Five words about AI trust that get used interchangeably — and shouldn’t
Confidence, accuracy, calibration, verification, and attestation all get used as if they mean the same thing. They don’t, and the differences are usually where the overstated claim hides.
Five words show up constantly in any conversation about whether an AI can be trusted — confidence, accuracy, calibration, verification, attestation — and they get used almost interchangeably in casual writing, including plenty of marketing copy. They’re not interchangeable. Each one makes a different, specific claim, and the gap between them is usually exactly where an overstated claim hides. The full glossary has formal definitions for these and related terms; this is the short version of why the distinctions matter.
Confidence is a feeling, stated as a number
Confidence is simply how sure a system says it is — “I’m 90% certain.” On its own, that number is a claim about the system’s internal state, and nothing more. It tells you nothing about whether the 90% is earned.
Accuracy is a claim about the past, in aggregate
Accuracy answers “how often is the answer right?” across a set of questions. It’s useful and it’s also incomplete: a system can be reasonably accurate overall while being badly overconfident on the specific subset of hard questions where being wrong actually costs you something.
Calibration is the relationship between the two
Calibration asks whether confidence and accuracy actually line up — do the answers a system called “90% sure” turn out to be right about 90% of the time? This is the axis that catches the failure accuracy alone can’t see: a system that’s accurate on average but overconfident on the questions that matter most.
Verification is checking a specific claim
Verification is narrower than all three — it’s the act of confirming one particular claim is true, independent of how the system that made it scores in general. A well-calibrated system can still make one wrong claim; verification is what catches that specific instance.
Attestation is a claim plus the means to check it
Attestation is the strongest of the five, because it isn’t just a claim — it’s a claim packaged with a hash-locked, recomputable record that lets someone else verify it independently, rather than trusting the person or system making it. A number without an attestation is an assertion, however well-calibrated the system behind it might be; an attestation is what turns “trust me” into “check this yourself.”
Why the distinctions matter in practice
Marketing copy has every incentive to blur these into one reassuring impression — “highly accurate,” “trusted,” “verified” — without specifying which of the five claims is actually being made, or whether it’s been measured at all. Once you know the words are doing different jobs, a claim like “99% accurate” stops being reassuring on its own — the next question is always which of the other four claims it’s standing in for, and whether anyone can check.
Every claim AEQUARA makes about its own systems is built to survive exactly that question: the AI Trust Index reports calibration, not just accuracy; the Calibration Attestation is a claim you can independently re-derive, not just one we assert. See the full glossary for the formal definitions, with sources.