Can inference be wrong?

Predictions can be wrong, and logical systems can receive false premises.

Can a system combine rules and learned models?

Yes, provided their roles remain clear.

AI Glossary / Graphs and symbolic AI / Advanced

Inference

Definition

Inference is the process of obtaining conclusions or outputs from available information using a specified method. In knowledge-based AI, it means deriving statements from stored facts and rules. In machine learning, it commonly means applying a trained model to an input to produce a prediction or generated response. These uses share a name but provide different kinds of justification.

What it means

The source glossary emphasizes logical inference: a knowledge-based agent derives consequences of what it has been told. A sound logical procedure produces conclusions that follow from its premises under the chosen semantics.
Machine-learning inference uses learned parameters to compute an output. Calling that output an inference does not make it a deductive consequence of verified facts. A generated explanation can be fluent while containing unsupported statements.
The context therefore matters. A team discussing inference speed may mean the time required to run a neural network, while a team discussing an inference proof may mean a chain of logical derivations.

How it works

For logical inference, a knowledge base supplies facts and relationships. A procedure applies permitted rules to establish conclusions. Forward chaining starts with established facts and derives consequences. Backward reasoning starts with a query and seeks supporting premises.
Soundness concerns whether derived statements are logical consequences. Completeness concerns whether the procedure can establish every consequence within the relevant setting. Neither property, by itself, verifies that the supplied premises accurately describe the real world.
For machine-learning inference, an application prepares an input, runs the trained model, and interprets its output. A classifier might return scores or labels; a language model might generate a sequence of tokens. Ordinary model inference uses the learned parameters rather than necessarily updating them through training.

Three practical examples

A workflow engine derives an order's eligibility from recorded approvals and a written rule. A proof trace can identify which facts supported the conclusion.
An image classifier applies a trained model to a new photograph and returns a category score. That model inference is a prediction, not a formal proof of what the photograph depicts.
A writing assistant generates an answer from a prompt. Generation is often called inference, but factual claims in the answer still need support from reliable information.

Think of it this way

Think of two ways to answer whether a parcel will leave today. One checks a checklist whose rules define permission to dispatch. Another predicts departure from patterns in historical records. Both process information to produce an answer, but “permitted by the rules” and “predicted from experience” are different claims.

Worked example

Consider an invented order workflow with five propositions:
R: the order has been reviewed.
A: the order has been approved.
E: the order is eligible.
B: transport has been booked.
G: the order has dispatch authorization.
The initial knowledge base contains facts R and A. It also contains two rules: R AND A implies E, and E AND B implies G. Assume ordinary propositional logic for this example.
Forward reasoning applies the first rule because both required facts are present. It derives E. The second rule still requires B, which has not been established. Therefore, this knowledge base entails E but does not entail G.
Failure to establish G does not mean its negation has been established. One interpretation satisfying the knowledge base has R, A, and E true, with B and G false. Another has all five propositions true. Because the premises permit both possibilities for G, they entail neither G nor not-G.
Now add B as a fact. The second rule's two conditions are supported, so the system derives G. The explanation can point to B, the derived E, and the rule connecting them. No statistical confidence score is needed for this derivation within the stated model.
For comparison, imagine a trained toy model that predicts the label “dispatch expected.” It encodes R, A, and B as 1 when recorded and 0 otherwise, and computes score = 0.4R + 0.3A + 0.2B - 0.5. Its decision rule predicts the positive label when the score is at least zero.
Using R = 1, A = 1, and B = 0 gives a score of 0.2, so the model predicts dispatch expected. This is a model output, not a derivation of G from the workflow rules. In particular, encoding an unrecorded booking as zero for a feature does not add a logical premise that transport is definitely unbooked.
The model's prediction and the rule engine's conclusion answer different questions. The prediction must not silently become dispatch authorization. The score also is not a probability merely because it is numerical.
Finally, suppose A was entered incorrectly. The logical derivation after adding B still follows from the recorded premises, but the record no longer reliably represents the real order. A valid derivation cannot repair a false input fact.

Illustrative case study

Fictional Cedar Dispatch displays predicted departure beside formal authorization. Staff interpret a positive prediction as permission to release a parcel.
Cedar changes the display to identify each output's meaning and supporting information. Authorization shows the required recorded conditions and derivation, while the predictive result is labeled as an estimate of an outcome.
The team tests missing booking records, corrected approvals, and disagreements between prediction and authorization.

Common misconceptions and mistakes

“Inference always means deduction.” Machine-learning usage also includes prediction and generation.
“A sound procedure guarantees real-world truth.” Its premises must also be accurate.
“Not proved means disproved.” That conclusion requires additional assumptions.
“A model score is a probability.” Its interpretation depends on the model and processing.
“Running inference necessarily retrains the model.” Applying a model and updating its parameters are different operations.

Business use

State what each output represents and what justifies it. For rule-based conclusions, retain the relevant premises and derivation. For predictive systems, evaluate performance on the intended operating conditions. Separate an output from the action it may inform, and specify how missing, conflicting, or outdated information affects the result.

Check your understanding

Why is E initially derivable?

Answer: Both R and A are premises.

Why is G initially unresolved?

Answer: B is not established, and the premises allow either value of G.

Why does score 0.2 not authorize dispatch?

Answer: The model predicts a different target.

Related terms

Entailment
Completeness
Forward-chaining
Prediction
Training
Deduction
soundness
model serving

Learn next

Add only G to a fresh copy of the initial knowledge base. Determine whether B follows, and test your answer by constructing a satisfying interpretation with B false.

Knowledge base

Sources

AIMA glossary, inference: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Poole and Mackworth, proofs and inference procedures: https://artint.info/2e/html2e/ArtInt2e.Ch5.S3.SS2.html Google Machine Learning Glossary, inference: https://developers.google.com/machine-learning/glossary/#inference

Previous: Industry standard data model Next: Inference attack