Verifying fabricated facts in AI-enabled knowledge work

Turn fluent prose into atomic claims

Extract claims before trying to improve the writing. Split compound sentences into who or what, action or property, amount, unit, population, place, period, comparison, and causal language. Mark quotations, named entities, product capabilities, policy requirements, statistics, dates, and “first,” “only,” “always,” or “never” statements for priority review.

Classify each claim as externally verifiable, internally verifiable, calculated, clearly hypothetical, opinion, or synthetic example. Then name the evidence needed. A real webpage is not enough if it does not contain the claimed fact. Agreement between two generated answers is not independent confirmation.

The NIST Generative AI Profile uses the term confabulation for confidently presented erroneous or false generated content and discusses information-integrity risks. The profile supports treating fluent output as something to verify, not as evidence of truth, while remaining voluntary risk-management guidance.

Sources: [1]

Complete the fact verification matrix

Give every material claim a row with the draft wording, claim type, consequence, required evidence, authoritative source, relevant passage or data, version and date, calculation, scope limits, reviewer, and disposition. Useful dispositions include verified as written, verified after narrowing, calculation corrected, conflicting evidence, unverifiable, hypothetical label required, or remove.

Verify entity identity before attributes. Check the current version before quoting. Reproduce calculations from labeled inputs. Trace a causal statement to evidence that actually tests causation rather than correlation. For product claims, require approved first-party evidence for the exact offering, plan, configuration, and date. Preserve rejected statements so later edits do not silently restore them.

The NIST AI RMF Measure Playbook recommends context-relevant testing, documentation of limitations, independent assessment, and monitoring. Those practices support a repeatable review process but do not determine which source is authoritative for every subject.

Sources: [2]

Worked matrix: a synthetic procurement memo

A synthetic procurement memo says a vendor was founded in 2018, supports eleven languages, reduces review time by 42 percent, stores all data in Singapore, and is the only supplier certified to a named standard. None of the statements includes a source. The reviewer creates five matrix rows instead of searching the paragraph as one block.

The official company profile supports the founding year. Current product documentation lists nine supported interface languages, so the number is corrected and its date is added. The 42 percent claim appears only in an undated reseller post with no method; it is removed. The security documentation describes selectable regions but not universal Singapore storage, so the final wording is narrowed to the documented configuration and marked for technical confirmation. The certification directory lists multiple suppliers, so “only” is deleted.

The revised memo states what was verified, what depends on configuration, and what remains unresolved. The matrix retains every source and change. This example is synthetic; it makes no allegation about a real vendor and does not establish that the resulting procurement decision is correct.

Challenge the finished document

Have an independent reviewer sample high-consequence, quantitative, comparative, and product-specific claims. Give them the final document and matrix, then require them to locate the evidence, reproduce calculations, confirm scope and date, and explain any unresolved conflict. Search for facts that escaped the matrix, especially in headings, captions, tables, footnotes, and generated summaries.

Recheck claims when the source, product, regulation, dataset, or document audience changes. Broken-link monitoring is helpful but insufficient because an accessible page can change or fail to support the statement. Report removed, narrowed, conflicting, and inaccessible claims separately rather than presenting one false precision score.

Two Prune can observe how a participant identifies and verifies unsupported factual claims inside a bounded assessment. It does not certify factual accuracy across a domain, provide legal or procurement conclusions, prove complete error detection, or make employment decisions.