The Three Signatures of AI Hallucination

Why you need named patterns

Plausible is not the same as verified. Claude writes fluently whether it is right or wrong, so you cannot lean on tone or confidence to flag an error. The Claude Certified Associate - Foundations (CCAO-F) exam responds to this by teaching the specific signatures of failure, so that you can spot them quickly instead of reading every line with equal, exhausting suspicion. Knowing the pattern is what turns a vague unease into a targeted check.

There are three hallucination signatures to recognise. They are not degrees of the same thing; they are distinct shapes that a fabricated or ungrounded claim can take, each with its own tell. Learning them by name is the first step toward catching them in the wild.

Plausible-but-unsupported claims

The first and most dangerous pattern is a statement that sounds reasonable, fits the topic perfectly, and has no basis in the source you provided or in fact. Nothing about it looks wrong, because it was generated to fit, not to be true. There is no jarring detail to catch your eye, no obvious error to trip over. This is why it is the hardest of the three: the only reliable defence is checking the claim against a reference rather than waiting for it to feel off.

Fabricated specifics

The second pattern is invented detail presented with precision: a statistic, a date, a name, a quotation, a citation. Specificity reads as authority, which is exactly why fabricated specifics are persuasive. A number like "63 percent" or a named source carries an air of rigour that a vague statement does not. But precision is a property of the writing, not proof of the fact, and an invented precise figure is more dangerous than a hedged one precisely because it invites trust it has not earned.

Confident tone masking uncertainty

The third pattern is about voice. Claude rarely hedges in proportion to its actual certainty, so a genuine guess and a well-grounded fact can arrive in the same assured tone. A jurisdiction-specific legal question and a piece of arithmetic get answered with the same steadiness, even though one deserves heavy qualification and the other does not. Assurance is not evidence. When an answer that should hedge does not, the missing hedge is itself the signal.

What the CCAO-F exam trips candidates on

The first trap is believing that a hedge-free, confident answer is more likely to be accurate. The taxonomy says the opposite about confidence: because tone does not track certainty, a smooth, unhedged answer to a genuinely uncertain question is a warning, not a reassurance. The credited answer treats confidence as neutral-to-suspicious, never as evidence.

The second trap is assuming an error will "sound wrong" and therefore be easy to catch on a normal read. All three patterns are, by the nature of fluent generation, built to pass a casual read. An output that sounds fine has told you nothing. The exam rewards actively checking against sources and watching for the named signatures rather than waiting to be alerted by a jarring note that never comes.

Common misreadings to avoid

How this shows up on the exam

Domain 2 questions on this knowledge point show you an output exhibiting one of the three signatures and ask you to name the pattern or the risk. The dependable reading is to match the observation to a signature, precise-but-uncited detail to fabricated specific, an unhedged answer to a hedge-worthy question to confident-tone masking, an on-topic claim with no basis to plausible-but-unsupported, and to treat fluency as irrelevant to correctness throughout.

This taxonomy is the foundation for the rest of task statement 2.2. It leads into spotting fabricated specifics via the precision-without-source signal, internal contradiction detection in long outputs, confirmation bias in framing, and capability hallucination. It also pairs with the three reference points for evaluation, since checking against a source is how these signatures are confirmed.