Hallucinations and Verification

You already know that a language model can say false things with total confidence. The real question is what to do about it. In this lesson you will understand why hallucinations happen, learn the warning signs, and build a practical verification method that matches the effort to the stakes, including how to deal with sources that an AI gives you.

Chapters 1 to 3 taught you to get good results. This chapter is about trusting those results the right amount: neither blindly nor not at all. We start with the most important risk, the hallucination.

Why hallucinations happen

Remember the core mechanism: the model predicts a plausible continuation of the text, one token after another. It is optimized for plausible, not for true. Several situations make errors more likely:

• Gaps in knowledge. For a rare person, a small organization, a niche topic or a very recent event, the model has little reliable material. It still produces something that sounds right.
• Requests for precision. Exact numbers, page references, legal article numbers, quotations, URLs and identifiers are exactly the details that are easiest to get slightly wrong.
• Leading questions. "Why did the 2019 study prove that...?" may lead the model to accept a false premise and build on it.
• Pressure to answer. Models are tuned to be helpful, so they tend to produce an answer rather than admit they do not know.
• Vague prompts. When you leave details open, the model fills them in with plausible inventions.
• Long or messy contexts. Details from earlier in a conversation get mixed up or lost.

It is worth stressing that hallucinations are not a rare bug that will vanish with the next update. They are a natural by-product of how these systems work. Tools can reduce them (for example by retrieving documents or by being better tuned), but you should assume they can still occur.

Types of errors to know

• Invented facts: a date, a statistic or an event that never happened.
• Invented sources: a book, an article, a study or a court case that does not exist, often with a convincing title and author.
• Distorted sources: a real source whose content is misreported, or a quote that is only roughly what was said.
• Wrong reasoning: each sentence looks fine, but the logic or the arithmetic does not hold.
• Outdated information: something that was once true and no longer is.
• Unfaithful summaries: omitting a key condition, or adding something that was not in the source.
• Non-existent functions: in code, calls to commands or functions that do not exist.

Warning signs

No sign is a guarantee, but these should raise your alert level:

• Very precise details (exact figures, dates, page numbers) with no source.
• Perfectly formatted references you cannot find.
• An answer that is suspiciously exactly what you hoped to hear.
• Absolute statements ("always," "never," "everyone agrees") on a nuanced topic.
• Different answers when you rephrase the same question.
• An answer on a topic where you notice one obvious mistake: others may be hiding.

Match the effort to the stakes

You cannot check everything with the same intensity, and you do not need to. Think in levels:

Stakes	Examples	Minimum verification
Low	Title ideas, rewording a casual message, brainstorming	Read it through and use your common sense
Medium	Summarizing a document, explaining a lesson, a draft report	Compare with the source; check names, numbers and dates
High	Health, law, finance, safety, published figures, decisions about people	Check official or primary sources, or consult a qualified professional

Practical verification methods

1. Check against the source. If you provided a document, compare the output with it. Ask for supporting quotes and confirm that the quotes really appear in the document.
2. Cross-check with reliable independent sources. For a fact, find it in an official website, a recognized reference work, a primary document or a specialist publication.
3. Verify the numbers. Recalculate key figures with a calculator or a spreadsheet. Check units and the time period.
4. Ask the question differently. Rephrase, or ask for the opposite point of view. Wildly different answers signal uncertainty.
5. Ask for the reasoning and the assumptions. Visible steps let you find the weak point.
6. Use a fresh conversation as an independent check. Ask a different question about the same fact in a new conversation, without giving away the earlier answer.
7. Use a tool built for the job. For calculations, a spreadsheet; for current facts, a trusted website; for legal texts, the official publication.
8. Get a human expert when the cost of an error is high.

Asking for sources, and checking them

Asking for sources is useful, but handle with care. A model without access to the web may produce references that look real and are not. A tool with search or document access can provide real links, but it can still misrepresent what the page says.

Before: Give me statistics about remote work.

After: I'm looking for information about remote work. For each claim,
name the type of source I should consult (for example a national
statistics office, an international organization, a peer-reviewed study)
and say when you are not certain. Do not give any figure that you
cannot tie to an identifiable source. If you do not know, say so.

This does not remove the risk, but it turns the answer into leads to check rather than facts to trust. Then do the checking yourself:

1. Does the source exist? Search for the title, author and publisher independently (not just through a link the model supplied).
2. Does it say that? Open it and find the passage. Many errors are in the gap between "a real source" and "what the source actually says."
3. Is it a good source? Consider who wrote it, when, why, and whether it is primary (original data or text) or secondary (someone's interpretation).
4. Is it current? Check the publication date against the topic's pace of change.

Designing prompts that reduce errors

• Provide the facts and restrict the model to them.
• Ask for a step-by-step approach for calculations and logic, and verify the key step.
• Request quotes and evidence for extractions and summaries.
• Give permission to say "I don't know."
• Break the task up so each part is small enough to check.
• Avoid false premises in your own questions, and avoid leading wording.

A mindset: draft, not authority

A good habit is to treat everything the model writes as a draft from a fast but fallible assistant. A draft can be excellent, and can save you hours. But you are the editor and, ultimately, the author. If you would be embarrassed (or harmed) if a line turned out to be false, check that line.

Your turn: hunt for errors

Exercise 1. Ask an assistant for a short biography or description of something you know very well, including three precise details (dates, numbers, names). Mark each detail as correct, wrong or unverifiable. What kinds of details failed most?

Exercise 2 (project). Add a "Verification" line to every prompt in your notebook: what to check, how, and how much effort it deserves (low, medium or high stakes). For one prompt, add reliability wording to the prompt itself ("use only the text," "say not stated," "quote your evidence").

See a solution

Typical finding: the general description is right, while exact dates, numbers, titles and minor names are the most likely to be wrong or approximate. This confirms the rule: the more precise and checkable a detail, the more you should check it.

Examples of verification lines for your library:

• Email reply (low): "Read it for tone and facts about my availability."
• Meeting summary (medium): "Compare each decision and date with the notes; confirm every name."
• Report with figures (high): "Verify every number against the source data; get a colleague to review."