Why Claude gets things wrong, and why it sounds so sure

Hallucination is the word the industry settled on and it is a poor one. A Claude AI hallucination is not the tool malfunctioning or slipping a gear. It is the system doing precisely what it was built to do, and producing something that happens not to be true. The word implies an aberration, a rare episode, a fault someone will eventually patch. What you are looking at is a side effect of the ordinary process.

Confidence tells you nothing

This is the uncomfortable part. A confident wrong answer is produced by exactly the same machinery as a confident right one. There is no internal alarm, no shift in tone, no careful hedging that appears when the ground gets thin. The prose arrives equally assured either way.

So the signal most of us have relied on all our working lives, how sure someone sounds, carries no information here at all. Senior people are unusually exposed to this. They have spent decades reading confidence as competence across a meeting table, and that reflex does not switch off in front of a screen. I have watched a chief executive accept a fabricated figure in half a second because the sentence around it was well built.

The three places it happens most

Three kinds of question account for nearly everything I see go wrong.

• Specific facts it was never given. Ask what the termination clause in your supplier agreement says, without supplying the agreement, and you will get something that reads exactly like a termination clause.
• Recent events. Training stopped at a point in the past. A question about last week's announcement gets answered from the shape such announcements usually take.
• Anything precise. Figures, dates, page references, case names. Precision is easy to produce plausibly and hard to produce accurately, and it is the thing people quote onward without checking.

Two of those overlap with the limits no release will remove. A few jobs are better kept away from it altogether.

What helps

Anthropic's own documentation tackles this directly, and it is franker than most vendor material. Four things in it change the odds, and none of them needs any technical skill.

Give it explicit permission to say it does not know. Anthropic calls this a simple technique that can drastically reduce false information, and in my experience it is the highest-return sentence you can add to a prompt. Type it plainly: if you are not sure, say so and stop there.

With a long document, ask for the relevant passages word for word before you ask for anything else. Anthropic attaches that advice to anything over about twenty thousand tokens, tokens being the units of text it counts in, which in practice means a tender pack and not a two-page note. Quoting first anchors the work to what is on the page rather than to what a page like that usually contains.

Then ask it to cite a source for each claim, and afterwards to go back over its own answer and withdraw anything it cannot support. It will withdraw things. That is the point of asking.

The fourth is the bluntest. Tell it to use only the documents you have supplied and to ignore its general knowledge. That turns a silent invention into a visible gap, which is the trade you want on anything leaving the building.

Anthropic's own caveat sits at the bottom of that page, and I read it aloud in sessions:

while these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions.

Anthropic, Reduce hallucinations

Where this leaves you

None of this makes Claude unusable. It makes it a tool with a known failure mode, which is a perfectly normal thing for a tool to be.

The working standard is duller than any of that. You would not sign a contract because a junior colleague told you the clause was fine. You would read the clause. The same applies here, and it costs about a minute.