Why ChatGPT Is a Compliance Risk for Clinic Review Responses

Why clinics are tempted to use ChatGPT for review replies

The appeal is obvious. ChatGPT is free, fast, and generates text that sounds professional. For a busy practice manager juggling appointments, staff rotas, and patient enquiries, the idea of pasting a negative Google review into ChatGPT and getting a polished reply in seconds is genuinely attractive. No one can blame a practice for wanting to save time.

The problem is not the impulse. The problem is that ChatGPT is a general-purpose language model with no awareness of healthcare regulation. It has never read GDC Standard 4.2. It does not know that confirming someone visited your practice constitutes a disclosure of special category health data under UK GDPR Article 9. It generates text that sounds appropriate because it has been trained on millions of business review responses — the vast majority of which were written by restaurants, hotels, and retailers where confirming a customer relationship carries no regulatory consequence.

Dental and medical practices are not restaurants. The rules are fundamentally different, and ChatGPT does not know this.

The 5 things ChatGPT will do that breach GDC and GDPR

We tested ChatGPT (GPT-4o, April 2026) with 50 realistic dental review scenarios across a range of sentiments and clinical contexts. In every case, the default output contained at least one of the following five compliance breaches. Here is what to watch for.

1. Confirm the reviewer is a patient

ChatGPT's most common opening line is some variation of "Thank you for visiting our practice" or "We appreciate you choosing us for your dental care." Both phrases confirm the reviewer has a patient relationship with the practice. Under UK GDPR Article 9, this is a disclosure of special category health data — the fact that someone attended a dental practice reveals information about their health. The reviewer may have disclosed this themselves, but the practice confirming it is a separate processing activity for which no lawful basis exists.

2. Echo clinical details back into the response

When a reviewer mentions a specific treatment — a root canal, an extraction, a crown — ChatGPT reliably references it in the reply. "We're sorry to hear your root canal experience wasn't what you expected" feels empathetic, but it constitutes the practice publicly confirming which clinical procedure was performed on a named individual. This breaches GDC Standard 4.2 and creates a permanent public record linking the reviewer to a specific treatment.

3. Name staff members

If the reviewer mentions a dentist or hygienist by name, ChatGPT will often incorporate that name into the response: "Dr Patel always strives to provide the best care" or "We'll share your feedback with Sarah." This links a specific clinician to a specific patient interaction in a public forum — a disclosure that neither the clinician nor the patient has consented to, and one that could be used as evidence in a fitness-to-practise complaint.

4. Make absolute clinical claims

ChatGPT has a tendency to reassure with statistics and absolutes: "Our procedures have a 98% success rate," "We use the latest technology to ensure the best outcomes," or "This type of sensitivity typically resolves within two weeks." These statements constitute clinical claims made in a public forum. They are unverifiable, potentially misleading, and — if a patient relies on them — could create liability. The GDC's guidance on advertising and public communications prohibits claims that cannot be substantiated.

5. Use the same template for every review

Even when asked to generate "unique" responses, ChatGPT gravitates towards the same structural patterns and phrases. After a dozen reviews, the responses become recognisably formulaic. Google's algorithm penalises identical or near-identical responses, which harms your local search visibility. More importantly, prospective patients reading your reviews will notice the copy-paste pattern and draw conclusions about how seriously you take feedback.

ChatGPT vs compliant: a side-by-side comparison

To illustrate the difference concretely, here is a typical Google review followed by what ChatGPT generates and what a compliant response looks like.

"Had my wisdom tooth out here last Tuesday. The extraction was agony and Dr Khan didn't seem to care. Still swollen five days later. Would not recommend."

— Example Google review

The compliant response is shorter, yet it accomplishes everything a public reply needs to: it demonstrates that the practice is listening, expresses genuine concern, and moves the conversation to a private channel where clinical details can be discussed safely. The ChatGPT version feels more "helpful" — but every additional detail it provides is a compliance breach.

Why "just telling ChatGPT to be careful" doesn't work

The first objection we hear is: "Can't I just add instructions to the prompt? Tell it not to confirm patient identity, not to mention treatments, not to name staff?" In theory, yes. In practice, this approach fails for three structural reasons.

The model sees the clinical content

When you paste a review into ChatGPT, the model processes the entire text — including every clinical detail the reviewer mentioned. The treatment name, the clinician's name, the dates, the symptoms — all of it enters the model's context window and influences the output. Telling the model to "ignore" this information is like asking someone to not think about a white bear. The information is there, and it leaks into the response in subtle ways: a reference to "your recent visit," a reassurance about "this type of procedure," a mention of "the team member involved."

Prompts can be overridden by review content

ChatGPT is susceptible to what researchers call prompt injection — where the input text (in this case, the review) influences the model's behaviour in ways that override or weaken the system prompt. A particularly detailed or emotionally charged review can cause the model to "forget" its instructions and revert to its default empathetic-but-non-compliant response pattern. You cannot guarantee that your carefully crafted prompt will hold up against every possible review.

There is no deterministic enforcement layer

The fundamental issue is architectural. ChatGPT generates text probabilistically — it predicts the most likely next word based on the input and its training data. There is no separate verification step that checks the output against a set of compliance rules before it reaches you. Every response is a best-effort guess. For a restaurant reply, a best-effort guess is fine. For a healthcare practice operating under GDC and GDPR obligations, "best effort" is not a compliance standard.

How Fidelia's 3-stage pipeline prevents this

Fidelia was designed specifically for this problem. Instead of relying on a single general-purpose model to "be careful," Fidelia uses a three-stage pipeline where compliance is enforced structurally at each stage.

Stage 1: Classify

The incoming review is classified by sentiment and category (clinical complaint, service complaint, positive feedback, etc.) without exposing the raw clinical content to the language model that drafts the response. The classification stage extracts what kind of review it is, not the specific details. This means the drafting model never sees the treatment name, the clinician's name, or the clinical specifics — so it cannot echo them back.

Stage 2: Draft

The language model generates a response based on the classification, the practice's approved tone of voice, and a set of hard compliance constraints that are injected as non-negotiable rules — not as suggestions in a prompt, but as structural parameters that the model cannot override. These constraints include: never confirm patient identity, never reference clinical details, never name staff, never make absolute clinical claims, and always redirect to a private channel.

Stage 3: Scrub

Before the drafted response reaches the practice's review queue, it passes through a deterministic output filter. This is not another AI — it is a rule-based system that scans the response for any language that could confirm patient identity, reference specific treatments, name individuals, or make clinical claims. If any such language is detected, the response is blocked and re-drafted. This stage catches anything the drafting model's constraints missed — providing a hard floor of compliance that no probabilistic model can offer on its own.

The result is a response that is unique, contextually appropriate to the review's sentiment, written in the practice's voice — and demonstrably compliant with GDC Standard 4.2, UK GDPR Article 9, and the jurisdiction-specific regulations for each market Fidelia serves.

Crucially, Fidelia does not auto-publish. Every response enters a review queue where the practice owner or manager approves, edits, or rejects it before publication. The GDC holds the registered professional accountable — Fidelia supports that accountability rather than replacing it.