Where we are taking Cryterio

Transcription is the first step, but it isn't enough. A parliamentary chamber or a television studio carries background noise, applause, chatter and microphones that open and close. Speakers use different accents and delivery, often fast, and sometimes talk over each other. A system that transcribes word by word without accounting for this produces dirty text, and on dirty text every later check is fragile.

We work so that speech recognition stays stable in these conditions: separating voice from noise, handling overlaps and reconstructing coherent sentences even as the audio degrades. We also care about attributing who is speaking, that is, working out which speaker each sentence belongs to, because a claim only makes sense when it is tied to the person who made it.

A political speech mixes very different things: data and facts, but also opinions, promises, questions and rhetoric. Only part of it can be checked against a source. “Unemployment is at ten per cent” can be verified; “this government doesn't care about young people” cannot. The first task is to tell one from the other, avoiding both checking what isn't a fact and letting a figure slip through just because it is buried in rhetoric.

On factual claims we normalise the elements that matter: numbers, percentages, dates, institutions and units, so that “ten thousand”, “10,000” and “ten-thousand” become the same comparable value. We also discard repetitions, because in a live broadcast the same figure comes back several times and there is no point in checking it from scratch each time.

Finding a figure isn't enough: it has to be linked to the correct statistical series. An employment figure may refer to different age bands, periods and definitions, and picking the wrong table leads to a wrong verdict while still staying inside the right source. The link between what is said and the reference data point is one of the most delicate parts of the work.

That is why we maintain a curated index of primary sources, the public institutions and registers the data comes from, so we know in advance where to look. When a live search is needed, we keep it restricted to a set of authoritative public domains rather than opening it to the whole web: a wrong source is worse than no source.

A verdict is only worth something if it is calibrated. Many figures are ambiguous: rounding, implicit reference periods, definitions that change from one source to another. In these cases claiming a certainty that isn't there is a mistake, just as much as getting the figure wrong. We work so that caution is proportionate to the evidence available, not an automatic reflex.

There are four public verdicts: true, imprecise, unconfirmed, unverifiable. The underlying rule is that a claim is never called false without a primary source cited, and that between an error and an “unverifiable” we choose the latter. Getting a verdict wrong costs more than admitting we don't have enough to go on.

Live, timing matters: a verdict that arrives when the topic has already moved on is of little use. We cut the latency between the spoken word and the verdict by working on every stage, from transcription to source retrieval, so that the check appears while the sentence is still current.

The difficulty is that speed and accuracy pull in opposite directions. Shortening the timeline must not mean giving up the source: the verdict stays anchored to a cited data point even when it arrives quickly. We look for the point where the check is fast enough to be useful and solid enough to be trusted.

No verification system is free of errors. What matters is how they are handled. When a verdict turns out to be wrong we correct it in the open, without quietly removing it, because a visible correction is part of the method and not an exception to hide.

Every case that slips through is also a chance to enrich the source index and refine the link between claim and data, so that a gap that surfaced once does not recur. The criterion by which we decide what to check and how to phrase a verdict is set out in the service's ethical principles.

The method began with Italian, but its steps aren't tied to a single language: understanding speech, isolating claims, linking them to a source and calibrating the verdict hold everywhere. We are building the foundations to apply it to other European languages.

Each language, though, brings its own national primary sources: statistical institutes, registers and public archives differ from one country to another. Bringing verification into a new language therefore means not just translating, but building the index of the right sources for that context.

Not everything worth checking is live. A rally, an interview or a statement can be filmed or uploaded as a file, and the same method applies to that material: transcription, isolating the claims, linking to sources, verdict.

Compared with a live broadcast, the time constraint changes. A recorded video can be processed asynchronously, without the pressure of real time, which allows for deeper checks. It is a direction of work that extends verification beyond the moment of broadcast.

Let us say up front what this direction is not. We are not trying to tell from someone's face whether they are lying: the idea that lying has a recognisable facial signature has no scientific basis, and we do not claim it does. Any work in this area starts from that limit.

What we are evaluating is narrower: whether visual signals such as emphasis, a hesitation or a change of pace might help to contextualise a claim or attribute it more precisely, for example telling apart who is speaking in a confused passage. It would remain a secondary signal, never a judgement on the person.

The scientific and ethical limits are explicit. This would involve biometric data, with the privacy implications that carries, and a real risk of bias toward certain faces or behaviours. That is why it is a direction under evaluation and not a feature: we would only proceed if it were consistent with the rule of anchoring to the data and never to the person.