A couple of days ago I wrote about how AI meeting scribes might be gamed in organisational settings, where people could exploit systematic vulnerabilities in how LLMs process language. In this post I want to explore how clinical AI scribes could be exploited because of the same technical limitations, but where the power dynamics are different and potentially more consequential.
Physicians have historically controlled the medical record, shaping diagnosis, treatment pathways, and institutional memory. The person who writes the notes determines what “counts” as the patient’s problem; a dynamic with consequences for patients who are systematically dismissed or misunderstood.
Clinical AI scribes might exhibit the same predictable vulnerabilities we see in meeting summarisers: over-weighting content at the start and end of consultations, struggling to distinguish emphasis from ordinary speech, responding to formulaic phrases like “key concern.” Patients who understand these patterns could strategically position their symptoms, use repetition at transition points, or employ high-signal language that ensures their perspective appears prominently in the official record.
This redistribution of narrative power has contradictory implications. For patients who’ve been historically marginalised—women with chronic pain, or people whose symptoms don’t fit clean diagnostic categories—this might provide a mechanism to push back against medical gaslighting. Their lived experience could gain weight in systems that have traditionally privileged clinical interpretation over patient testimony.
But the same mechanisms enable symptom fabrication or strategic framing to obtain particular diagnoses or treatments. The medical record stops being “objective” documentation and becomes another site of strategic positioning, potentially eroding the trust that’s necessary for effective care.
The tensions clinical AI scribes introduce
The immediate challenge is that clinical AI scribes transform the consultation from a space where physicians control documentation into one where narrative authority is negotiated in real-time through language choices that both parties may not fully understand they’re making. This creates several unresolved tensions.
First, there’s the question of whether clinicians should develop awareness of how these systems weight certain types of speech. Doing so might enable more accurate documentation, but it also risks creating defensive practices where physicians actively counter patient attempts to emphasise certain symptoms, recreating the dismissive dynamics that already plague healthcare interactions.
Second, there’s uncertainty about whether transparency helps or hinders. Making AI scribe operation visible by acknowledging that they capture certain information better than others, could create space for collaborative documentation. But it might also provide a roadmap for strategic gaming, turning every consultation into a negotiation about what the official record should contain.
Third, there’s the educational dimension. Health professions training would need to address these dynamics, but toward what end? Teaching clinicians to recognise and counter gaming patterns risks positioning patients as adversaries with arms race dynamics. Teaching collaborative navigation of AI-mediated documentation acknowledges the reality but normalises what might be a problematic shift away from authentic clinical dialogue.
The governance challenge
This technology reminds us that clinical documentation was never purely objective; it’s always been shaped by who holds authority in the consultation. Clinical AI scribes don’t resolve this tension; they redistribute the mechanisms through which it plays out. Whether that redistribution serves equity or undermines care likely depends on context: the same dynamic that helps a chronically dismissed patient get proper attention could enable harmful manipulation in other circumstances.
Healthcare organisations implementing these systems face a choice: design governance frameworks that acknowledge narrative control explicitly, or assume documentation is neutral and act surprised when newer, more subtle versions of age-old power dynamics emerge.