Is identity uncertainty limited to people?

No. It applies to objects, publications, organizations, and other entities.

Can a model represent uncertainty without merging records?

Yes. It can keep probabilities or alternative mappings separately from source records.

Does more data always resolve the issue?

No. Additional observations may be redundant, unreliable, or unable to distinguish similar entities.

AI Glossary / Graphs and symbolic AI / Advanced

Identity uncertainty

Definition

Identity uncertainty is uncertainty about whether different observations, names, or records refer to the same underlying entity. An AI system may know that two observations occurred while remaining unsure whether they concern one object or two.

What it means

Identity concerns which individual is represented, not merely its category or attributes. Two vehicles can both be blue vans without being the same van. Conversely, one vehicle may appear differently across observations because of lighting, viewpoint, damage, or recording errors.
Identity uncertainty also differs from existence uncertainty. The former asks whether references denote the same entity; the latter concerns which entities exist or how many there are. The questions can interact, because deciding that two observations share an identity changes how many distinct entities they represent.
Different database identifiers establish different records, not necessarily different real-world entities.

How it works

The system considers alternative mappings from observations to entities. One hypothesis may assign two records to one entity; another may assign them to different entities. Evidence can include descriptive fields, timing, location, relationships, and the reliability of the sources.
A probabilistic model compares how likely the observations would be under these alternatives and combines that evidence with prior assumptions. The result can retain uncertainty instead of immediately selecting one permanent mapping. BLOG is an example of a research language designed to express models involving unknown objects and uncertain identities.
With many records, the mappings must be coherent. If A and B denote the same entity, and B and C do too, then A and C must also denote that entity within the same hypothesis. Independently thresholding pairwise scores can produce contradictory decisions unless the overall grouping is handled appropriately.

Three practical examples

Citation matching: A literature system receives references with abbreviated author names, variant titles, or missing publication details. It must determine which citations identify the same publication before aggregating their information.
Equipment histories: A maintenance platform receives records from several systems with differing asset labels. Similar model numbers identify a product type, while a reliable serial number may identify an individual unit. Confusing those roles can attach repairs to the wrong asset.
Object tracking: A camera system observes an object before and after a temporary occlusion. It must assess whether the later detection belongs to the earlier track or to another object. Appearance alone may be insufficient when several objects look similar.

Think of it this way

Think of two photographs with different filenames. They could show the same bicycle from different angles or two similar bicycles. Comparing filenames answers a question about stored files; comparing evidence about the photographed objects addresses the identity question.

Worked example

Consider two equipment records. Let S mean that they refer to the same unit, and D mean different units. In this deliberately simplified example, assume prior probabilities P(S) = 0.1 and P(D) = 0.9. Evidence E is a particular agreement pattern in their recorded descriptions.
Suppose the model assigns P(E given S) = 0.8 and P(E given D) = 0.05. These are illustrative assumptions, not measured matching rates. The joint weights are 0.1 times 0.8 = 0.08 for S and 0.9 times 0.05 = 0.045 for D.
Normalize the weights: P(S given E) = 0.08/(0.08 + 0.045) = 0.64. The posterior probability of different units is 0.36. Thus an agreement pattern that is much more likely for matching records still does not prove identity.
Now use a lower prior match probability of 0.01 while leaving the same conditional evidence probabilities. The weights become 0.008 and 0.0495. The posterior match probability is 0.008/0.0575, approximately 13.9%.
For the first model, the number of distinct units represented by the pair is 1 under S and 2 under D. Its posterior expected count is 0.64 times 1 + 0.36 times 2 = 1.36. That is an expectation over alternatives, not a claim that a fractional physical unit exists.
If two fields were copied from the same source, treating their agreement as two independent pieces of evidence could exaggerate confidence. A further update would need a justified joint or conditional model rather than simply multiplying convenient scores.

Illustrative case study

Fictional Willow Assets combines inventory and repair records. An initial rule merges entries whenever their manufacturer and model descriptions match. Reviewers find that the company owns many identical models, so this rule combines the histories of different physical units.
The team separates type information from individual identifiers, records source provenance, and retains alternative links when evidence is insufficient. A review process handles uncertain matches before they change the authoritative maintenance history. Original records remain available so a mistaken link can be corrected.
It treats identity resolution as an inference problem with consequences for downstream reports, rather than as a cosmetic cleanup of duplicate-looking text.

Common misconceptions and mistakes

“Identical descriptions prove identical entities.” Distinct objects can share recorded attributes.
“Different names prove different entities.” One entity can have aliases or inconsistent records.
“A high similarity score is automatically a probability.” Its interpretation depends on how the score was produced and evaluated.
“An expected count of 1.36 describes a fractional object.” It summarizes uncertainty across possible integer counts.
“Pairwise decisions always form a coherent grouping.” Identity hypotheses must respect relationships such as transitivity.

Business use

Define the entity being resolved: a physical item, organization, publication, or another unit of analysis. Document which identifiers are reliable, which fields describe only categories, and which observations share a source.
Choose review and merge policies according to the impact of incorrect links and missed matches. Retain provenance and support corrections. When unresolved identity affects totals or other analysis, show that uncertainty rather than silently treating every tentative match as a fact.

Check your understanding

Why is the first posterior 64% rather than 80%?

Answer: It incorporates both the prior and the evidence likelihood under different-unit records.

Why does the second posterior fall?

Answer: Matching pairs are assumed much rarer beforehand.

What does agreement on a product model establish?

Answer: Shared type information, not necessarily a shared individual identity.

Related terms

Record linkage
Existence uncertainty
Bayesian inference
Probabilistic model
Entity resolution
data association
coreference
provenance

Learn next

Construct several matching and nonmatching record pairs with similar descriptions. Separate identifier evidence from category evidence, then explore how prior match rates and dependent fields affect the conclusions.

Sources

Stuart Russell, Identity Uncertainty: https://people.eecs.berkeley.edu/~russell/papers/ifsa01-identity.pdf Milch and colleagues, BLOG: Probabilistic Models with Unknown Objects: https://aima.cs.berkeley.edu/~russell/papers/ijcai05-blog.pdf Pasula and colleagues, Identity Uncertainty and Citation Matching, abstract: https://papers.neurips.cc/paper_files/paper/2002/hash/d30960ce77e83d896503d43ba249caf7-Abstract.html

Previous: Identity matrix Next: IDistance