing?

Yes. New information can change the assessment of an already determined event.

Is degree of belief the same as fuzzy truth?

No. Probability concerns uncertainty; fuzzy truth concerns graded satisfaction under a chosen semantics.

AI Glossary / Graphs and symbolic AI / Advanced

Degree of belief

Definition

A degree of belief is the probability an agent assigns to a proposition in light of its information and assumptions. In the Bayesian interpretation used here, it expresses confidence that the proposition is true on a scale from zero to one. It describes uncertainty about truth, not how partially true the proposition is.

What it means

An agent might assign probability 0.7 to the proposition that a parcel has arrived. The parcel's arrival is treated as a definite event; the agent lacks complete information about it. The number describes the agent's assessment, not an event that is “70% true.”
The assessment should state its evidence. A probability before checking a delivery notice may differ from one after receiving the notice. Different agents can reasonably assign different probabilities when their information or assumptions differ.
Calling probability subjective does not make every number equally well justified. Evidence, model assumptions, consistency, and evaluation still matter. A confident-looking score from an AI system should not be treated as a reliable degree of belief without understanding what it represents.

How it works

Define the proposition and the information available. Assign a prior probability, representing the assessment before a particular new observation. A likelihood model describes how probable that observation would be under the proposition and relevant alternatives.
Bayes' rule combines these quantities to produce a posterior probability after the observation. For a hypothesis H and evidence E with positive probability, P(H given E) = P(E given H) multiplied by P(H), divided by P(E).
The calculation does not supply its own assumptions. Priors and likelihoods need justification, and duplicated or dependent evidence must not be treated as independent without support. Probability assignments also obey consistency requirements: for a proposition and its negation, their probabilities sum to one. Updating a belief changes the assessment, not necessarily the underlying world.

Three practical examples

A robot maintains probabilities over possible locations because its sensor readings are noisy. A new reading changes the distribution. The robot is not physically spread across those locations; the probabilities represent its uncertainty about where it is.
A document system estimates whether two records refer to the same item. Additional matching information can strengthen that belief, while a conflicting identifier can weaken it. The interpretation depends on the evidence model.
A planning assistant assigns a probability to a shipment arriving before a deadline. A manager combines that assessment with the consequences of delay when choosing a plan. The probability alone does not determine the best action.

Think of it this way

Think of having a sealed envelope on your desk and being uncertain whether a requested document is inside. You can revise your confidence after learning who sent it or reading a contents note. The document is either present or absent under the stated question; your information about it changes.

Worked example

Consider a constructed system evaluating whether a machine needs inspection, denoted H. Before an alert, the model assigns P(H) = 0.20. Its assumed alert model is P(alert given H) = 0.75 and P(alert given not H) = 0.25.
The probability of an alert is (0.75 × 0.20) + (0.25 × 0.80) = 0.15 + 0.20 = 0.35.
After observing an alert, the updated degree of belief is P(H given alert) = 0.15/0.35 = 3/7, approximately 42.9%.
A frequency-style illustration gives the same arithmetic: among 100 hypothetical comparable cases, the model assigns 20 to H and 80 to not H. It expects alerts in 15 and 20 of those cases respectively. Of the 35 expected alerts, 15 correspond to H.
The posterior is not 75%. That number answers the reverse question: how often an alert would occur when H holds. The calculation is conditional on the invented prior and alert model. It neither establishes an actual machine's condition nor determines whether inspection is worthwhile without considering consequences.

Illustrative case study

Fictional company Willow Deliveries develops an assistant that estimates whether a parcel has arrived. Staff initially treat a displayed probability as a final status and mark some uncertain parcels as delivered.
The team clarifies the distinction between an inferred status and a confirmed observation. It documents which evidence contributes to the probability, checks the model on relevant cases, and keeps a process for resolving uncertain records.
It also discovers that two notifications sometimes originate from the same underlying scan. Treating them as independent confirmations would exaggerate the evidence, so the model and reporting are revised accordingly.
This fictional case illustrates interpreting and updating belief under uncertainty. It does not claim that a particular probability threshold is suitable for every delivery decision or that a documented model is necessarily accurate.

Common misconceptions and mistakes

“A probability of 0.7 means the proposition is 70% true.” It represents uncertainty about the proposition in this interpretation.
“The likelihood is the posterior.” P(E given H) and P(H given E) answer different questions.
“Subjective probability is arbitrary.” A useful assessment should be reasoned, coherent, and informed by evidence.
“Repeating the same evidence strengthens it independently.” Duplicate reports can come from one source.
“The most likely outcome is always the best basis for action.” Decisions also depend on costs, benefits, and alternatives.

Business use

When presenting probabilities, identify the proposition, time horizon, and evidence conditions. Distinguish inferred assessments from confirmed facts. Ask whether a confidence score is intended as a probability and how that interpretation was evaluated.
Document important assumptions and test sensitivity when they are uncertain. Keep track of evidence provenance to avoid counting the same observation twice. Use probability estimates alongside consequences when setting review or action policies. Reassess their usefulness when operating conditions or data sources change.

Check your understanding

Why is the posterior about 42.9% rather than 75%?

Answer: Alerts also occur when inspection is not needed, and the prior matters.

What does an additional identical notification require checking?

Answer: Whether it supplies independent evidence.

Why is a probability insufficient to choose an action?

Answer: Different errors and actions can have different consequences.

Related terms

Learn next

Choose a binary proposition. State a prior and a hypothetical evidence model, calculate a posterior, and explain which assumptions most affect the result. Keep the belief assessment separate from the decision made using it.

Posterior

Sources

Russell and Norvig, AIMA glossary, degree of belief: https://github.com/aimacode/aima-glossary/blob/master/glossary.md UC Berkeley CS188, Probability lecture, beliefs, conditional probability, and Bayesian inference: https://inst.eecs.berkeley.edu/~cs188/fa22/assets/lectures/cs188-fa22-lec11.pdf

Previous: Degree heuristic Next: Degree of belonging