A sentence can look like one thing and contain five things that must be checked separately.

Consider this invented sentence:

“The city will open three new libraries by 2028, creating 150 jobs and giving every resident access to a library within fifteen minutes of home.”

At first glance, that looks like one claim. In fact, it contains several propositions with different evidence requirements. There is a plan to open libraries. The number is three. The deadline is 2028. There is a forecast about 150 jobs. There is a further claim about universal fifteen-minute access. Each component may have a different source, time status, unit, assumption and degree of uncertainty.

If we send the whole sentence to a fact-checker, researcher or AI system as one indivisible object, a strange result can occur. Three parts may be well supported, one may be a prediction and one may be unsupported. A single label—“true” or “false”—then hides the structure that matters.

Before information can be verified well, it has to be cut at the right joints.

The answer in one paragraph

A claim is an assertion that presents something as being the case, or in some contexts presents something as deserving acceptance, evaluation or action. For factual verification, the most useful unit is a statement precise enough to be supported, contradicted or left unresolved by evidence. A good verifiable claim preserves the necessary subject, relation, time, scope, attribution, quantities and conditions while avoiding unrelated material. Long sentences often need to be decomposed into several claims; short sentences sometimes need context added back before they can be checked responsibly.

This article extends the public foundation sequence on types and representations, identity and versions, missing information and facts and verification. The present article owns the step immediately before verification: turning language into the right checkable unit.

The examples below are educational constructions. They are not reports about particular people, schools, companies or deployed AI systems. External sources are used where established research or standards are discussed.

1. Claim is not the same as sentence

A sentence is a grammatical unit. A claim is a semantic or argumentative unit.

One sentence can contain several claims:

“The school opened in 2010, moved campus in 2018 and now serves 900 students.”

• The school opened in 2010.
• The school moved campus in 2018.
• The school currently serves 900 students.

Each claim can be checked separately. The first is historical. The second is historical and geographic. The third is current and quantitative. One could be wrong while the others are right.

The reverse can also happen. A claim may span several sentences:

“Attendance fell last term. The timetable had been changed two weeks earlier. The change caused the decline.”

The causal claim is carried by the final sentence, but its meaning depends on the preceding observations and timing. Sentence boundaries are useful. They are not reliable claim boundaries.

2. Claim is not the same as fact

A claim is something asserted. A fact concerns what is actually the case.

“The bridge is closed” is a claim. If the bridge is in fact closed at the relevant time and under the relevant access definition, the claim is factual. If it is open, the claim is false. If the evidence is insufficient or conflicting, the claim may remain unresolved.

This distinction matters because claim extraction must happen before truth assessment. The extraction stage should preserve what the source actually says without silently correcting it, strengthening it or making it more plausible.

A fact-checking pipeline that rewrites a source into a stronger claim before checking it can create its own error.

3. Claim is not the same as proposition—but they overlap

Philosophers use the word proposition in several technical ways. The Stanford Encyclopedia of Philosophy notes that propositions are commonly treated as sharable objects of attitudes and as primary bearers of truth and falsity, while also warning that the concept plays several roles in philosophical theory.

For practical information work, we do not need to settle the metaphysics of propositions. We need a usable distinction:

• sentence — an expression in language;
• claim — an assertion made through language or another representation;
• propositional content — roughly, what is being asserted;
• evidence — what bears on whether that assertion is supported;
• fact status — the outcome of evaluating the claim against appropriate evidence.

This practical grammar is sufficient for reading, research, education, fact-checking and AI evaluation.

4. Not every utterance is a factual claim

Language does many things besides assert facts.

• “Close the door.” — command.
• “Is the door closed?” — question.
• “I wish the door were closed.” — expression of desire.
• “Please close the door.” — request.
• “The door is closed.” — factual claim about a state.
• “The door should remain closed.” — normative claim.
• “Closing the door will reduce noise.” — predictive or causal claim depending on context.

A verifier should not force every sentence into the same truth-checking template. A question can contain presuppositions. A command can imply beliefs about the world. A joke can contain factual material. But their communicative function differs.

Correct classification prevents wasted checking and false certainty.

5. What is a factual claim?

A practical factual claim asserts something about the world that can, at least in principle, be evaluated using evidence appropriate to the question.

Examples:

• “The meeting began at 9:12 am.”
• “The river is 280 kilometres long under this measurement convention.”
• “Version 3 contains a new eligibility clause.”
• “The organisation announced the project on 2 May.”
• “Six of the ten observed samples exceeded the threshold.”

Factual claims are not necessarily easy to verify. Historical claims may depend on incomplete archives. Scientific claims may require difficult experiments. Statistical claims may depend on sampling design. A factual claim can therefore be checkable in principle while currently unresolved in practice.

The phrase factual claim also appears in argumentation and fact-checking literature with slightly different boundaries. The useful common ground is that the claim makes an assertion whose support can be assessed rather than functioning only as a preference or instruction.

6. Claims of fact, value, definition, cause and policy

Writing instruction often distinguishes several claim families. One widely used teaching scheme includes claims of fact, definition, value, cause, comparison and policy. The exact taxonomy varies, but the categories help because different claims ask for different evidence.

Claim type	Typical question	Example
Fact	What happened? Does it exist?	The train arrived at 8:14.
Definition	What counts as X?	This activity meets the stated definition of plagiarism.
Value	How good, fair or important is it?	The policy is unfair.
Cause	Did A produce B?	The timetable change reduced attendance.
Comparison	Is A greater, better or different from B?	Method A is faster than Method B.
Policy	What should be done?	The school should extend library hours.

A value or policy claim often contains factual subclaims. “The school should extend library hours because 40% of students cannot reach the library before closing” combines a recommendation with an empirical premise. Verification can check the premise even when the recommendation also depends on values and trade-offs.

7. Why long-form AI answers create a claim-extraction problem

Long-form answers can contain dozens of factual propositions mixed with explanation, hedging, examples and transitions. A whole-answer label hides where any error sits.

Research on long-form factuality increasingly handles this by decomposing responses. The FActScore work introduced evaluation through atomic facts: smaller factual units whose support can be checked individually. Google DeepMind’s SAFE work similarly breaks a long-form response into individual facts and evaluates them using search-supported verification. Microsoft Research’s Claimify work treats extraction quality itself as a major problem because poor claim extraction can corrupt later fact-checking.

These research programmes differ in implementation and evaluation, but they share a powerful idea: verification quality depends on the quality of the unit being verified.

A paragraph is often too large. A token is far too small. The useful middle layer is a proposition-like unit that preserves enough context to be checked.

8. What is an atomic claim?

An atomic claim is a claim made small enough that its central factual assertion can be evaluated independently of unrelated assertions.

Take:

“The museum opened in 1998 and receives more than one million visitors each year.”

A useful decomposition is:

• The museum opened in 1998.
• The museum receives more than one million visitors per year.

Now each assertion has a clearer evidential target.

Atomic does not mean grammatically short. “It opened then” is short but unusable without context. Atomicity is about separating independent truth conditions, not simply reducing word count.

9. Too atomic can become contextless

Breaking language into smaller pieces creates a new danger: the claim can lose the information needed to interpret it.

Suppose a paragraph says:

“The Riverside pilot began in 2024. It reduced average waiting time by 18% in the first six months.”

Extracting “It reduced average waiting time by 18%” produces a smaller sentence but destroys the referent, time window and baseline context. The verifier no longer knows what “it” means.

A better decontextualised claim is:

“The Riverside pilot reduced the stated average waiting-time measure by 18% during its first six months in 2024.”

The rewritten version adds necessary context while preserving the original assertion. It does not add a cause, mechanism or judgement that the source did not make.

10. Atomicity and decontextualization pull in opposite directions

This tension is central to modern claim extraction. Make the claim too large and several factual units become entangled. Make it too small and the claim cannot stand alone.

Recent research describes this problem explicitly. Microsoft Research evaluates extracted claims using dimensions including coverage and decontextualization. Work on “molecular facts” argues that fully atomic facts can lose important context and proposes balancing minimality with enough information to make the claim interpretable.

The practical target is therefore not “shortest possible sentence”. It is “smallest useful independently interpretable assertion”.

That is a much better design criterion for humans too.

11. Decontextualization: add back only what the claim needs

Decontextualization rewrites a claim so that it can be understood without requiring the reader to recover essential information from surrounding text.

Common missing elements include:

• the subject hidden behind a pronoun;
• the time hidden in the previous sentence;
• the location defined earlier;
• the relevant version of a document;
• the comparison baseline;
• the population;
• the source of an attributed claim;
• the condition in an earlier clause.

The rule is conservative: restore what is necessary for interpretation, not what seems likely to be true.

“She won in 2024” can become “Candidate A won the stated contest in 2024” if the preceding context clearly identifies Candidate A and the contest. It should not become “Candidate A won the national election by a landslide in 2024” unless those details are actually present or independently established for a later verification step.

12. Faithfulness: extraction must not rewrite the source’s meaning

A claim extractor can become an accidental author.

Original:

“The study suggests the intervention may improve recall.”

Bad extraction:

“The intervention improves recall.”

The extractor removed both suggests and may. It has converted cautious evidence into certainty.

A faithful extraction preserves the epistemic force:

“The study suggests that the intervention may improve recall.”

Faithfulness matters before verification because fact-checking the stronger rewritten claim would unfairly judge a source for something it did not say.

13. Coverage: do not extract only the easy claims

An extraction can be perfectly faithful and still fail if it misses important assertions.

Imagine a long answer with twenty factual claims. An extractor returns five beautifully written claims—all correct. If the fifteen omitted claims contain the important errors, the later fact-check looks excellent while the answer remains unreliable.

This is the problem of coverage. The extraction should capture the factual content relevant to the verification objective rather than selecting only statements that are easy to isolate.

Coverage does not mean extracting every adjective or rhetorical phrase. It means not losing materially checkable content.

14. What is a check-worthy claim?

A text may contain many claims but limited verification time. Fact-checking organisations and researchers therefore use the idea of check-worthiness: which claims deserve priority for checking?

Recent multilingual research continues to treat check-worthy claim detection as a distinct stage in automated fact-checking. The important insight for everyday use is that extractable and worth checking now are different questions.

Prioritisation may depend on:

• potential harm if wrong;
• public importance;
• specificity;
• novelty;
• spread or prominence;
• whether the claim is already well established;
• whether evidence is realistically obtainable;
• whether the claim affects a decision.

A trivial claim can be verifiable but not worth urgent checking. A consequential claim can deserve verification even when checking it is difficult.

15. Attribution belongs inside the claim

Compare:

• “The project will finish in December.”
• “The contractor says the project will finish in December.”

The first looks like a prediction presented directly. The second is first a claim about what the contractor said, and only secondarily a prediction that may itself be evaluated.

Removing attribution changes what can be verified. The contractor’s statement may be easy to confirm from a dated publication even when the eventual completion date remains unknown.

Useful claim records therefore keep:

• the speaker or author;
• the exact or normalised claim;
• when it was made;
• where it appeared;
• whether the current text is quoting, paraphrasing or independently asserting it.

Attribution is not decorative metadata. It changes the proposition being checked.

16. Time belongs inside the claim

“The station is closed” is incomplete if the station reopened yesterday.

Claims about roles, prices, rules, availability, leadership, enrolment, rankings, operating status and many other subjects are time-bounded.

A decontextualised claim should therefore restore relevant time information when the source supplies it:

• observed at;
• valid from;
• valid until;
• announced on;
• effective from;
• during the study period;
• as of the publication date.

Time prevents an old factual statement from being mistaken for a current one.

17. Scope belongs inside the claim

“Students improved” can hide a population boundary. Which students? All enrolled students? Only those who completed both assessments? One class? One school? One country?

Compare:

• “Students improved.”
• “Among the 24 students with both pre- and post-test scores, the mean score increased.”

The second claim is longer but more verifiable because its population is visible.

Scope can include:

• population;
• place;
• jurisdiction;
• time period;
• dataset;
• version;
• measurement procedure;
• inclusion and exclusion rules.

A claim can be true within one scope and false when generalised beyond it.

18. Quantifiers are small words with large consequences

Words such as all, every, some, most, none, often and usually change the evidence burden.

“Some students used the resource” can be supported by one established instance. “Every student used the resource” requires coverage of the full relevant population. “Most students used the resource” requires a defined denominator and a count sufficient to establish a majority.

An extractor that deletes the quantifier changes the claim.

An extractor that adds one does the same.

19. Negation must survive extraction

“The review did not find evidence of fraud” is not equivalent to “The review found no fraud occurred.”

The first is a statement about an evidence outcome. The second is a stronger claim about the world.

Negation interacts with missingness and search coverage. A source can truthfully report that it found no evidence without thereby establishing non-existence.

When extracting negative claims, preserve both the negative operator and the object it applies to. “No significant difference was detected” should not become “The groups were identical.”

20. Modality must survive extraction

Modality marks possibility, probability, necessity, obligation and ability.

• may
• might
• could
• probably
• likely
• must
• should
• can

“The treatment may reduce symptoms” is not the same claim as “The treatment reduces symptoms.” “The system can process 1,000 records per minute” is a capability claim, not proof that it actually processed that volume in a particular run.

Removing modal language during simplification creates certainty that the source did not provide.

21. Conditionals need their condition

“If the temperature falls below 0°C, the control system activates the heater” cannot safely be extracted as “The heater activates.”

The condition is part of the claim structure.

Likewise:

“Students who complete all four practice sets are eligible for the revision workshop.”

The eligibility claim applies under a condition. Removing the condition turns a restricted rule into a universal one.

22. Comparisons need a baseline

“The new method is 30% faster” raises immediate questions:

• faster than what?
• under which workload?
• measured by latency, throughput or completion time?
• mean, median or another summary?
• over which sample?

A comparison without a baseline is often not independently checkable.

A decontextualised claim should restore the comparison target if the source supplies it. It should not guess the baseline from common sense.

23. Numbers need units and denominators

“The rate rose to 12” is not a complete quantitative claim.

Twelve what? Percent? Cases per thousand? Millimetres? Dollars? Minutes? A score out of twenty?

Percentages also need denominators. “Failure fell by 50%” could describe a fall from 2% to 1%, from 40 failures to 20 failures, or another precisely defined comparison. Each communicates a different scale.

A good claim carries the quantity, unit, denominator where needed, population, time period and comparison basis.

24. Pronouns and coreference can hide the subject

Natural language economises by using pronouns:

“The committee reviewed the proposal. It rejected it on Friday.”

There are two instances of it. The first refers to the committee as grammatical subject? No—the sentence “It rejected it” contains one pronoun for the committee and one for the proposal. A human reader resolves this easily. A claim extractor must do so explicitly before producing:

“The committee rejected the proposal on Friday.”

Wrong coreference creates wrong facts even when every remaining word is copied accurately.

25. Entity resolution comes before claim aggregation

If two sources mention “Alex Tan”, we cannot safely combine their claims until we know whether they refer to the same person.

This is why claim work depends on identity work. A claim should ideally point to resolved or at least distinguishable referents:

• person;
• organisation;
• document version;
• place;
• event;
• dataset;
• product model;
• policy or law.

An accurate claim attached to the wrong entity becomes misinformation.

For the full identity problem, use When Are Two Things the Same Thing?

26. Ambiguity should sometimes stop extraction

A claim extractor should not always force an answer.

“Jordan told Alex that he had won.”

Who won—Jordan or Alex? Without more context, the pronoun is ambiguous.

A careless rewrite chooses one interpretation and manufactures a factual claim. A responsible system preserves the ambiguity or declines to extract the contested proposition until context resolves it.

Microsoft Research’s Claimify work explicitly highlights ambiguity handling as a central extraction problem: a high-quality extractor should avoid producing a claim when it lacks confidence in the correct interpretation of the source.

Unknown is sometimes a better output than invented clarity.

27. Presuppositions are claims hiding in the floorboards

Consider:

“Why did the company stop hiding the defect?”

The grammatical form is a question. It presupposes that the company had been hiding a defect and that it stopped doing so.

A search or summarisation system that treats the presupposition as established fact can spread an unsupported accusation without ever seeing a declarative sentence.

Presupposition-aware reading asks whether the background assumption has independent support before carrying it forward.

28. Implicature is not the same as assertion

Language can imply more than it states.

“Some of the students passed” often conversationally suggests that not all passed. But in formal logic, “some” does not by itself entail “not all”.

A claim extractor designed for factual verification should be conservative about converting implicatures into explicit assertions unless the task specifically requires pragmatic inference.

What was literally asserted and what was pragmatically suggested are both meaningful. They should not be silently merged.

29. Causal claims need more than temporal sequence

“Attendance fell after the timetable changed” and “The timetable change caused attendance to fall” are different claims.

The first states sequence plus measured change. The second asserts causation.

If a source says only the first, an extractor must not produce the second. If a source explicitly claims causation, the causal relation should be retained so that the verification stage can demand evidence appropriate to a causal claim.

Claim type therefore determines evidence routing.

30. Predictions are claims about possible future states

“The project will create 150 jobs” may be a forecast rather than a current fact.

A good claim representation separates:

• who made the forecast;
• when;
• the forecast quantity;
• the target date or period;
• conditions or assumptions;
• whether the outcome has since occurred.

After the target period, the prediction can be compared with an outcome. Before then, verification may assess whether the forecast was genuinely issued and whether its assumptions are accurately represented, but it cannot convert the future into an observed event.

31. Normative claims need different evidence

“The library should stay open later” is not verified in the same way as “The library closes at 6 pm.”

The first is a recommendation. It may be supported by factual claims about demand, staffing, safety, costs and student schedules. It also depends on values and priorities.

A claim decomposition can therefore separate:

• normative conclusion;
• empirical premises;
• assumed objectives;
• trade-offs.

This is useful in essays, policy arguments and debate because disagreement may lie in the evidence, the values or both.

32. Definition claims set the boundary of later facts

“A participant counts as a completer after attending eight sessions and submitting the final task.”

This is a definition or operational rule. The later claim “82% completed the programme” depends on it.

If a second report defines completion as attending all ten sessions, the two percentages may differ without either arithmetic calculation being wrong.

Claims about counts often inherit their meaning from earlier definition claims. Good extraction keeps that dependency available.

33. Compound claims should be split where truth can diverge

A practical test for atomicity is:

Could one part be true while another part is false?

If yes, consider splitting.

“The new station opened in May and cost $200 million.”

• The new station opened in May.
• The new station cost $200 million.

The opening date could be correct while the cost is wrong.

But do not split blindly. “The average fell from 20 minutes to 15 minutes” should normally preserve both endpoints because the comparative change depends on them.

34. Claim decomposition is not sentence simplification

Sentence simplification aims to make language easier to read. Claim decomposition aims to isolate truth-evaluable content.

These goals overlap but are not identical.

A sentence can be simplified beautifully while merging two claims. Another sentence can remain grammatically complex while containing one coherent relational claim.

For verification, preserve semantic structure before optimizing style.

35. Claim extraction is not summarization

Summarization selects and compresses information. Claim extraction identifies assertions that may need checking.

A summary may legitimately omit minor details. A verification extraction may need those details if they are the exact points most likely to be wrong.

Conversely, not every rhetorical sentence deserves extraction merely because it appears in the document.

Document-level claim-extraction research has explored the relationship between central-sentence selection, decontextualisation and evidence retrieval, but the objectives remain distinct. A good workflow states whether it wants important claims, all factual claims, check-worthy claims or claims relevant to one question.

36. Claim extraction is not information extraction in general

Information extraction may identify names, dates, locations, quantities and relationships without reconstructing a complete assertion.

From:

“Dr Lim presented the revised model in Tokyo on 4 April.”

Entity extraction may return Dr Lim, Tokyo and 4 April. Relation extraction may identify presented(Dr Lim, revised model). Claim extraction aims to preserve the proposition that can be checked: Dr Lim presented the revised model in Tokyo on 4 April.

Entities are ingredients. Claims are assertions built from those ingredients.

37. A claim can contain other claims

“The report says the intervention reduced waiting time by 18%.”

This sentence contains at least two levels:

• The report says X.
• X = the intervention reduced waiting time by 18%.

The outer attribution can be checked by reading the report. The inner claim requires the report’s evidence or independent evidence.

This nested structure appears constantly in news, academic writing and AI summaries. A statement about a claim is not the same as endorsement of that claim.

38. Claim graphs are often better than flat lists

Long explanations have dependencies.

Example:

• C1: The bridge inspection found corrosion in two support elements.
• C2: The corrosion reduced the measured safety margin.
• C3: The reduced safety margin justified temporary load restrictions.

C3 may depend on C2, which depends on C1 plus engineering analysis. Treating the claims as an unordered bag loses the reasoning path.

A claim graph can represent relations such as:

• supports;
• contradicts;
• explains;
• depends on;
• defines;
• qualifies;
• supersedes;
• attributes;
• is evidence for.

This becomes especially useful for research synthesis and long AI answers.

39. Evidence retrieval improves when the claim is well formed

Search systems work better when the query preserves discriminating information.

Bad query:

“It fell by 18%.”

Better query:

“Riverside pilot average waiting time reduced 18% first six months 2024.”

The second query carries entity, metric, magnitude and time. Decontextualised claims therefore help not only verification but also evidence retrieval.

Research on document-level claim extraction has reported this connection directly: decontextualising claims can improve retrieval because the evidence search receives a clearer target.

40. Evidence alignment: one citation per paragraph is not enough

A paragraph can contain several claims but end with one citation. Which claims does the source actually support?

Good evidence alignment asks:

• Which exact claim is the citation attached to?
• Does the source support all of it or only part?
• Does the source concern the same entity and time period?
• Is the citation primary evidence, secondary synthesis or merely an appearance of the claim?
• Does the source use the same definitions?

A citation is a link between information objects. Claim extraction makes the target of that link visible.

For the wider editorial process, see How Editorial Fact-Checking and Source Verification Work. This article stays one layer earlier: defining what exactly gets checked.

41. ClaimReview shows why specific claims matter on the web

Schema.org defines ClaimReview as a fact-checking review of claims made or reported in a creative work. Its properties include a short summary of the claim reviewed. Google’s fact-check documentation likewise emphasises identifying the specific claim, attributing it to a traceable origin and keeping the fact-check analysis transparent about sources and methods.

The technical implementation may evolve across search products, but the information-design lesson is durable: a fact check needs a clearly identified claim object, not merely a topic label.

“Climate” is a topic. “Global mean surface temperature increased over period P according to dataset D” is a checkable claim.

42. Source appearance and claim origin are different

A claim can appear in many places.

A newspaper may quote a speech. A blog may quote the newspaper. A social post may screenshot the blog. A search result may summarize the social post.

There may now be four appearances but one originating claim event.

For provenance, distinguish:

• claim content;
• claim author or speaker;
• first known appearance;
• later appearances;
• transformations or paraphrases;
• fact-checks of the claim.

This prevents copy inflation from masquerading as corroboration.

43. Paraphrase can preserve meaning—or move it

Original:

“We expect the first phase to be completed before the end of the year.”

Faithful paraphrase:

“The organisation expects the first phase to be completed by year-end.”

Unfaithful paraphrase:

“The first phase will be completed by year-end.”

The second version has removed attribution and forecast status. It is shorter and stronger. That is precisely why it is dangerous.

Claim extraction should preserve semantic force before optimizing elegance.

44. Quotation is not decomposition

Copying a sentence verbatim preserves wording. It does not guarantee that the sentence is the right verification unit.

A long quotation can contain multiple claims, rhetorical framing and unresolved references. Conversely, a short quotation may lack the subject identified in the previous sentence.

Use quotations when exact wording matters. Use claim representations when verifiability matters. Keep a pointer back to the original wording so the transformation can be audited.

45. A claim should be reversible to its source context

A high-quality extracted claim should not become an orphan.

Preserve enough provenance to answer:

• Which document did this come from?
• Which passage?
• Who authored or uttered it?
• What context was added during decontextualization?
• Was the extracted wording quoted or rewritten?
• Which version of the source was used?

The W3C PROV family provides general-purpose provenance concepts for entities, activities, agents and derivations. The exact schema can vary, but the principle is simple: transformations should leave a trace.

46. Contradictions should attach to claims, not whole documents

Document A can agree with Document B on nine claims and disagree on one.

Calling the entire documents “contradictory” loses precision.

Instead:

• C1 is supported by both.
• C2 is supported only by A.
• C3 is contradicted by B.
• C4 concerns a later time and therefore does not conflict.

Claim-level comparison makes disagreement diagnosable.

47. “Supported” is not the same as “true”

An evidence source can support a claim without settling it conclusively.

Suppose a witness says a vehicle was blue. The testimony supports the claim that the vehicle was blue. A photograph under unusual lighting may support another interpretation. Additional evidence may later resolve the colour.

This is why useful claim statuses include:

• supported;
• strongly supported;
• contradicted;
• mixed evidence;
• unresolved;
• not verifiable from available evidence;
• superseded;
• not applicable to the question.

The final vocabulary should match the domain and its standards. The deeper principle is to avoid pretending that one citation mechanically converts a claim into truth.

48. The claim extraction quality test

A practical extracted claim can be evaluated against six questions inspired by current research and adapted for general use.

1. Atomicity: Can independent truth conditions be separated further without losing necessary structure?
2. Faithfulness: Does the claim preserve what the source actually asserted?
3. Decontextualization: Can a reader understand the claim without hidden essential context?
4. Coverage: Did the extraction retain the materially checkable content?
5. Minimality: Was only necessary context added?
6. Fluency: Is the claim readable enough to search, compare and verify?

These dimensions can conflict. Adding context can improve decontextualization while reducing minimality. Splitting aggressively can improve atomicity while damaging interpretability. Good extraction is therefore a controlled trade-off, not a word-count contest.

49. A worked decomposition: one sentence, six claims

Return to the opening sentence:

“The city will open three new libraries by 2028, creating 150 jobs and giving every resident access to a library within fifteen minutes of home.”

A careful decomposition might produce:

1. The city has announced or adopted a plan to open new libraries.
2. The plan specifies three new libraries.
3. The plan specifies completion by 2028.
4. The plan projects 150 jobs associated with the libraries.
5. The plan claims the completed library network will place every resident within a fifteen-minute access threshold.
6. The fifteen-minute threshold uses a particular, still-to-be-specified travel mode or accessibility definition.

The sixth item may not be an asserted claim if the original gives no definition. It may instead be an identified information gap: fifteen minutes by walking, public transport or car? The extractor should not invent the answer. It should mark the unresolved scope needed to verify the fifth claim.

This is a critical distinction. Claim extraction can reveal that a statement is not yet sufficiently specified to check.

50. Worked example: education

Sentence:

“After four weeks of practice, the class improved dramatically because the new worksheets were better.”

Candidate claims:

• The class completed four weeks of practice.
• A measured class outcome improved after the practice period.
• The magnitude of improvement can be described as “dramatic” under some unstated criterion.
• The new worksheets caused the improvement.
• The worksheets were better under some unstated evaluative criterion.

Notice how the original sentence moves from chronology to magnitude to causation to value without signalling the evidence burden for each step.

A good teacher can turn this into a reasoning exercise: which claims are descriptive, quantitative, causal and evaluative? What evidence would each require?

51. Worked example: science

Sentence:

“The treated plants grew 12% taller, proving that the fertilizer accelerates growth.”

Candidate claims:

• Under the experiment’s measurement procedure, the treated plants were 12% taller on the reported outcome than the comparison plants or baseline.
• The experiment establishes a causal effect of the fertilizer on growth.
• The causal effect is appropriately summarised as accelerating growth.

The first claim can be checked against the dataset and calculation. The second depends on experimental design, controls, replication and uncertainty. The word proving may overstate what one experiment supports.

The extraction stage should preserve the source’s strong wording. The verification stage can then challenge it.

52. Worked example: statistics

Sentence:

“Eighty per cent of users prefer the new interface.”

To make the claim verifiable, we need to know who counts as users, how preference was measured, the sample size, the survey period and whether the result refers to respondents or the entire user population.

If the source actually says “80% of 250 survey respondents preferred the new interface”, the decontextualised claim should preserve that denominator rather than inflate the sample result into a universal population claim.

53. Worked example: history

Sentence:

“The reform failed because citizens opposed it.”

This compact sentence hides at least three questions:

• What counts as failure?
• What evidence establishes citizen opposition?
• What evidence establishes that opposition caused the failure rather than merely accompanying it?

Historical claims often depend on sources created by actors with different positions, incomplete archives and later interpretation. Decomposition makes the argument inspectable without pretending the past is simple.

54. Worked example: news

Headline:

“Experts say new policy will save families $500 a year.”

Extract:

• Identified experts or a named analysis estimates annual savings.
• The estimated saving is $500 under specified assumptions.
• The population described as “families” needs definition.
• The saving is a forecast, not yet an observed annual outcome.

A useful fact-check can now investigate the model, assumptions, household types and source attribution instead of arguing about the headline as one undifferentiated object.

55. Worked example: AI answer

Imagine an AI answer:

“The tower was completed in 1931, was the tallest building in the region for twenty years, and its architect later designed the central station.”

The answer contains at least three factual claims:

• The tower was completed in 1931.
• The tower was the tallest building in the relevant region for twenty years.
• The tower’s architect later designed the central station.

The second claim also hides a definition of region, a height metric and a twenty-year interval. The third requires identity resolution for the architect and the station.

If two claims are supported and one is wrong, “the answer is mostly right” is less useful than knowing exactly where the error is.

56. A practical claim-extraction pipeline

1. Freeze the source. Record the exact text, version and provenance before rewriting.
2. Identify candidate assertions. Separate assertions from questions, commands, rhetoric and pure transitions.
3. Resolve referents. Identify pronouns, aliases and entities where context permits.
4. Separate independent truth conditions. Split compound claims where one part could be true and another false.
5. Restore necessary context. Add subject, time, scope, baseline and attribution when present in the source context.
6. Preserve epistemic force. Keep may, likely, according to, estimates, alleges and similar qualifiers.
7. Preserve negation and conditions. Do not drop logical operators.
8. Mark ambiguity. Do not invent a single reading when the source supports several.
9. Link back to source span. Make the transformation auditable.
10. Classify claim type. Factual, causal, predictive, quantitative, normative and so on.
11. Prioritise if needed. Separate all extracted claims from check-worthy claims.
12. Only then retrieve and evaluate evidence.

The order matters. Evidence search performed against a malformed claim can return impressive sources for the wrong proposition.

57. A compact claim record

A public, domain-neutral claim record can be imagined with fields such as:

• claim text — the independently interpretable assertion;
• source span — where it came from;
• source — document, speech, dataset or other origin;
• speaker/author — who asserted it;
• subject — entity or event concerned;
• predicate/relation — what is asserted about the subject;
• object/value — relevant entity, state or quantity;
• time — when the assertion applies;
• scope — population, place, jurisdiction or version;
• modality — asserted, possible, likely, required, predicted;
• claim type — descriptive, quantitative, causal and so on;
• extraction confidence — how confident we are in the interpretation;
• evidence status — not checked, supported, contradicted, mixed or unresolved;
• provenance — how the normalised claim was derived from the source.

The exact schema will vary. The important design decision is to keep extraction uncertainty separate from truth uncertainty. We can be unsure what a sentence means even before we ask whether that meaning is true.

58. Two kinds of uncertainty

Interpretation uncertainty: What is the source actually asserting?

World uncertainty: Is that assertion true?

These can combine in four ways:

Interpretation	Evidence about world	Situation
Clear	Strong	Easy verification case
Clear	Weak	Claim understood but unresolved
Ambiguous	Strong for one reading	Need to resolve meaning before verdict
Ambiguous	Weak	Do not manufacture a confident claim

This distinction is extremely useful for AI systems because a confident verifier cannot repair an incorrectly extracted proposition.

59. What not to extract as a factual claim

Depending on the task, avoid treating the following as ordinary factual claims:

• pure questions without asserted presuppositions;
• commands;
• greetings;
• discourse markers such as “however”;
• purely expressive exclamations;
• placeholder examples clearly marked as fictional;
• quoted claims when the current author merely reports them—unless attribution is retained;
• ambiguous fragments that cannot be safely resolved;
• definitions that need their own rule-aware treatment;
• normative recommendations when the verification task only concerns empirical propositions.

The list is task-dependent. A legal analysis may care deeply about commands and obligations. A rhetoric analysis may care about questions and implicatures. The extraction objective should be stated before the labels are applied.

60. Common failure: over-splitting

Original:

“The average wait fell from 20 minutes to 15 minutes.”

Bad over-split:

• The average wait was 20 minutes.
• The average wait was 15 minutes.

Without time or baseline labels, the two claims appear contradictory. The change relation has disappeared.

Better:

“The stated average wait decreased from a baseline of 20 minutes to 15 minutes in the comparison period.”

Atomicity should preserve the relation that gives the numbers meaning.

61. Common failure: under-splitting

Original:

“The programme began in January, reached 5,000 users by June and cut costs by 30%.”

This contains at least three independently checkable claims. Keeping them fused means one verdict cannot explain where disagreement lies.

Under-splitting is especially dangerous when a sentence combines an easily verified fact with a difficult causal or evaluative claim. The verified part can lend borrowed credibility to the unsupported part.

62. Common failure: context invention

Original:

“The rate rose by 12%.”

If the source context does not define the rate or comparison period, extraction should not invent them.

Better output:

“The source asserts that an unspecified rate rose by 12%; the metric or baseline is not established in the available context.”

Sometimes the best extraction includes a missing-information flag.

63. Common failure: qualifier deletion

Words that are easy to delete are often exactly the words that control truth conditions.

• approximately;
• at least;
• up to;
• may;
• according to;
• in the sample;
• during the first quarter;
• under laboratory conditions;
• preliminary;
• estimated.

Removing them creates a different claim.

64. Common failure: attribution deletion

“Officials estimate costs will reach $2 billion” becomes “Costs will reach $2 billion.”

The source of the estimate and its forecast status vanish. A later system may present the rewritten sentence as an observed budget fact.

Preserve who says what.

65. Common failure: source wording mistaken for world wording

“The report describes the programme as successful” is a claim about the report’s language.

“The programme was successful” is an evaluative claim about the programme.

The first can be verified by reading the report. The second requires a definition of success and evidence about outcomes.

This distinction is one of the simplest ways to prevent summaries from becoming stronger than their sources.

66. Common failure: check-worthiness mistaken for truth

A claim can be worth checking because it is consequential, surprising or widely shared. None of those properties makes it false.

Likewise, an ordinary-looking claim may be false but low priority.

Check-worthiness is a triage decision. Truth assessment is an evidential decision. Keep the two labels separate.

67. For students: the six-part claim test

1. Who or what is this about?
2. What exactly is being asserted?
3. When and where does it apply?
4. Which words limit the claim? All, some, may, likely, approximately?
5. Can one part be true while another is false? If yes, split it.
6. What evidence could support or contradict it?

This works in comprehension, argumentative writing, Science, Humanities and data interpretation.

68. For teachers: teach the boundary of the answer

Students are often told to “use evidence” before they are taught to identify the exact claim that the evidence must support.

Try reversing the order:

1. Underline the claim.
2. Circle the qualifiers.
3. Box the subject and time scope.
4. Split independent assertions.
5. Then match evidence.

Give students a paragraph containing one observation, one interpretation, one causal claim and one recommendation. Ask them to label each. The exercise trains both writing and reading.

69. For researchers: define the claim before selecting the evidence

Research quality depends on alignment among question, estimand or target, measurement, analysis and conclusion.

A statistical model can answer a narrow association question while a paper’s conclusion drifts into a broader causal statement. A laboratory measurement can establish an effect under controlled conditions while a discussion paragraph generalises beyond the tested environment.

Claim decomposition makes scope drift visible. Write the central conclusions as independent propositions and ask what exact evidence supports each one.

For systematic evidence synthesis, connect this article with How Systematic Reviews and Evidence Synthesis Work.

70. For editors: claim maps make fact-checking faster

Before publication, editors can turn a draft into a claim map:

• claim;
• source;
• source type;
• date;
• status;
• citation;
• remaining issue.

This exposes unsupported numbers, duplicated source lineages, version mismatches and paragraphs whose citations support only part of the prose.

The objective is not to turn writing into a spreadsheet. It is to make verification precise enough that the prose can remain fluid.

71. For AI systems: claim extraction should be inspectable

For AI-generated long-form answers, a useful architecture separates generation from factual inspection.

1. Generate or receive the answer.
2. Freeze the exact output.
3. Extract candidate factual claims.
4. Resolve essential context.
5. Mark ambiguous extractions.
6. Retrieve evidence for each claim.
7. Evaluate support and contradiction.
8. Return corrections or confidence at claim level.

The details can vary substantially. The important separation is conceptual: the component that decides what was said should not silently change the claim while the component that checks whether it is true is working.

This article describes a public reasoning pattern, not a claim about any particular deployed system.

72. For knowledge graphs: claim nodes can preserve disagreement

If a knowledge graph stores only “facts”, disagreement becomes difficult to represent.

A claim-aware graph can represent:

• Source A claims X.
• Source B claims not-X.
• Evidence E1 supports X.
• Evidence E2 undermines X.
• Review R assigns status disputed.
• Later record V supersedes an earlier claim.

The graph does not need to choose prematurely between erasing conflict and treating every statement as equally valid. It can represent the epistemic structure itself.

73. Why claim extraction belongs between language and evidence

The foundation sequence can now be seen as a chain:

1. Type: What kind of thing are we dealing with?
2. Identity: Which particular thing is it?
3. Missingness: What do we not know, and why?
4. Claim: What exactly is being asserted?
5. Evidence: What bears on that assertion?
6. Fact status: What conclusion is justified?
7. Decision: What should we do with the result?

Skipping the claim layer makes evidence matching unstable. A search engine cannot retrieve the right evidence for a proposition that has not been specified. A verifier cannot decide whether a citation supports wording whose scope keeps moving.

Claim extraction is therefore not a minor preprocessing trick. It is the hinge between language and verification.

74. The minimal rules

1. Sentence is not claim.
2. Claim is not fact.
3. One sentence may contain many claims.
4. One claim may depend on several sentences.
5. Split where truth can diverge.
6. Do not split away the relation that gives meaning.
7. Restore necessary context, not guessed context.
8. Keep attribution.
9. Keep time.
10. Keep scope and denominators.
11. Keep negation, modality and conditions.
12. Do not resolve ambiguity by invention.
13. Preserve the link back to source wording.
14. Separate extraction confidence from truth confidence.
15. Check the claim, not merely the paragraph around it.

75. The final claim checklist

1. What exactly is asserted?
2. Who asserts it?
3. Which entity or event is it about?
4. When does it apply?
5. Where or to whom does it apply?
6. What quantities, units or denominators matter?
7. What qualifiers change its strength?
8. Is it conditional?
9. Is it one claim or several?
10. Can it stand alone without hidden context?
11. Did the extraction add anything the source did not say?
12. Is any ambiguity unresolved?
13. What evidence could support it?
14. What evidence could contradict it?
15. Can a reviewer trace it back to the exact source span?

If those questions have good answers, the claim is ready for serious verification.

Conclusion: the unit of truth has to be built before it can be checked

Language arrives in paragraphs, headlines, transcripts, charts, conversations and long AI answers. Evidence does not naturally attach itself to those containers. It attaches to specific propositions about specific things under specific conditions.

That is why claim extraction matters.

A good claim is small enough that disagreement can be located, large enough that meaning survives, faithful enough that the source recognises itself, explicit enough that evidence can find it, and traceable enough that a reviewer can return to the original context.

Once that unit exists, the rest of the knowledge process becomes cleaner. Sources can be matched. Contradictions can be attached to the right proposition. Statistics can keep their denominators. Predictions can remain predictions. Attributions can remain attributions. Unknowns can remain unknown until evidence resolves them.

Before asking whether a statement is true, first decide exactly what statement you are asking the world to answer.

Sources and further reading

For current work on extracting factual claims from long-form model outputs, see Microsoft Research, Towards Effective Extraction and Evaluation of Factual Claims. For long-form factuality evaluation through individual facts, see Google DeepMind, Long-form factuality in large language models. For document-level claim extraction and decontextualisation, see the ACL Anthology paper Document-level Claim Extraction and Decontextualisation for Fact-Checking. For the structured web representation of fact-checked claims, see Schema.org ClaimReview and Google Search Central’s fact-check structured-data documentation. For provenance, see the W3C PROV Overview. For philosophical background on propositions, see the Stanford Encyclopedia of Philosophy.

Within eduKateSingapore, continue through What Is a Fact?, How Editorial Fact-Checking and Source Verification Work, How Research Methods and Source Evaluation Work, and Knowledge Graphs and Semantic Data.