The Hidden Cost of Letting Generic AI Write Your Clinic's Content

Can you just point ChatGPT, or one of the new medical AI writers, at your clinic’s blog and let it publish? Technically yes. Safely, no. The tools are good at producing fluent, confident, medical-sounding text in seconds, and that is exactly the problem. Fluent and confident is not the same as true, sourced, legally defensible, or built to get found. For a cash-pay clinic in hormones, peptides, or weight loss, the gap between those two things is the gap between a content asset and a liability with your name on it. Here is what generic AI actually gets wrong, why most medical-specific AI tools fix the easy part and skip the hard part, and what publishing the right way looks like.

TL;DR

• Generic AI invents sources. In a peer-reviewed study, 47% of ChatGPT’s medical references were completely fabricated and another 46% were real citations pointing to the wrong thing, leaving only 7% both authentic and accurate.
• A confident, unsupported health claim is a regulatory problem. The FTC requires real scientific evidence behind a claim before it is made, and an AI does not know or care whether your evidence exists.
• Anonymous, unreviewed AI text fails the exact trust signals Google and AI answer engines now reward, so it does not rank and does not get cited.
• Most medical-specific AI writers fix tone and specialty vocabulary. They still do not ground claims in real sources, screen against FDA and FTC guidance, put an accountable physician behind the words, or keep content current as guidance changes.
• The answer is not to avoid AI. It is to wrap AI in the controls that make its output safe and effective: real-source grounding, claim screening, physician sign-off, verifiable provenance, and structure built for SEO and GEO. That full pipeline is what we built.

What generic AI actually gets wrong

The failure modes are not edge cases. They are built into how a general-purpose language model works, and each one lands directly on a clinic.

It invents its sources. This is the one that should stop every clinic owner cold. Researchers asked ChatGPT to generate medical content with references, then checked the references against the literature. They found that of the citations it produced, “47% were fabricated, 46% were authentic but inaccurate, and only 7% were authentic and accurate” [1]. Read that again. Ninety-three percent of the sources were either invented outright or real papers cited for claims they do not support. A model writes a citation because the text looks like it should have one, not because it read and verified a study. For medical content, where the citation is the whole point, that is disqualifying on its own.

It states claims it cannot back up. A language model is built to produce the most plausible next sentence, not the most defensible one. Ask it about a treatment and it will tell you it works, often in stronger terms than any clinician would put in writing. The marketing rules do not grade on fluency. The FTC requires “competent and reliable scientific evidence” behind a health claim before it is made [2], and its authority to act on claims that fall short sits in Section 5 of the FTC Act, which prohibits “unfair or deceptive acts or practices in or affecting commerce” [3]. An AI has no idea whether the evidence for a given claim exists, and it will write the claim anyway.

It reads as anonymous and unaccountable. Even when the facts happen to be right, raw AI output has no author who reviewed it and no expert who stands behind it. That is the precise pattern Google’s quality system is built to discount. Its guidance asks publishers whether content is “written or reviewed by an expert or enthusiast who demonstrably knows the topic well” and whether a reader can see “who authored your content” [4]. Unreviewed AI text answers no to both. As we covered in Google knows you used AI and does not care as long as you did it right, the problem was never the tool. It is shipping the tool’s output with no human expertise on top.

It leaves fingerprints, and it goes stale. Default AI prose carries tells that both readers and ranking systems have learned to spot, and it reads like marketing rather than medicine. Worse, it is frozen at the moment it was written. Medical and regulatory guidance moves, and a static article published once and forgotten drifts out of date with no mechanism to catch it.

”But we use a medical AI writer”

The honest pushback is that the specialized tools are better than raw ChatGPT, and they are, at the easy part. A medical-specific AI writer will fix the specialty vocabulary, match a clinical tone, and avoid the most obvious howlers. That is real, and it is also not the part that protects you.

The load-bearing problems survive the upgrade. Most medical AI writers still generate citations from the model rather than grounding every claim in a verified source, so the fabrication problem persists in a more convincing wrapper. Most do not screen the finished draft against current FTC and FDA guidance. Most have no physician of record who reviews and signs the work, which is the actual basis of trust and the actual location of liability. Most produce no verifiable record of who wrote what, when, and on what evidence. And most publish once and move on, with no system to re-check the piece when guidance shifts. A better-sounding draft with the same five holes is not a safer draft. It is a more persuasive one, which is arguably worse.

The three approaches, side by side

	Generic AI	Most medical AI writers	Authoritize
Citations	Often fabricated	Model-generated, usually unverified	Grounded in real PubMed sources, verified
Marketing-claim safety	None	Rare or generic	Screened against FDA and FTC guidance, every piece
Accountable author	None	None	The clinic’s own physician reviews and signs
Provenance	None	None	Hash-chained Attestation Badge per article
Built for SEO and GEO	Incidental	Partial	Structure and schema baked in by default
Stays current	No	No	Monthly re-audit as guidance changes

The pattern is the same down every row. The cheaper tools handle the part that is easy to see and skip the parts that are easy to ignore right up until a regulator, a skeptical patient, or a ranking drop makes them impossible to ignore.

What publishing the right way looks like

We did not build a better AI writer. We built the pipeline of controls that has to sit around any AI writer before its output is safe to put a clinic’s name on. Each control answers one of the failures above directly.

• Real-source grounding. Every article is built on retrieved, real medical literature, not on whatever a model recalls. Claims trace to sources that actually exist and actually say what we cite them for. This is the direct answer to the 93% citation-failure problem.
• Marketing-claim screening. Before anything publishes, it is screened against FDA and FTC guidance using the same claim scanner our public audit runs. Flagged language is rewritten or removed, not shipped and hoped over.
• A de-AI and AI-citation pass. Two automated gates strip the fingerprints that get content demoted and confirm the structure that gets content cited, the question-and-answer openings, schema markup, and clean sourcing that AI answer engines reward.
• A physician of record. The clinic’s own credentialed physician reviews and signs anything touching a medical decision. That is the trust signal Google and the AI engines are built to reward, and it keeps clinical liability with the clinician, where it belongs. We screen the marketing language. The doctor owns the medicine.
• Verifiable provenance. Each article carries a hash-chained Attestation Badge, a tamper-evident record of what was published, when, and who stood behind it. If a claim is ever questioned, the answer is not a shrug. It is a record.
• Ongoing re-audit. A monthly process re-checks published content against current guidance, so a piece that was clean in June does not quietly become a problem in December.

No single one of these is exotic. The point is that they only protect a clinic when they all run, in order, on every piece. A general AI tool gives you the draft and none of the guardrails. We built the guardrails first, and the draft is the easy part.

If you want to see how your current content holds up on the signals that decide whether you rank, get cited, and stay clean, run a free audit. It scores your live site in a few minutes, no sales call attached.

FAQ

Is it actually unsafe to use ChatGPT for clinic content? Used as a private brainstorming tool, no. Published straight to a clinic’s site, yes. The documented problems are fabricated citations, unsupported claims, and no accountable reviewer. A study of ChatGPT-generated medical references found only 7% were both authentic and accurate [1]. None of that is safe to publish to patients without verification, claim screening, and a physician’s review on top.

Aren’t medical-specific AI writers good enough? They fix tone and vocabulary, which is the visible part. They generally do not verify citations against real sources, screen claims against FTC and FDA guidance, attach an accountable physician, or keep content current. A more convincing draft with the same underlying risks is not safer, and can be more dangerous because the errors are harder to spot.

So is Authoritize against using AI? No. AI does the drafting in our pipeline too. The difference is everything wrapped around it: real-source grounding, claim screening, a de-AI and citation-structure pass, a physician of record, an Attestation Badge, and monthly re-audit. AI without those controls is the liability. AI with them is a genuine advantage.

What makes this the safe and effective way rather than just one option? Each control closes a specific failure. Grounding closes fabricated citations, screening closes claim risk, physician sign-off closes the trust and liability gap, attestation closes provenance, and the structure pass closes the SEO and GEO gap. We are not aware of a general tool that runs all of them, in one pipeline, on every article. That combination is the standard high-quality medical content has to meet, and meeting all of it at once is what we built.

Will this hurt our search rankings the way AI content is rumored to? The opposite. Google has said it does not penalize AI assistance as such, only low-quality, unhelpful, unaccountable content. Content that is sourced, reviewed by a named physician, and structured for AI citation is exactly what its guidance rewards [4]. The risk is not using AI. The risk is publishing AI output raw.

Citations

1. Bhattacharyya M, Miller VM, Bhattacharyya D, Miller LE. “High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content.” Cureus. 2023;15(5):e39238. https://pmc.ncbi.nlm.nih.gov/articles/PMC10277170/
2. Federal Trade Commission. “Health Products Compliance Guidance.” 2022. https://www.ftc.gov/business-guidance/resources/health-products-compliance-guidance
3. Federal Trade Commission Act, Section 5, 15 U.S.C. § 45. https://www.ftc.gov/legal-library/browse/statutes/federal-trade-commission-act
4. Google Search Central. “Creating helpful, reliable, people-first content.” 2024. https://developers.google.com/search/docs/fundamentals/creating-helpful-content