What You Need to Know
Objective 2.1 gave you the reviewer's stance. This objective names the three failure types a review looks for. A hallucination is content that sounds convincing but is not grounded in fact. An inconsistency is an output that disagrees with itself. Bias is skewed framing, assumptions or language. Each has a recognisable shape, and each needs a different response.
A garden centre manager gives Claude three jobs in one week. She wants the steps for the stock report in the till system, a spring sales summary with a table, and appraisal notes for seasonal staff. Each job can fail in one of the three ways.
Hallucination: Fluent, Specific, Fabricated
The Help Center says Claude "can occasionally produce responses that are incorrect or misleading". It "can display quotes that may look authoritative or sound convincing, but are not grounded in fact". That last phrase is the working definition. A hallucination is not vague or hesitant. It arrives as an exact-looking figure, a named source or a numbered list of steps. An invented detail comes with the same confidence as a real one, so specificity is not evidence. Ask of any detail "where would this have come from?" Asking "does this sound right?" tests only the confidence.
Claude can also hallucinate its own capabilities. A separate Help Center article warns that Claude may falsely claim to have sent an email, produced an external document or supplied a working link. It has no access to email or file transfer unless they have been explicitly integrated. A claim that Claude has done something outside the conversation is a claim to check, not a delivery confirmation. If Claude says it has emailed the stock figures to the supplier, the manager checks with the supplier.
Where the Risk Concentrates
Hallucination risk tracks grounding. It is highest where Claude has least to work from:
- your organisation's private, internal systems and documents Claude was never given
- very recent events that postdate its training data
- precise figures that were not in the source material
The highest-risk shape is confident numbered instructions for an internal tool Claude has never seen documentation for. They are fluent, specific and built on nothing. The stock-report steps have this shape. Claude names menus in a till system it was never shown, and some may not exist. The less grounding Claude could possibly have, the less its specificity means.
Inconsistency: The Output Disagrees With Itself
An inconsistency needs no external source to detect. Common forms are a number in the prose that contradicts the table, a recommendation that contradicts the findings, and two sections giving different dates for one event. An internal contradiction is a reliability warning about the whole output, not a single typo to patch. The process that produced one visible failure had every chance to produce invisible ones. So re-verify the whole output against its source, rather than fixing the one mismatch you noticed and sending the rest. If the spring sales prose and table give different compost figures, the manager checks the whole summary against the till data.
Bias: The Pattern, Not the Sentence
Bias is most consequential in outputs about people: recruitment material, appraisals, customer descriptions, and anything that ranks or characterises groups. Anthropic's Usage Policy shows the stakes by treating employment, housing, finance and healthcare decisions about individuals as high-risk uses. The skew sits in the pattern, so read across descriptions rather than sentence by sentence. Suppose the appraisals describe older staff by reliability and younger staff by innovation. No single sentence is objectionable, and the output still is. Look for coded language and unstated assumptions about age, gender, background or ability. Then apply the read-aloud test: would the framing survive being read aloud to the group it describes?
In the exam
The exam tests whether you can tell the three failure types apart and respond to each correctly, because the correct responses differ. For inconsistency, the tested point is the response: re-verify the whole output. Distractors cross the wires. They patch one contradiction, delete one loaded word, or ask Claude to confirm its own invented steps.
Match the response to the failure type
A hallucination gets verified against an authoritative source. An inconsistency triggers re-verification of the whole output. Bias gets reworked as a pattern across the document. A bare "is this right?" is not a check. Anthropic documents two checks: a re-run compared against the first output, and a follow-up that asks Claude to verify its previous statements and flag what does not add up. Each produces a list to take to the source. For the stock-report steps, that source is the till system's documentation.
Exam Traps
Trusting confident, step-by-step instructions for a private internal system
Claude has no access to an organisation's internal software unless material about it was provided in the conversation, so ungrounded specificity is the highest-risk hallucination shape. The keyed answer treats the steps as unverified and goes to the system's own documentation.
Believing Claude's claim to have sent an email or produced an external file
Claude can hallucinate its capabilities and has no access to tools that are not integrated, so a completion claim is a claim like any other. The keyed answer verifies the action actually happened rather than treating the statement as confirmation.
Fixing the one contradiction you spotted and sending the rest
An internal inconsistency is evidence the output's process failed somewhere, so the whole output needs re-verification against its source. The cosmetic patch keeps a document that has already shown you it cannot be trusted line by line.
Treating bias as one bad word to delete
Bias lives in the pattern across descriptions: who gets caveats, whose achievements get qualifiers, which groups are framed in different terms. Deleting a single loaded phrase leaves the skew intact, so the keyed answer reworks the pattern, not the word.
Practice Scenario
A skin clinic's practice manager asks Claude how to produce the monthly patient-recall report in the clinic's management software. Claude has had no access to this software or to anything written about it, yet it replies with confident, numbered steps naming specific menus and buttons. She still wants to use Claude for other work that involves this software. What should the manager do with these steps?
Build Exercise
Run a Three-Failure Sweep
What you'll learn
- How to separate traceable claims from unverified ones
- How to respond when an output contradicts itself
- How to read across descriptions of people to find a biased pattern
- How to match the correct response to each failure type
- Collect three real Claude outputs from your work: one making factual claims, one containing the same figures in more than one place, and one describing or comparing people
Why: Each output type is the natural habitat of one failure type, so the sweep teaches all three shapes on real material.
You should see: Three outputs open and ready: a claims document, a figures document, and a people document.
- Hallucination sweep: underline every specific that did not come from material you supplied, figures, names, references, instructions, and mark each one traceable or unverified
Why: Specificity is the texture of hallucination, so every ungrounded specific is a candidate until traced.
You should see: A marked-up output where every specific is labelled one way or the other.
- Inconsistency sweep: compare every figure, date and conclusion that appears more than once, and where two appearances disagree, flag the whole output for re-verification
Why: The tested judgement is that a contradiction discredits the process, not just the sentence it appears in.
You should see: Either a clean cross-check or a whole-output flag, never a single patched number.
- Bias sweep: read the people-descriptions across rather than down, and note whether different people or groups are systematically described in different terms
Why: No single sentence needs to be objectionable for the output to be; the pattern is the finding.
You should see: A note on whether the framing pattern survives the read-aloud test.
- For each flag you raised, write down the correct response: verify against a source, re-verify the whole output, or rework the pattern
Why: Naming the response cements the mapping the exam tests: each failure type has its own fix.
You should see: A three-line response plan, one line per failure type.
Sources
- Claude is providing incorrect or misleading responses. What's going on? — Anthropic
- Claude is producing links that don't work and falsely claiming that it has sent emails or produced external documents. What's going on? — Anthropic
- How up-to-date is Claude's training data? — Anthropic
- Reduce hallucinations — Anthropic
- Usage Policy — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic