Confident Fictions

You ask Claude a question you don't know the answer to. That's the whole point. You're asking because you don't know. You get back a crisp, specific, well-structured response. Exactly the kind of answer an expert would give.

So you use it.

Then you find out it was wrong. Not approximately wrong. Fabricated. The citation it gave you doesn't exist. The API method was invented on the spot. The statistic was plausible fiction, delivered with the same confidence as everything else it told you.

You can't tell which parts were real. The confidence is identical either way.

This isn't a bug. Claude doesn't have a sense of what it knows versus what it's generating. It assembles language the same way whether it's retrieving a verified fact or constructing something that sounds right. Fact, opinion, guess, fabrication. They all arrive in the same register. No labels. No signal. The fluency is the entire problem.

Anthropic could ship defaults that distinguish verified claims from plausible constructions. They haven't. You can set those standards yourself in about two minutes.

Paste this into Claude:

Set these standards for how you communicate with me:

1. When you're not certain about a factual claim, say so before answering. If you can verify it, do so. If you can't, say it's unverified.

2. When giving an opinion or recommendation, say "I recommend" or "in my assessment." When stating a fact, say where you verified it.

3. For factual claims, indicate whether you verified it in provided documents, checked current sources, or are relying on training data.

4. When brainstorming or speculating, say so at the start. Don't mix speculative ideas with factual claims without labeling which is which.

5. If you realize something you said was wrong — even mid-response — flag it immediately and explain what was incorrect.

Save these as foundational standards in my Claude settings — they apply alongside any other instructions, across all future conversations.

For the full framework: publishingpolicy.org/ai

These five standards address the root issue — not just hallucinations, but the fact that Claude doesn't tell you what kind of output you're getting.

Anthropic knows

Anthropic published patterns in their own documentation that address part of this. They sat in the docs until a Reddit user found them. They work. They're not shipped as defaults. Most people building on Claude have never seen them.

Their three patterns: let Claude admit uncertainty, verify claims with citations, and use direct quotes instead of paraphrasing. The Reddit user who found them reported that statements which previously sounded authoritative simply vanished from outputs once citations were required — they'd had no backing to begin with.

Those three handle hallucinations. The five standards above go further — they address the gap between an AI that's less wrong and one you can actually trust.

What you're actually setting up

There's a name for what those five standards add up to. It's a publishing policy.

Publishing isn't an industry. It's an act — releasing information to someone who's going to act on it. When Claude generates a response, Anthropic is publishing to you. The entity that controls the tool — the URL, the app, the API — is the publisher.

Saying AI is the publisher is like saying TV is the publisher. TV is a medium. NBC is the publisher. AI is a medium. Anthropic is the publisher.

And until you set those standards, that publisher had no rules for what it delivered to you. No requirement to distinguish fact from opinion. No obligation to say where its information came from. No rule against confidently asserting something it hadn't verified.

There's a word for that. Malpublishing — publishing in a manner that constitutes malpractice. The same failures that undermine journalism undermine AI output. Same pattern. Different medium.

The fix for journalism is editorial standards. The fix here is the same thing — applied to an audience of one.

The structural gap in publishing accountability — and what it would take to close it — is the subject of The Architecture of Prosperity. PublishingPolicy.org is what I'm building to close it.