The Scribe Dilemma

AI documentation tools are getting blamed for the wrong problem. The real failure isn’t accuracy. It’s authorship.

By Luis Cisneros, CEO | Serelora

Something strange is happening to clinical notes across American medicine, and almost nobody is talking about it honestly.

Run five charts from five different clinicians through the same AI scribe and you will get five notes that read like they were written by the same cautious, mid-level committee member who has never once touched a patient. The phrasing is immaculate. The tone is beige. The reasoning, if you can still call it that, has been sanded down to a smooth regulatory finish that says everything a payer wants to hear and almost nothing another clinician actually needs to know.

This is not a glitch. It is the product working exactly as designed. And that is the problem.

The entire AI scribe industry has organized itself around one question: how do we make the note more accurate? Meanwhile, the question that actually matters has gone almost entirely unasked: whose voice is this, and why does it no longer belong to the person who was in the room?

The Authorship Collapse

A clinician who says, “Your back pain is just compensation, it’s secondary, we’ve got to sort your knee out and the pain will disappear,” is not being imprecise. That sentence is doing extraordinary work. It names a biomechanical cause-effect chain. It assigns clinical priority. It expresses a level of confidence that communicates urgency to the patient and, just as importantly, to every downstream clinician who reads the note. The word “just” is doing more labor than an entire paragraph of hedged prose. The phrase “will disappear” is a clinical commitment, a decision marker that says: I am confident enough to stake my reasoning on this sequencing.

Now watch what the scribe does with it.

“Lumbar symptoms are considered secondary to primary knee pathology. Addressing the knee dysfunction is expected to resolve back pain.”

Same facts. Entirely different document. The causal chain has been replaced by passive attribution. The confidence has been swapped for bureaucratic hedging. The clinical priority, which was the whole point, has been flattened into two equally weighted clauses that tell the next reader almost nothing about what this clinician actually thinks, how certain they are, or what they would do first if forced to choose.

That is not documentation. That is a translation error with a malpractice tail.

And it is happening millions of times a day, in every specialty, in every setting where an AI scribe sits between the clinician’s mouth and the medical record. The note stops being an artifact of clinical thought and becomes an artifact of linguistic processing. The author changes. The voice disappears. What remains is compliant, defensible, and dangerously hollow.

Recent research published in JAMIA confirms what clinicians have been feeling in their bones — the notes don’t sound like them. And that is not a cosmetic complaint. When the voice changes, the signal changes. When the signal changes, the next clinician reading that note is working with degraded information dressed up in professional formatting. It looks like a note. It parses like a note. But the thing that made it medically useful, the reasoning, the stance, the hierarchy of concern, has been quietly removed and replaced with something that reads more like a compliance document than a clinical one.

Two Languages, One Impossible Note

To understand why this keeps happening, you have to understand that healthcare runs on two completely different languages that were never meant to occupy the same sentence.

Clinical language is built for decision-making under uncertainty. It is direct, often conversational, full of shorthand and implicit priority markers. When a physician writes “classic presentation” or “looks septic” or “this is the knee talking,” they are compressing enormous amounts of clinical reasoning into phrases that other experienced clinicians can instantly decompress. The messiness is not a flaw. The messiness is the fidelity. Real clinical reasoning is nonlinear, probabilistic, and full of confident bets made under incomplete information. The language reflects that, and it should.

Administrative language serves an entirely different master. It is built for justification after the fact, optimized for billing rules, audit defense, prior authorization logic, and utilization review. It is structured, coded, and deliberately cautious because its audience is not another clinician but a payer, a regulator, or a plaintiff’s attorney. This language is not wrong. It is simply answering a different question. Clinical language asks what do I think is happening and what should we do? Administrative language asks can we prove this was necessary and get paid for it?

Two languages. Two purposes. Two entirely different audiences.

Now here is where it gets ugly, and where it gets personal.

I grew up between Spanish and English. I speak both. I live in both. And like anyone who has ever navigated two languages in the same household, the same neighborhood, the same sentence, I know Spanglish intimately. Not from a textbook. From the kitchen table.

Spanglish is what happens when two languages collide in the same environment and neither one fully wins. It borrows Spanish grammar with English nouns, English sentence structure with Spanish slang, mixed verb conjugations, half-translated idioms that land perfectly if you grew up inside the culture and land nowhere if you did not. It is vivid, economical, and deeply expressive within its native context. Among people who share the hybrid, Spanglish is a superpower. It says things neither language can say alone.

But here is the thing about Spanglish that nobody romanticizes — it only works when everyone in the room shares the code.

A pure English speaker hears Spanglish and catches fragments. They follow the nouns, maybe the verbs, but the grammar feels wrong and the context is missing. They get the shape of the meaning but not the weight. A pure Spanish speaker hears it and winces, because the phrasing is unnatural, the words don’t carry their proper emotional gravity, and it sounds like someone is breaking the language on purpose. Both groups understand pieces. Neither group fully trusts the output. Spanglish becomes this strange middle zone where the words are familiar enough to follow, foreign enough to confuse, and inconsistent enough to feel unreliable.

That is beautiful when you are ordering cafecito from your tía at the counter. It is catastrophic when you are communicating a differential diagnosis across a care team that never met each other.

AI scribes are producing clinical Spanglish. They take the clinician’s raw reasoning and inject billing vocabulary into the narrative of the visit. They wrap compliance language around clinical judgment. They thread administrative phrasing through the clinician’s causal reasoning like English nouns crammed into Spanish grammar. What comes out is not an optimized note. It is a third dialect, a hybrid that is neither a real clinical note nor a real administrative artifact, and it is being generated at industrial scale by systems that have no idea they are doing it.

And now only one group can interpret it fluently — the tiny population of people who live in both worlds at once. The clinic manager who used to be a biller. The physician who taught herself coding. The administrator who once worked bedside. Everyone else is reading fragments and filling in the gaps with assumptions.

The clinician reads the note and thinks: this doesn’t sound like me. Where is the decision? Where is the priority? Why does it read like a lawyer drafted it? The administrator reads the same note and thinks: this is still vague. It doesn’t clearly justify medical necessity. It doesn’t map cleanly to the rules I need it to map to. Both parties catch pieces. Neither party fully trusts what they are reading. And nobody says anything, because the note looks professional and the AI produced it quickly and questioning the machine feels like questioning progress itself.

No. That is the worst possible outcome. A note that loses clinical signal and still fails administrative clarity. A document that satisfies nobody and risks everyone.

Spanglish is powerful socially because it creates intimacy among people who share the hybrid. But it is dangerous for precision because it produces a third language that only a small subset of speakers can truly interpret. Healthcare cannot afford a documentation language that only hybrid speakers understand. In medicine, misunderstanding is not merely annoying. It breaks care. It breaks reimbursement. It breaks trust. And unlike a fumbled cafecito order, nobody laughs it off at the end.

This is the scribe dilemma in its purest form… one document is being forced to perform two fundamentally incompatible jobs, and the AI is making it worse by being very, very good at producing text that looks like it solved the problem while quietly inventing a dialect that nobody asked for and nobody can fully rely on.

The Downstream Wreckage

The consequences of authorship collapse do not stay inside the note. They propagate.

When a clinician’s confidence markers are stripped, the next provider reading that note cannot distinguish between a finding the original clinician was certain about and one they were merely considering. “Expected to resolve” does not carry the same clinical weight as “will disappear.” One is a hedge. The other is a bet. That distinction matters when the next provider is deciding whether to pursue additional workup or trust the prior assessment. Flatten it, and you have introduced diagnostic noise into a system that was already drowning in it.

When causal reasoning is replaced by passive summarization, the clinical logic becomes invisible. A note that says “lumbar symptoms are considered secondary to primary knee pathology” does not tell you why the clinician believes this, what evidence led to that conclusion, or how strong the inferential chain is. It tells you only that a conclusion exists, stripped of its supporting architecture. This is the medical equivalent of showing someone the answer to an equation without the work. It is technically informative and practically useless for the next person who needs to build on that reasoning.

When every note sounds the same, pattern recognition across a patient’s chart degrades. Clinicians reading longitudinal records rely, more than they often realize, on subtle shifts in language, tone, and emphasis to detect when a colleague became more or less worried, when a diagnosis crystallized, when a treatment plan shifted from exploratory to committed. Homogenize the voice and you homogenize the signal. The chart becomes a flat, undifferentiated wall of text where every visit reads like every other visit, regardless of what actually happened.

And then there is liability. The legal fiction of AI-scribed notes is that they represent the clinician’s documentation. But when the phrasing, tone, and reasoning structure have been fundamentally altered by an intermediary, what exactly is the clinician signing? They are attesting to a document they did not write, in a voice they do not recognize, expressing a level of certainty they may not have intended. Plaintiff’s attorneys have not yet fully discovered this seam, but they will. And when they do, the question will not be whether the note was accurate. It will be whether the note faithfully represented the clinician’s actual reasoning, or whether an AI quietly rewrote the standard of care.

The Fix Nobody Wants to Hear

The solution is not better prompts, fancier templates, or more sophisticated natural language processing applied to the same flawed architecture. The solution is separation.

Preserve the clinical note as a clinical artifact. Keep the clinician’s voice, reasoning chain, prioritization, and confidence markers intact. The rawness is not noise. It is often the only place in the entire healthcare data ecosystem where clinical intent is truly visible. A note that sounds like the clinician who wrote it is not an unpolished note. It is a high-fidelity note. Those are not the same thing, and the industry’s confusion on this point has been the single largest driver of documentation decay.

Then generate administrative outputs separately, downstream, as distinct artifacts purpose-built for the workflows they serve. Instead of rewriting the clinical note to satisfy billing, prior authorization, utilization review, or claims defense, translate clinical intent into administrative logic as an explicit, auditable step. The billing letter is not the note. The prior auth justification is not the note. The coding summary is not the note. They are derived products, and they should be produced by reasoning layers that understand both the clinical source and the administrative target without corrupting either.

This is the architectural insight that most of the scribe market has missed entirely. You do not fix authorship collapse by making the AI write more like a clinician. You fix it by stopping the AI from rewriting the clinician at all. You let the note stay a note. You let the system produce everything else.

Serelora is built around this separation. The clinical record stays faithful to the clinician and the patient. The reasoning, the confidence, the voice, all of it preserved, because that is where the medical value lives. Administrative outputs are generated downstream through structured reasoning, explicit justifications, and workflow-specific artifacts that match how healthcare actually operates. Not by laundering clinical language into something a payer might accept, but by maintaining two clean signal paths where one corrupted hybrid used to be.

The scribe dilemma was never about whether AI could document a visit. It can. The dilemma is that documentation is not transcription, and a note is not a form. A clinical note is a compressed, authored representation of one human being’s reasoning about another human being’s health. The moment you let a machine silently replace that authorship, you have not automated documentation. You have automated its destruction.

Preserve authorship. Translate downstream. Stop blending two languages into something nobody fully trusts and everybody quietly edits on their way out the door.

Whoever figures this out does not just win the scribe market. They fix the record.