Blog · 2026-10-15 · baryreview editors

AI-written legal, medical and financial content: the extra checks

Why these topics get a different review

Most AI-written blog posts can be wrong without hurting anyone. A furnace post with a made-up statistic embarrasses the company. A post about tax deadlines, drug interactions or tenant rights can cost the reader money, health or a legal position. An AI written legal blog fact check, or its medical or financial equivalent, therefore starts from a different assumption: every specific claim is presumed wrong until a primary source says otherwise.

Google treats these topics differently too. Its guidance on creating helpful, reliable, people-first content says its systems “give even more weight to content that aligns with strong E-E-A-T for topics that could significantly impact the health, financial stability, or safety of people, or the welfare or well-being of society,” and names them: “We call these ‘Your Money or Your Life’ topics, or YMYL for short.” The same page says that for YMYL topics “the content must be highly accurate and consistent with established expert consensus,” and that of the E-E-A-T qualities, “trust is most important.” That is the search-engine reason. The reader-harm reason is bigger, and it is the one that should drive the review.

The failure mode is the same as on any AI draft: the model produces fluent, specific, confident text, and nobody checks it. What changes in these fields is what the specifics are. Below are the patterns we see most often in each, and then the protocol that catches them.

Legal: invented statutes, invented cases, real-sounding citations

The legal profession has produced the best-documented record of AI fabrication anywhere, because courts write down what happened. A public database maintained by Damien Charlotin tracks court decisions worldwide involving AI-fabricated citations in filings; as of its 10 October 2026 update it listed 2,184 cases. Scientific American has covered why trained lawyers keep doing it: the citations look exactly like real ones, with plausible party names, reporters, volumes and page numbers.

A blog post is not a court filing, but the same model writes both. The patterns in AI legal content:

• Cases that do not exist, cited in standard form. A citation like “Hartwell v. Meridian Logistics, 2021” reads as real. The only way to know is to look it up.
• Real cases that say something else. The model remembers that a case is about employment and invents the holding.
• Statutes with the wrong number, the wrong section, or the wrong jurisdiction. State law varies; a draft that gives one rule as “the law” has usually picked one state, or blended several.
• Deadlines and thresholds stated as fixed. Statutes of limitation, filing windows, small-claims limits and notice periods differ by state and change by legislation.
• “Generally” doing a lot of work. “Employers are generally required to…” is often a hedge over a claim the model is unsure of. The reader hears a rule.

The business consequences are not theoretical. In 2024 a Canadian tribunal held Air Canada liable for a refund policy its customer-service chatbot invented, rejecting the argument that the chatbot was a separate entity (American Bar Association summary). In 2025, Deloitte Australia reissued a government report after it was found to contain fabricated citations and a fabricated court quotation, and refunded part of the fee (The Conversation). Both are cases of an organization publishing what a model produced, under its own name, without a check.

Medical: dosage and guideline drift

Medical drafts fail more quietly. The model rarely invents a drug. What it does is drift: a number or a recommendation that was current at some point, or was true for a different population, or was a rounded version of a range, gets stated as the rule.

• Dosages. A draft will give a maximum daily dose, a frequency or a pediatric figure with total confidence. It may be the adult figure applied to children, an older guideline, or a figure for a different formulation. Any dose in an AI draft needs the label or the current prescribing information beside it.
• Guidelines that have moved. Screening ages, blood-pressure thresholds, vaccination schedules and dietary recommendations change. The model’s version is from whenever its training data ended, and it does not know that.
• Studies that are real but misdescribed. “A 2019 study found that…” often refers to a study that exists and found something narrower, in mice, or in a population the sentence does not mention.
• Symptom lists that mix conditions. Fluent, plausible and wrong in a way only a clinician would catch.
• Missing “see a doctor” lines, or lines placed where they read as decoration. For a page a worried person will find at 2 a.m., the escalation advice is the most important sentence on it.

An editor can catch the structural signs: a number with no source, a guideline with no date, a study with no citation. An editor cannot tell you whether the dose is right. That requires a clinician, and the protocol below says so.

Financial: thresholds that expire

Financial content ages fastest. Contribution limits, standard deductions, tax brackets, filing deadlines, interest-rate assumptions and benefit thresholds are revised every year, sometimes more often. An AI draft will state last year’s figure, or the year before’s, in the present tense. It will also:

• Blend jurisdictions, giving a federal rule with a state’s number in it, or a US figure in a post for a UK audience.
• Round into error. A threshold that is a precise dollar amount becomes “about $20,000,” which is wrong in both directions for the people near it.
• Present a strategy as universal advice. “You should max out your retirement contributions before paying down debt” is a decision that depends on the reader’s rate, income and goals, and reads as a recommendation from whoever publishes it.
• Invent performance figures. Average returns, fund fees and historical rates arrive with decimal places and no source.
• Skip the disclaimer or write a generic one that does not match what the page actually says.

The fix for financial drift is mechanical: every figure gets a tax year or a date next to it, and a link to the agency or institution that sets it. If the draft cannot say which year a number belongs to, the number comes out.

The review protocol

This is what the review looks like on a YMYL draft. It is stricter than a standard review, and slower.

1. Primary sources only. For legal claims, the statute, the regulation or the court opinion itself, not a law firm’s blog post about it. For medical claims, the prescribing information, the guideline body, or the study. For financial figures, the agency that sets the number. A secondary source is a lead, not a verification.
2. Dates on everything. Every figure, threshold, guideline and rule gets a date or a year beside it in the text, so the reader can judge whether it is still current, and so the next person to update the page knows what to check. “As of the 2026 tax year,” “under the guideline published in 2024,” “in Illinois, as of this writing.”
3. Jurisdiction on everything legal and financial. Which country, which state, which court. A rule with no jurisdiction is flagged as a rule with no jurisdiction.
4. Every citation looked up. Not “does this case exist” but “does it say what the sentence says it says.” Same for studies.
5. Strip advice-shaped sentences or label them. “You should” becomes “many people in this situation consider,” or gets an explicit disclaimer, or comes out. For a business that is not licensed to advise, this is the line between content and liability.
6. Disclaimers that match the page. A generic “this is not legal advice” footer on a page that walks the reader through a legal decision is not a disclaimer, it is a contradiction. The disclaimer should say what the page is, who wrote it, and what the reader should do before acting.
7. Author and reviewer named. Google’s guidance asks whether it is “self-evident to your visitors who authored your content,” and for these topics that includes who checked it. A byline, a reviewer line, and a date.
8. Professional sign-off before publishing. A lawyer for legal content, a clinician for medical, a licensed adviser or accountant for financial. The editor’s job is to get the draft into a state where that professional’s time is spent on judgment, not on catching made-up cases.

What we do and what we do not do

We should be plain about where our review ends on these topics.

What we do. A professional editor reads the whole page and flags every claim that looks wrong, outdated, unsourced or invented, every number without a date, every rule without a jurisdiction, every citation we could not find or that does not say what the draft says. We flag advice-shaped sentences and missing or mismatched disclaimers. We check the mechanics and the structure as on any article. Each flag comes back as a numbered correction, quoting the passage, usually asking you for a source or telling the AI to remove the claim. The corrections go to you and to your AI; the decision about what to keep is yours and, for these topics, your professional’s.

What we do not do. We are not lawyers, doctors or financial advisers, and we do not pretend to be. We do not confirm that a dose is correct, that a legal rule applies to your reader, or that a strategy is sound. We do not research primary sources on your behalf beyond confirming whether a cited source exists and says what the draft claims; if a claim needs a source, we ask you for it. We do not add disclaimers for you, and we do not sign off on a page as compliant. A review from us is the step before professional sign-off, not a substitute for it. If you are publishing in these fields without a professional in the loop, the most useful thing our editor may tell you is that you need one.

The practical version

If your business publishes AI-assisted content on legal, medical or financial topics, treat the draft as a list of claims to verify rather than an article to polish, put a date and a source beside every number, and get a licensed professional to read it before it goes live. A human review is the fastest way to find out how long that list of claims really is.