Does belief revision imply that an AI has human beliefs?

No. The term refers to a formal representation of accepted information and rules for changing it.

Is contraction the same as accepting the opposite statement?

No. Giving up a commitment to p does not necessarily mean accepting not-p. The system may become undecided.

Does the AGM framework specify which source is trustworthy?

No. Source priorities and acceptance policies require additional modeling or application decisions.

AI Glossary / Computer vision and multimodal AI / Intermediate

Belief revision

Definition

Belief revision is the systematic change of a set of accepted statements when new information arrives, especially when that information conflicts with existing beliefs. In AI, it studies how a knowledge state can accommodate evidence while preserving as much justified prior information as appropriate.

What it means

An information system may conclude that a delivery is at a depot, then receive a reliable correction showing that the earlier location report was wrong. Simply adding the new statement can leave incompatible conclusions in the system.

Belief revision asks which commitments should change and which should remain. It treats beliefs as representations held by a system, without implying that the software has human experiences or feelings.

Classical logical approaches differ from probabilistic updating. A logical belief set contains accepted propositions, while a probabilistic model can assign degrees of belief. Both address changing information, but their representations and consistency requirements are not identical.

How it works

The influential AGM framework, named after Alchourrón, Gärdenfors and Makinson, distinguishes expansion, contraction and revision. Expansion adds a statement, contraction gives up a commitment to a statement, and revision incorporates new information while making the changes needed to preserve consistency when possible.

Classical AGM belief sets are closed under logical consequence. Removing one explicit sentence may therefore be insufficient: other retained statements might still imply it. Revision rules must consider those relationships, not just delete a matching text string.

Different priorities or selection rules can lead to different acceptable revisions. The framework supplies rationality conditions rather than a universal answer about which real-world source deserves trust. Practical systems may represent finite belief bases, justifications or source preferences instead of maintaining every logical consequence explicitly.

Three practical examples

A rule-based assistant withdraws a conclusion when a trusted correction invalidates one of its supporting premises.
A knowledge-base workflow resolves an inconsistency between an earlier equipment-status report and a verified replacement report.
A planning system revises assumptions about resource availability before producing an updated plan.

Think of it this way

Think of correcting a carefully organized notebook. A new fact may require changing not only one line but also conclusions that depended on it, while leaving unrelated notes intact. The analogy is limited: formal belief revision uses explicit logical relationships and change policies, whereas a human notebook may leave dependencies unstated or tolerate unresolved contradictions.

Worked example

Suppose a simple belief base contains three statements: all items in Room A are fragile; parcel P is in Room A; and courier C is available. The first two imply that P is fragile.

Now a verified correction says P is not fragile. Keeping both original premises together with the correction produces a contradiction. One possible revision withdraws the claim that P is in Room A while retaining the general room rule and courier availability. Another withdraws the general fragility rule while retaining the location claim.

Choosing between these outcomes requires priorities or additional evidence. Merely removing the derived sentence P is fragile is not enough if the retained premises still imply it. The unrelated courier statement need not be discarded just because a different part of the belief base was corrected.

Illustrative case study

Fictional example: Depotwise Planning maintains facts about parcels, storage rooms and available staff. A verified audit corrects an earlier item classification that the planning rules used to select handling procedures.

The team traces which conclusions depended on the old classification and revises those commitments according to its source-priority policy. It records the correction and the affected conclusions, then checks that unrelated staffing facts remain available.

The revised plan is reviewed against the updated evidence. This fictional case illustrates dependency-aware correction, not a claim that a logical revision framework can independently establish which report is true or replace the organization's evidence checks.

Common misconceptions and mistakes

Belief revision is not simply appending new text to a database. Contradictions and logical consequences may require changes elsewhere. It is also not the same as forgetting everything and starting over; preserving appropriate information is a central concern.

Minimal change does not always mean deleting the fewest written sentences, because representation and priority matter. A logically consistent result can still be factually wrong. In a useful technical distinction, revision often concerns new information about an unchanged world, while belief update concerns a world that has changed. Applications should state which situation they model.

Business use

The ideas are useful for knowledge systems, rule-based assistants and planning workflows that must remain coherent as evidence changes. They encourage explicit handling of corrections instead of silently accumulating contradictory assertions.

Record provenance and dependencies so that a correction can be traced through derived conclusions. Define which inputs are accepted, how competing sources are prioritized and what happens when the conflict cannot be resolved automatically. Evaluate the system using realistic correction sequences. Logical properties help assess consistency, but they do not remove the need to validate input evidence and business consequences.

Check your understanding

Why is deleting the sentence P is fragile insufficient in the worked example if both supporting premises remain?

Answer: Those premises still logically imply that P is fragile. The revision must address the relevant support structure, not only the explicit wording of the conclusion.

Related terms

Knowledge base
AGM theory
belief contraction
belief expansion
logical consistency
nonmonotonic reasoning
truth maintenance
belief update

Learn next

Write a small set of rules and facts with one derived conclusion. Introduce a conflicting correction and list several ways to restore consistency. State which information each option preserves or sacrifices, then add a source-priority rule. Compare withdrawing a belief with accepting its negation.

Sources

Alchourrón, Gärdenfors and Makinson, On the logic of theory change: Partial meet contraction and revision functions, original publication: https://doi.org/10.2307/2274239 Accessible original paper: https://fitelson.org/piksi/piksi_22/agm.pdf The parcel example is an original finite-base illustration, not a claim to implement every AGM postulate or an automatic truth-verification procedure.

Previous: Belief propagation Next: Belief state