Fact Generalization Error

Goal Knowledge Freshness Frequency Very Common Category Knowledge Retrieval Published View source on GitHub ↗

Issue

An agent takes a fact that is true only under narrow, specific conditions — a particular study population, a specific product configuration, a specific regulatory jurisdiction — and presents it as a general truth applicable broadly. The source fact isn’t misquoted; the error is in stripping away the scope that made it narrowly true and applying it as if it held universally.

Frequency: Very Common

Symptoms

  • Agent output states a narrowly-scoped fact without any of its original qualifying conditions
  • The fact is technically sourced correctly, but the source explicitly limits it to a subset the agent’s answer doesn’t mention
  • Errors increase when the query is phrased in general terms even though the best-matching retrieved source is narrow
  • Users apply the generalized claim to cases the original narrow finding never covered, with predictably worse-than-expected results

Root Cause

Retrieval and summarization compress source text, and specificity markers — sample descriptions, population definitions, configuration prerequisites — are exactly the kind of detail that gets trimmed first when a passage is condensed to fit a response, because they read as secondary to the “headline” claim. Generation models are also trained on a strong prior toward confident, general-sounding statements, since hedged and narrowly-scoped claims are less common and less rewarded in typical training distributions. The combination — compression that strips scope markers plus a generation bias toward unqualified statements — means narrow findings default to sounding general unless something actively preserves and asserts the boundary.

Example

A source study finds: "Among trial participants aged 65+ with prior
cardiovascular history, Drug X reduced recurrence risk by 34% over 24
months." The study explicitly did not enroll or draw conclusions about
patients under 65 or without prior cardiovascular history.

A user without cardiovascular history, age 40, asks a health-information
agent whether Drug X reduces recurrence risk. The agent retrieves the
study and answers: "Yes, studies show Drug X reduces recurrence risk by
34% over 24 months" — generalizing a finding specific to an older,
cardiovascular-history population to the general population, dropping
both qualifying conditions in the process.

The user's actual risk-reduction from the drug, given their different
risk profile, may be substantially different from the cited figure,
which was never established for someone in their situation.

Statistics

FindingContext
An estimated 25-40% of statistics cited from population-specific or condition-specific studies are presented without their scoping population/condition in agent-generated summariesEstimated from summarization-fidelity audits of scientific/medical source material
General-phrased user queries produce a higher rate of unscoped generalization than queries that explicitly name the population/condition of interestTypical pattern observed in retrieval-QA evaluation
Explicit scope-preservation instructions in the summarization step reduce generalization errors substantially in tested pipelinesReported range across teams that added scope-preservation prompting/checks

Mitigations

  1. Scope-marker extraction and forced retention: At ingestion, explicitly extract population/condition/configuration scope markers from source facts as structured metadata, and require the generation step to include them whenever the fact is cited, rather than treating them as droppable prose.
  2. Scope-mismatch flagging at query time: Compare the scope of the best-matching source fact against the scope implied by the user’s query, and explicitly flag when they don’t match (e.g. “this finding is specific to patients 65+ with cardiovascular history — is that your situation?”).
  3. Anti-generalization prompting: Instruct the generation step to preserve rather than compress scope qualifiers, and to explicitly state “this is specific to X” rather than defaulting to unqualified phrasing for narrow findings.
  4. Scope-aware confidence calibration: Lower the agent’s confidence framing when citing a narrowly-scoped fact for a query outside that scope, rather than presenting the extrapolation with the same confidence as an in-scope citation.
  5. Post-generation scope audit: Automatically compare generated claims against source scope metadata and flag mismatches for review before the response is finalized, particularly in health/financial/legal domains.

Production Signals

Key Metrics

MetricDescriptionAlert Threshold
scope_qualifier_retention_rateShare of cited narrow-scope facts whose scope qualifier survives into the final responseAlert if < 85%
out_of_scope_citation_rateRate at which narrowly-scoped facts are cited for queries outside their documented scope, without a scope mismatch flagAlert if > 5%
generalization_correction_rateRate of expert/user corrections identifying an over-generalized claimTrack trend; alert on sustained increase

Alerts

AlertConditionSeverityResponse
Unscoped citation in high-stakes domainA narrowly-scoped medical/financial/legal fact is cited without its qualifier for an out-of-scope queryHighRetract/correct the response, add case to scope-audit test set
Scope qualifier retention dropscope_qualifier_retention_rate falls below threshold after a summarization pipeline changeMediumReview recent summarization prompt/pipeline changes for scope-stripping regressions
  • Fact Partial Truth - closely related; generalization error is the specific case where the dropped qualifier is the fact’s applicable scope
  • Knowledge Scope Assumption Wrong - shares the scope-mismatch mechanism, framed at the level of jurisdiction/version rather than population/condition
  • Fact Probabilistic Mismatch - a related over-confidence failure where a conditional/probabilistic finding is stated as a flat certainty