Generative AI has made coherent, plausible content cheap to produce. That erodes the signals automated fact-checking has relied on, and leaves the field's claim-centric default mis-calibrated for today's online environments.
The context around a claim and its retrieved evidence carries information that content alone does not: its temporal grounding, its source attribution, and its position within web networks. Automated systems should consider both the claim and these contextual signals.
We develop this position along four arguments. Machine learning methods should account for the temporal nature of verification, which is limited by the evidence available at the moment it happens. Contextual signals are useful for verification broadly, particularly when retrieved evidence supports conflicting judgments, and for detecting varying degrees of verifiability, ambiguity and subjectivity in a claim. We find that such nuanced claims make up considerable proportions of popular benchmarks, and outline considerations for designing and evaluating holistic verification systems.
Verify the claim and what surrounds it
Rather than focusing on individual claims' factuality, verification systems should incorporate data-driven contextual information around the claim and the retrieved evidence. Current approaches, whether supervised, semi-supervised or unsupervised, mostly rely on signals inside the text. Factuality stays necessary. It is no longer sufficient.
Source attribution
Who said the claim, and in what context? Who published the supporting or refuting evidence? Primary or secondary? Any conflict of interest?
Evolving topological context
How have the claim's source and its evidence evolved over time? Who do they cite, who cites them, and how often? Have any been compromised?
Temporality
When was the claim made, and when was it verified? When was each piece of retrieved evidence published?
What professional fact-checkers do differently
Expert fact-checkers select claims by public relevance and potential harm, go to primary sources, interrogate each source for credibility and time frame, triangulate, and say plainly what is known and what is uncertain. Automated pipelines diverge at every stage.
- Select, filter and prioritize claims
- Domain research
- Source interrogation
- Primary evidence
- Temporal weighting
- Triangulation
- Write explanation
- Cite sources
- Confidence
- Check-worthiness detection
- Information retrieval
- Relevant evidence
- Secondary sources
- Veracity prediction
- Evidence alignment
- Classification
- Explanation generation
- LLM justification
Verification is constrained by its temporal context
The truth of a claim, the evidence used to assess it, and that evidence's reliability all depend on when verification happens. We separate four temporal axes. Each has its own failure mode, and a system can fail along one of them while appearing competent overall.
Reference time of the claim
The moment the claim's truth is anchored to. Failure: aligning evidence to the wrong period and confidently judging a different proposition.
Retrieval time
When evidence is fetched. Failure: rankings favor popularity over recency, so an established article can outrank a newer correction.
Publication time of evidence
When each piece of evidence was written. Failure: weighting outdated and recent evidence equally gives order-dependent, misleading conclusions.
LLM knowledge cutoff
The end of the model's training data. Failure: hallucinated or outdated answers about events past the cutoff.
Average human attention span fell from 12 seconds in 2000 to 8 seconds by 2013, shorter than a goldfish's.
Reported within days, attributed to a Microsoft Canada study.
The Microsoft report did not measure attention in seconds and did not mention goldfish. The figure came from Statistic Brain, whose sources did not support it.
The correction does not exist at retrieval time. A 2015 verifier cannot find it.
Retrieval in 2015 surfaces only the coverage from established outlets (Axis 2).
Only if the system weighs the 2017 correction by its publication date instead of treating it as one more vote (Axis 3).
RAG does not resolve temporality. Retrieval surfaces recent evidence but does not make the model use it correctly. Search augmentation has proven insufficient on fast-changing or false-premise queries, and frontier reasoning models struggle with claims that require reasoning over specific time ranges.
Build living fact-checking benchmarks
Benchmarks should reflect a changing world, as LLM evaluation already does with regularly updated sets. Because verdicts evolve, treat labels as revisable. Archive and timestamp claims and evidence, and version releases to keep results reproducible.
Source context complements claim factuality
We augment a RAG fact-checking pipeline (OpenFactCheck's RARR and FactTool implementations) with two kinds of source context and evaluate on six widely used benchmarks. Topology gives the model the domain names of each evidence source's 1- or 2-hop neighbors over three months of the CrediBench webgraphs. Source attribution either prompts the model to reason about each source's domain and motives, or supplies a numerical domain-reliability score.
Models search the web for up to 5 rounds of 5 documents per query. Results below are for GPT-5 Mini, mean ± std over 5 seeds. Qwen3-1.7B and FactTool results are in Appendix B.
| Web topology | Source attribution | |||||
|---|---|---|---|---|---|---|
| Dataset | RAG | + 1-hop | + 2-hop | + Reasoning | + Reliability score | Best gain |
| RUMORS | 0.70±.02 | 0.75±.01 | 0.72±.01 | 0.70±.04 | 0.70±.02 | +5% |
| LIAR | 0.49±.02 | 0.50±.01 | 0.50±.01 | 0.49±.01 | 0.58±.01 | +9% |
| ISOT | 0.49±.02 | 0.52±.02 | 0.53±.01 | 0.50±.01 | 0.54±.02 | +5% |
| ClaimsKG | 0.46±.01 | 0.46±.01 | 0.46±.02 | 0.47±.01 | 0.48±.01 | +2% |
| FakeCovid | 0.46±.03 | 0.57±.05 | 0.53±.05 | 0.47±.02 | 0.50±.05 | +11% |
| Benjamin | 0.63±.03 | 0.68±.04 | 0.67±.02 | 0.63±.04 | 0.62±.04 | +5% |
Topology helps on most benchmarks, with gains of up to 10.7% for GPT-5 Mini and 11.1% for Qwen. 1-hop and 2-hop perform comparably, which suggests most of the gain comes from adding structure at all. A numerical reliability score improves every dataset except Benjamin on GPT (up to 8.5%) and every dataset except RUMORS on Qwen (up to 12.4%). Prompted source reasoning is negligible on GPT but strong on Qwen, up to 12.2% on FakeCovid.
Source topology also matters because adversaries already exploit it: SEO manipulation and networks such as Pravda push low-quality sites into search rankings and scraping pipelines.
Use topology-aware methods
Hyperlink structure encodes signals that content cannot recover. Combining graph neural networks or graph transformers with LLM text embeddings is a promising route to multimodal verification.
Integrate source attribution modules
Attach context reasoning and numerical reliability signals to the retrieved evidence and to the claim itself, as fact-checking practice does. Stronger justifications can also build trust in models.
Context can help decide which sources to trust under conflict
Conflicting evidence is common. RARR aggregates predictions made from single pieces of evidence, so we can measure directly how often the retrieved evidence disagrees. In four of the six benchmarks, more than 10% of queries do. Conflicts can be temporal (sources from different periods) or factual (sources asserting opposite facts). Current LLM-RAG pipelines rank evidence by content relevance and largely ignore credibility. Telling primary from secondary sources, which source context informs, is a first step toward resolving the conflict.
Use context-aware conflict resolution modules
Professional fact-checkers accept that concurrent narratives can exist. Automated systems should handle conflicting evidence in principled ways and express the disagreement or uncertainty, using context to arbitrate. The Pope case file below shows one such resolution.
Pope Francis endorsed Donald Trump for President of the United States.
“Pope Francis Shocks World, Endorses Donald Trump for President, Releases Statement.” Republished with the same headline on Ending the Fed.
“I never say a word about electoral campaigns.” (2 October 2016)
No corresponding endorsement appears in Vatican communications. No reputable outlet reported one.
Matched on content alone, one article supports the claim and another refutes it. Both are fluent and on-topic, so text similarity gives no way to choose between them.
Two contextual moves settle it. The primary source is the Vatican Press Office, which says nothing. The only outlet reporting the endorsement is a two-week-old domain that describes itself as fantasy news.
Subjectivity on a Spectrum: Nuances are Prevalent in Fact-Checking
A verifier that strips context can give a confident verdict on a claim that was never fully specified. We audited seven benchmarks with GPT-5 Mini plus web search for three kinds of nuance.
| Dataset | % Unverifiable | % Ambiguous | % Opinion |
|---|
of LIAR claims are ambiguous, even though each is scored against one correct label.
inter-annotator agreement among five expert annotators labelling 200 claims from LIAR and ClaimsKG. Experts struggle to classify nuance too, though they report similar overall rates (Appendix A.3).
Set comprehensive benchmarking standards
Label claims for verifiability, ambiguity and subjectivity at release, and whenever the benchmark is updated, and timestamp all claims and evidence. Models can then learn to withhold confident verdicts on unverifiable claims, and accuracy can be compared across claim types.
Isn't content-only verification safer?
We take the strongest counter-position seriously and answer it point by point.
- Genetic fallacy. Judging claims by provenance risks rejecting true claims from low-reputation sources and accepting false ones from reputable sources.
- Attack surface. Backlinks and citation patterns can be manipulated, as the Pravda network and SEO campaigns show.
- Reproducibility. A content-only verdict is reproducible from the claim and a fixed knowledge base. Context depends on an unstable web.
- Scale. Context needs billion-scale temporal webgraphs, reliability databases and versioned benchmarks.
- Combine signals. Generative models erase the discriminative power of content alone. Triangulating across channels, as experts do, means no single signal decides.
- Coordinated attacks cost more. Gaming content, structure and timing together is far more expensive than gaming any one channel.
- Knowledge bases drift too. The “fixed” baseline is itself a moving target. The answer is explicit versioning and transparent, contestable reliability signals.
- Scale cost to stakes. Shared resources like CrediBench and DomainDemo are spreading the infrastructure cost. Low-stakes uses can run lighter pipelines, while elections and public health justify the full stack.
Cite this work
@inproceedings{kondrup2026claims,
title = {From Claims to Context: Holistic Information
Verification Requires Contextual Signals},
author = {Kondrup, Emma and Eldifrawi, Islam and Wu, Tong and
Yang, Zachary and Sabry, Sebastian and Bronstein, Michael and
Godbout, Jean-Fran{\c{c}}ois and Rabbany, Reihaneh and
Huang, Shenyang},
booktitle = {Advances in Neural Information Processing Systems
(NeurIPS), Position Paper Track},
year = {2026}
}