Professor Raj Rattan, MPS global adviser, recently argued in a previous MedicalBrief article that we should stop treating human oversight as the immutable gold standard of safety. He said we need to be honest about human judgment, its strengths and its limitations. That means building systems that can evolve as work shifts between people and machines. He was writing about AI in the consulting room, but we’ve been noticing the same thing play out in the complaints process, write Dr Yash Naidoo, dento-legal consultant at MPS,and Dr Volker Hitzeroth, medico-legal consultant at MPS.
Naidoo and Hitzeroth write:
A while ago, a complaint landed against one of our members. Under the section of the HPCSA form headed Details of complaint, which the patient completes, sat a tidy, lettered analysis: A. Violation of “Do no harm” (non-maleficence). B. Failure of beneficence.
It cited ethical principles and offered a quotable line for emphasis: ethical harm is measured by impact, not intent. And right at the top, before any of it, three words the complainant failed to remove: “Your account describes…”
That’s the tell. Not the bioethics, but the pronoun. The AI was doing the talking, narrating the patient’s story back to them, and the patient had pasted the whole lot into the complaint form.
This case wasn’t a one-off. In another, the patient left the AI coaching in: “You did well including facts… you are building a very structured case… if you would like, I can now…” followed by a menu of next steps, including an offer to draft a parallel claim against the practitioner. In a third, the complaint arrived with an analysis of the lab results and a summary of missed diagnostic opportunities.
The AI had stopped reformatting a grievance and started writing the medicine. Used this way, AI doesn’t only pad thin complaints, it sharpens genuine ones.
In our experience, a complaint written by a patient usually reads one of two ways: disorganised and emotive, or heavy on chronology and still emotive. An AI-drafted complaint is different: calm, lettered, reaching for principles. You might wonder whether the patient had simply found themselves a good attorney, except that what gives these examples away is something no attorney would ever leave in.
The careless complainants leave the scaffolding showing. The careful ones will strip it out, and then the complaint just reads as unusually measured, and we are none the wiser.
For the average clinician, the tells are worth knowing:
Spotting an AI-drafted complaint: the common tells
Word choices. An over-reliance on verbs like delve, foster, elevate and leverage; buzzwords such as dynamic, comprehensive, landscape and testament; and stock openers like “In today’s fast-paced world” or “It is important to note”.
Structure and punctuation. Heavy use of the em dash to stitch thoughts together; a predictable rhythm of bold heading, colon, bullet list and a tidy summarising sentence; and sentences of near-uniform length, without the short jabs and long loops of natural writing.
Emotional flatline. Diplomatic hedging (“some may argue”, “on the other hand”) that never commits to a view; and flowery language around serious matters that never lands on a specific, lived detail.
Here is the part that should give us pause before we feel too smug about it. Is the AI even wrong? It was recently reported that South African researchers had benchmarked ward diagnoses at a large, well-known public hospital against a range of AI systems, using hundreds of in-patient files, and that the AI models consistently outperformed the hospital teams, at a tiny fraction of the cost of the expert doctors they were measured against.
The lead author was reported as describing this as the beginning of an era of cheap, reliable AI diagnosis. We haven’t seen the study ourselves and it carries the usual caveats. But the direction of travel is hard to ignore: when a complaint alleges a missed diagnosis, drafted with the help of a tool that may well out-diagnose us, it is hard to discount it. Sometimes the bot will be right.
What do AI powered complaints like this look like at scale?
Reana Steyn, who heads the National Financial Ombud Scheme, says submissions are ballooning to 150 pages, citing legislation and case law, some of which, on review, simply does not exist. The damage isn’t that complainants are getting help; it’s that the “help” buries the point.
In an interview with The Money Show, when asked how many of these AI-assisted complaints actually improved the complainant’s odds, her answer was: none, as far as she was aware. It lengthens the matter without changing the outcome, because someone must still trawl through the endless pages and dig out the key issues.
Steyn also points out that the bloat doesn’t only come from complainants. Some institutions answering them lean on AI too. One response her office received tested as 97% machine-generated. Translate that into the consultation room and the party drafting the response is the practitioner.
The temptation is obvious: a long, aggressive complaint lands, and the fastest way to deal with it is to feed the file to a chatbot and let it write back. Except that the moment you paste a patient’s confidential information into a public AI tool, you have a problem that has nothing to do with the quality of the reply.
Accountability doesn’t evaporate when the machine slips, it lands on the nearest human.
As Professor Martin Brand, an MPS medical adviser, noted in a recent article for Specialist Forum, liability for AI in healthcare stays murky right up until something goes wrong, at which point it concentrates on the clinician.
So, what’s the answer? Steyn offers a possible model almost in passing: her office once asked a complainant to summarise a sprawling submission. The complainant used AI to do exactly that, and it came back short, clear and workable for all. Same tool, different instruction.
Professor Rattan’s point was that we occasionally overrate the human factor and that we lean on clinician oversight as though it were infallible, when it never was. The complaints file proves him right from an unexpected direction. The patient who trusted the machine and never read it back was the weak link in their own case. The clinician who answers a complaint by pasting it into a chatbot is a keystroke away from being the weak link in their own defence.
As the technology advances, the tools in the clinical toolbox will change, but good patient care will still demand what it always has: clinicians who are kind and engaged, honest and reflective, open to new advances and willing to fold them into practice responsibly.
That is why we became dentists and doctors in the first place. A new technology has arrived, and how we adapt to it will reflect on us as much as on our patients. The tool isn’t the danger here. We are, unless we stay the kind of human in whom Professor Rattan still has faith: awake, accountable, and willing to read the thing back before it goes anywhere at all.
Last year we launched the AI Safer Practice Framework, aimed to help healthcare professionals integrate AI safely and responsibly into practice. The framework is made up of two parts: INFORMED and RECORDS.
INFORMED guides ethical decision-making using AI, while RECORDS documents AI-assisted decisions for accountability and clinical rationale. The framework has been structured around these acronyms for ease of use.
The AI Safer Practice Framework can be accessed here.
See more from MedicalBrief archives:
Ethical dilemmas as medicine intersects with AI chatbots
Growing role for AI in everyday medical interactions
SA doctors beaten by AI in hospital study
AI chatbots outstrip doctors in diagnoses – US randomised study