daBongo LMS AI Training Courses

Lesson Content

What hallucination looks like in practice.

Hallucination – Claude producing confident-sounding but incorrect content – does not always look like an obvious mistake. It often looks like:

• A plausible statistic with a specific-sounding but unverifiable source
• A named case study that "confirms" a point but cannot be found online
• A legislative citation that is formatted correctly but does not exist
• A biographical detail about a real person that is wrong
• A technical specification that seems reasonable but differs from official documentation
• A history of a company or event that mixes real and fabricated elements

The common pattern: the content is plausible, specific, and confident. The danger is that plausible + specific + confident = believable to a reader who does not check.

High-risk output categories (review these before use):

• Specific statistics with attributed sources
• Legal case citations and regulatory references
• Medical dosages, treatment protocols, drug interactions
• Software API specifications and function signatures
• Named research studies and academic citations
• Biographical facts about real, named individuals
• Recent events (anything near or after the knowledge cutoff)
• Company financial data and market share figures

Asking Claude to flag its own uncertainty.

You can instruct Claude to signal its uncertainty explicitly: "Where you are not confident in the accuracy of a specific claim, mark it with [VERIFY] so I know to check it before using this professionally."

This does not solve the hallucination problem – Claude does not always know when it is wrong – but it surfaces the areas where Claude itself is uncertain, which is a useful starting point for your verification checklist.

You can also ask after receiving a response: "Which specific claims in what you just wrote are you least confident about?" Claude's answer reveals the shakiest areas of the output and guides your verification effort.

Building a tiered verification habit.

Not every Claude output needs the same level of verification. A tiered approach is practical:

Tier 1 – Use as-is with light review: Brainstorming outputs, draft email tone checks, outline suggestions, idea generation, creative exploration. Low factual stakes.

Tier 2 – Review carefully before use: Internal documents, analyses, summaries, recommendations. Check for logical coherence, internal consistency, and plausibility. Spot-check key claims.

Tier 3 – Verify key claims from primary sources before distribution: External communications, reports with cited data, proposals with factual claims. Every specific statistic, citation, and named source should be verified.

Tier 4 – Require professional review before use: Legal documents, medical content, financial disclosures, compliance materials, safety procedures. Claude's output is a starting point only.

The verification workflow.

For Tier 3 content, use this workflow:

1. Ask Claude to generate a list of the specific claims in its output that should be verified
2. Prioritize by consequence: what claim, if wrong, would cause the most damage?
3. Verify the highest-priority claims from primary sources
4. Update or remove any claim that cannot be verified