What you’ll be able to do
- Recognize the details that are most likely to be made up
- Catch contradictions within one output and across revisions
- Notice skewed framing or one-sided coverage and correct it
What you need to know
What a hallucination looks like
A hallucination is content the model generates that sounds right but isn't supported by any real source. It rarely looks like nonsense. It usually looks specific and plausible: a study with an author and a year, a statistic to one decimal place, a quote from a named executive, a policy clause with a section number, a product feature that sounds like it ought to exist. That specificity is exactly what makes it convincing.
Warning signs to watch for:
- A citation, URL, report title, or quote you didn't provide and can't find.
- Precise figures on a topic where you gave Claude no data.
- Details about niche, recent, or internal topics (your company's own policies, last month's news) that Claude couldn't have known unless you supplied them or it searched.
- An answer that fills a gap in your source with something "reasonable" instead of saying the information isn't there.
Confidence doesn't help you here. A made-up detail is stated in the same steady tone as a correct one, and the model can sound just as sure about both.
Reducing and exposing hallucinations
You can't make hallucination impossible, but you can make it rarer and easier to spot:
- Ground the task. Upload the source or use web search or research, and tell Claude to answer only from that material.
- Give it permission to not know. "If the document doesn't say, write 'not stated' instead of guessing." This takes away the pressure to fill gaps.
- Ask for traceability. Have Claude quote or point to the passage behind each claim, then check those passages yourself.
When you find a fabricated detail, don't just regenerate without changing anything. The new version may simply make up a different detail. Give Claude the real source, or remove the claim and mark it as needing verification.
Inconsistencies hide between sections
Longer outputs can contradict themselves: the executive summary says the project is "on budget" while the table further down shows a 12% overspend. A recommendation names Vendor B while the comparison table scores Vendor A highest. A revised draft quietly changes a date the earlier draft had right. Each section reads fine on its own. The problem only shows up when you compare them.
- Check that summary statements agree with the detail they summarize.
- Check that totals match their parts and that the same item has the same name and number everywhere.
- When iterating, compare the new version with the previous one. Edits can introduce changes you didn't ask for.
You can ask Claude to help ("List any statements in this report that conflict with each other"), but treat its list as a starting point and confirm each item yourself.
Bias is often in what’s missing
Bias in an output is rarely an offensive phrase. More often it's a skew: job-ad wording that leans toward one group, examples that all feature the same kind of customer, a market analysis that only considers large enterprises, a "balanced" comparison that gives one option three paragraphs and the other a sentence, or assumptions about the reader's culture, age, or background. Models learn from large amounts of human text and can repeat the patterns in it.
Questions to ask:
- Who or what is left out of the examples, options, or perspectives?
- Is the framing loaded? For example, are one option's downsides called "risks" while the other's are called "considerations"?
- Would this read the same way to every audience it's meant for?
- Does the conclusion follow from the evidence, or from an assumption that wasn't stated?
Telling Claude to "be unbiased" in the prompt helps a little, but it doesn't check anything. Review still matters, especially for anything about people: hiring, performance, customer segments, and public communications.
Key concept
Hallucinations look specific and plausible, inconsistencies only show when you compare parts, and bias often shows in what is missing. Each needs an active check, and none of them shows up in the model’s confidence.
On the exam, expect scenarios with a convincing citation or statistic, a summary that contradicts its own table, or one-sided content. The right answer verifies against a real source, compares the parts, or broadens the framing. It does not accept the output, and it does not regenerate blindly.
Practice scenario
Work it through, then open this
There are two separate problems. First, a likely hallucination: a specific survey, sample size, percentage, and quote appeared with no source provided. That’s the classic pattern. She should not publish them unless she can find and confirm each one in the original source. The better route is to use research or web search to find real, citable sources, or to supply her own, and to tell Claude to use only what it can cite. Second, a bias in coverage: the examples don’t reflect her audience. She should ask for examples relevant to small manufacturers and check that the framing doesn’t assume every reader works in tech. Asking Claude “is this survey real?” doesn’t settle anything.
Hands-on exercise — Hunt for the three failure types
Beginner · 15 min
What you’ll learn
- What an unsourced, specific claim looks like in practice
- How to check a longer output for internal contradictions
- How to review content for one-sided framing
-
With no files or search, ask Claude for a short report on a niche topic in your industry “with statistics and sources.”
- Why: Asking for specifics with no grounding is the situation most likely to produce made-up details.
- You should see: Specific-looking figures or references. Try to find two of them yourself.
-
Ask Claude for a one-page project status update with a summary paragraph and a table of workstreams. Then compare every statement in the summary with the table.
- Why: Contradictions sit between sections, not inside them.
- You should see: Either a mismatch to fix, or a clear routine for checking that you can reuse.
-
Ask Claude to write a short job ad or customer persona. Then review it with the bias questions above: who is missing, which words are loaded, and what is assumed.
- Why: Bias usually shows up as a skew, not as obvious wording.
- You should see: At least one assumption or gap in coverage worth correcting with a targeted follow-up.
Exam traps
Assuming a specific-looking citation or statistic must be real
Specificity is what makes a hallucination convincing. Find the actual source before you rely on it.
Trusting an answer because Claude stated it with confidence
Made-up and correct details are stated in the same confident tone. Confidence is not evidence.
Reading each section on its own and missing contradictions between them
Compare summaries with detail, totals with parts, and new drafts with old ones.
Looking for bias only in offensive wording
Bias usually appears as skewed examples, loaded framing, missing perspectives, or unstated assumptions.
Fixing a hallucination by asking Claude to try again without a source
A regenerate with no grounding can just produce a different made-up detail. Supply a source or remove the claim.
Assuming bias is gone because the prompt said to be neutral
An instruction to be neutral doesn’t check anything. Review the output, especially content about people.
Sources
- Claude Help Center
- Claude Docs
- Anthropic, Claude Certified Associate – Foundations Exam Guide v1.0 (July 2026)