Why we force a fact-check warning on every health and finance outline (and you should too)

If you publish a blog post that says the wrong thing about a medication dosage, an interest rate, or a tax deadline, you can hurt someone. That's not a hypothetical. People read blog posts before they make decisions. The acronym Google uses for this category of content is YMYL — Your Money or Your Life. It covers anything where inaccurate information could meaningfully impact a reader's health, financial stability, or safety.

Most AI writing tools treat YMYL topics the same way they treat "best coffee shops in Brooklyn." They generate, the user copies, the user publishes, and the wrong number ends up in front of someone about to refinance a house. This post is about why we don't do that, what we do instead, and what you should be doing in your own editorial workflow even if you don't use our tool.

The problem with AI on YMYL topics

Language models don't know what's true. They know what's likely to come next in a sequence of tokens. On most topics, "likely to come next" and "true" overlap heavily enough that the model sounds authoritative without being wrong. On YMYL topics, the overlap narrows dramatically:

• Stale specifics. Tax brackets, contribution limits, interest rate ranges, FDA-approved indications — these change every year, and many models trained on older data will confidently produce last year's number.
• Confident over-precision. "A 30-year fixed mortgage is currently 6.42% APR" is the kind of sentence a model will produce. The right sentence is "rates vary by lender, credit profile, and day; check current quotes." Models are bad at the second kind.
• Plausible but wrong recommendations. "You can deduct mortgage interest on a primary residence up to $750,000 in mortgage debt" is roughly right for the US federal level, but state-level rules, marriage filing status, and AMT considerations can change the answer. A model rarely says "this depends."

The result is content that reads authoritative but is not safe to act on without a second source.

What OutlineAI does

Every outline our system generates is passed through a quality gate that classifies the topic into one of four risk categories: standard, health, financial, or legal. The classification happens in two places, and they have to agree.

1. The model's self-classification

Our system prompt asks the LLM to label its own output with a content_risk field. We don't trust this label. Models are confidently wrong about their own risk level on roughly 15-20% of YMYL topics in our testing. We use the model's label as a suggestion, never a final answer.

2. A deterministic keyword check

After the model returns an outline, we run a keyword check on the topic, the H1, and the primary keyword against a curated list. The list includes both English and common non-English terms — mortgage, roth, 401k, tax, homebuyer, credit score, health, medical, symptom, diagnosis, sleep, insomnia, pregnancy, law, legal, attorney, immigration, plus the same concepts in Spanish, French, German, Italian, Portuguese, Japanese, and Chinese. If either the model's self-classification or the keyword check returns a YMYL category, the outline is marked as YMYL.

We over-trigger YMYL on purpose. The cost of a false positive is a fact-check warning the user can ignore. The cost of a false negative is a reader acting on wrong information.

3. A non-dismissible warning

If the final risk category is health, financial, or legal, the outline page — and the Markdown export — both include a static warning at the top:

Financial topic: verify current limits, tax rules, eligibility requirements, and numerical claims against official guidance before publication.

Health topic: verify medical claims, outcome estimates, contraindications, and safety advice against authoritative clinical guidance before publication.

Legal topic: verify jurisdiction, current law, deadlines, and compliance requirements with an authoritative source before publication.

The warning is not a banner that can be X'd out. It is part of the document. It appears in the copied Markdown. It appears on the sample pages. It is part of the contract between us and the user: if you publish this without verifying, the warning was there.

What the warnings look like in practice

The full outlines are on the samples page. Two are YMYL:

• How to Build a Sustainable Sleep Schedule — health. Eight H2s, eight FAQs, all structured around the process of building a schedule, not specific medical claims. The warning sits above the outline and travels with the Markdown.
• First-Time Home Buyer's Checklist for 2026 — financial. Eight H2s including "Check and Improve Your Credit Score", "Save for Down Payment and Closing Costs", and "Buying Checklist: Questions to Ask Before You Order". Again, structured around the process, not specific 2026 numbers. The warning travels with the Markdown.

The reason both outlines can be generated at all without being dangerous is that we steer the model away from numerical specificity in the system prompt. The outline tells the writer what to research. The writer still has to do the research. The warning is the contract that says "you, the writer, have work to do here."

What you should do, even if you don't use our tool

If you publish YMYL content — health, finance, legal, anything where a wrong answer hurts someone — your editorial workflow should include:

1. A fact-check step before publication, by a human, against an authoritative source. Not a Google search. An actual primary source — a government page, a peer-reviewed paper, a statute, an official FAQ. AI search results are not authoritative.
2. A "last verified on [date]" line in the article or a visible "as of [date]" framing for any time-sensitive numbers. Rates change. Limits change. Rules change.
3. A specific-disclaimer block for medical, financial, or legal recommendations. Not a generic "this is not advice" — a specific one that names what the article does and does not cover.
4. An editor sign-off that is a different person than the writer, and who has flagged which claims were verified and how.
5. A revisit schedule. A 2024 health article probably needs a 2026 review. A 2024 tax article definitely does. Track when each YMYL article is due for re-verification.

These five steps are not optional for YMYL. They are the floor.

What we won't do

We won't silently generate YMYL outlines and let the user copy them with no warning. We won't add a generic "consult a professional" footer that nobody reads. We won't let the model's self-classification be the only safety net. We won't add a per-article opt-out that disables the warning.

If you find a YMYL outline in our output that should have a warning and doesn't, that is a bug. Email [email protected] with the topic and we'll trace it through the keyword list. The system is good but it is not perfect, and the failure mode we'd most regret is the one where the warning is missing.

You can see every YMYL outline we currently publish on the samples page. The two YMYL ones — sleep and home buying — both have the warning visible at the top, and the same warning is in the Markdown you copy. That is the standard.