Automated claim verification is a research field with more than a decade of history: the selection of claims worth checking was formalised by systems such as ClaimBuster (Hassan et al., 2017), and verification against an evidence corpus has established benchmarks such as FEVER (Thorne et al., 2018). Bringing these tasks live, on spontaneous speech and with an outcome expected within seconds, does not add a requirement: it changes the nature of the problem, because every link in the chain inherits the errors of the previous one and there is no time to correct them.
From speech to checkable claims
The flow begins with streaming transcription, segmented into utterances as it arrives, with punctuation and sentence boundaries reconstructed on the fly. The step that decides everything comes next: separating factual claims from opinions, predictions and rhetoric. «Unemployment is at 6%» can be checked; «unemployment is unacceptable» cannot; «unemployment will rise again» will be checkable, but not now. This selection task, the check-worthiness of the literature, is studied systematically in comparative evaluation campaigns such as CheckThat! (Barrón-Cedeño et al., 2020), and remains the qualitative bottleneck of the whole chain: a false negative here is a claim never examined, a false positive is compute spent on what has no answer.
From each selected claim the system extracts entities, quantities, dates and the reference interval, because a figure without a period compares with nothing: «at 6%» as of when, measured by whom, under which definition of unemployment.
The right sources for each type of claim
Evidence retrieval depends on the type of claim. A statistical figure is checked against the official source that produces it, not against an article that reports it; a past statement against the archives; a legal reference against the text of the provision in force. A graph of entities (people, organisations, places, measures) makes it possible to resolve who is talking about what and to rank sources by authority, giving primary sources precedence over secondary ones. The distinction is not documentary pedantry: in verification, distance from the primary source is a measure of accumulated error risk.
Latency and reliability: separate, don't choose
Every second saved increases the risk of losing context; every second spent moves the outcome closer to irrelevance, because the conversation is already elsewhere. The trade-off is not resolved by choosing a point on the curve but by separating two service levels: a preliminary outcome within seconds, accompanied by its declared confidence, and a consolidated verification that may arrive later, when slow sources have answered. Two presentation constraints keep this architecture honest: the preliminary is never presented as final and, when the consolidated verification contradicts it, the correction receives the same visibility as the initial outcome. A discreet correction after a visible verdict is, from the user's point of view, a form of uncorrected error.
Sources are the product
The system does not ask for trust. Every outcome exposes the sources consulted, the step that links the claim to the evidence and what could not be verified. A judgement without this chain would be an automated opinion, that is, exactly the object a verification system should fight; evidence traceability is not a premium feature, it is the very definition of the service.
The limits, declared
The hard cases remain: ambiguous claims, missing context, authoritative sources disagreeing with one another. In these cases the system exposes the disagreement instead of resolving it and flags the ambiguity instead of choosing an interpretation. It is less spectacular than a verdict, and it is the only correct behaviour when the available time is that of a live broadcast. This is the architecture Cryterio is built on: the hard part is neither transcription nor search, it is keeping what is known separate from what is still being verified, and saying so to the viewer.
References: Hassan N. et al., «Toward Automated Fact-Checking: Detecting Check-worthy Factual Claims by ClaimBuster», KDD 2017 · Thorne J. et al., «FEVER: A Large-scale Dataset for Fact Extraction and VERification», NAACL 2018 · Barrón-Cedeño A. et al., «CheckThat! at CLEF 2020», CLEF 2020.
Analytiko · 2 September 2026