Can You Trust AI Clinical Notes? Hallucinations, Assumptions and How to Stay in Control in 2026

Yes, you can trust AI clinical notes for daily use in 2026, but only under three conditions: the clinician reviews every note before it enters the record, the template explicitly constrains the AI to what was actually said in the transcript, and the audio going in is clean. With those three controls in place, a modern AI scribe produces notes accurate enough for routine clinical documentation, and does it in a fraction of the time. Without them, you are gambling. The important shift in thinking is this: trust comes from the workflow you wrap around the model, rather than from the model alone. No vendor, ourselves included, should ask you to trust the AI unconditionally, and any vendor who does is telling you something about how seriously they take the problem.

The question behind the question is usually "will it make things up in my notes?", and it deserves a precise answer rather than reassurance. NHS-scale procurement questionnaires now ask vendors directly how they handle "hallucination", and individual clinicians ask the same thing in different words during every trial we see. So here is the precise answer: AI notes fail in two distinct ways, with different causes and different fixes, and once you can tell them apart you can control both.

Everything below unpacks that table: what hallucination actually means, why it happens, the concrete fixes, what the regulators and the research say, and the questions to put to any vendor before you commit.

What does hallucination actually mean in a clinical note?

"Hallucination" has become a catch-all word, and that vagueness is part of the anxiety. In practice, clinicians reporting that "the AI made something up" are describing one of two different events.

The first is a transcription error. The speech-recognition layer misheard a word and wrote down a similar-sounding one. You said one drug name, the transcript shows another. You said "no history of", the transcript dropped the "no". This is an audio problem wearing a software costume: the AI is faithfully summarising a faulty transcript rather than inventing anything itself.

The second is a generation error, and this is hallucination proper. The transcript is accurate, but the summarising model, whose job is to turn a rambling twenty-minute conversation into a structured note, adds something that was never said. Common examples: a red-flag screen recorded as performed when it was never mentioned, an abbreviation expanded into claims you did not make, a normal finding "assumed" because it usually accompanies the findings you did describe. The model is doing what language models do by default, which is to produce the most plausible-looking clinical note, and plausible is not the same as true.

These two failure modes get conflated constantly, including in vendor marketing, and the conflation matters because the fixes are completely different. A better microphone does nothing about generation errors. A stricter template does nothing about mishearing. You need both, and you need to know which problem you are looking at when a note goes wrong.

Why does the AI mishear words?

Transcription accuracy is driven overwhelmingly by the quality of the audio you feed in. Three factors dominate in a consultation room:

• Distance and movement. Speech recognition degrades predictably with every metre between mouth and microphone. Clinicians who work at a treatment couch, behind the patient, or across the room from the laptop are recording under much harder conditions than someone dictating at a desk.

• Background noise. Corridor bleed, reception noise, treatment tools and air conditioning all raise the error rate, and built-in laptop microphones weight all of it roughly equally with your voice.

• Real-time transcription modes. Some scribes transcribe as you speak, streaming words onto the screen live. It looks impressive in a demo, but streaming modes typically trade accuracy for speed: the engine commits to words before it has heard the full sentence, and gets fewer chances to correct itself from context. If your scribe offers the choice, turn real-time transcription off where stability matters more than spectacle, which in clinical documentation is essentially always.

There is also research evidence that transcription engines can do something worse than mishear. A 2024 peer-reviewed study presented at the ACM Conference on Fairness, Accountability and Transparency ("Careless Whisper", Koenecke et al.) found that a widely used speech-to-text engine occasionally fabricated whole phrases that appeared in no part of the audio, in roughly one per cent of the audio segments studied, with a meaningful share of those fabrications potentially harmful in a care context. The practical lesson is that even the transcription layer deserves scepticism, which is exactly why the verbatim transcript needs to be checkable rather than hidden (more on that below).

The cheapest and most effective fix for the whole category is hardware: a wireless clip-on microphone keeps the mic centimetres from your mouth for the entire consultation, whatever the room and however you move. We have written up what the Motics team actually uses and recommends, with UK prices, and the short version is that an £89 twin kit does more for note accuracy than any software upgrade on any vendor's price list.

Why does the AI add things I didn't say?

Because that is the default behaviour of the technology, and it has to be actively constrained.

The large language models that turn transcripts into structured notes are trained to produce complete, fluent, well-formed text. Left uninstructed, faced with a transcript that never mentions a red-flag screen, a model "knows" that good notes for this presentation usually include one, and the path of least resistance is to write it in. The same instinct expands abbreviations, fills in normal findings that were not examined, and tidies your actual consultation into the consultation the model expects. None of this is malice or malfunction. It is a text generator generating the most likely text, and the fix is to change what "most likely" means by giving it explicit rules.

This is why template design matters far more than most clinicians expect. A template is a set of plain-English instructions the AI must follow, and instructions measurably change behaviour. Three lines do a disproportionate amount of work:

1. Lock the AI to the transcript. A sentence as plain as "do not make any assumptions that are not absolutely clear in the transcript" is a genuine guardrail, one you can write yourself in your own words. If it was not said, it does not go in the note.

2. Keep clinical abbreviations verbatim. Instruct the template to write "No THREADS" exactly as you said it, rather than expanding the acronym into a list of itemised negatives. Expansion is where a shorthand you use precisely becomes a set of specific claims you never made.

3. No duplication across sections. Tell the template not to repeat content between sections. Repetition pads notes, and padded notes hide errors, because a reviewer skims what looks familiar.

The same logic applies to context. There are legitimate reasons for a note to draw on information beyond today's transcript, previous sessions being the obvious one, but that should happen because you deliberately supplied the information, never because the model guessed. It is exactly why we built Context, which lets notes draw on prior documentation you choose to provide, as an explicit feature with explicit inputs rather than leaving the model to fill gaps from imagination. The principle to hold vendors to: every piece of information in the note should have a traceable source, either today's transcript or context you knowingly provided.

How do you stay in control? The three controls in practice

Pulling the threads together, trustworthy AI notes rest on three controls, and all three are workflow decisions you can inspect during a trial.

Control one: the clinician reviews every note. The generated note is a draft until you have read it and signed it off. This is also the position of NHS England's guidance on ambient scribing: responsibility for the clinical record stays with the clinician, whatever tool produced the first draft. A good product makes this review fast rather than optional, by keeping the note tight, structured your way, and easy to check against the transcript. If review takes longer than writing the note yourself would have, the template needs work, and that is fixable.

Control two: the template constrains the AI. The guardrail instructions above, written in plain English, tested and refined against your own consultations. Crucially, template changes should be testable safely: in Motics, the template playground runs any change against your past sessions, so you can see exactly what a new instruction does to real notes without going anywhere near a live patient record. You can also compare different AI models on the same session, because models genuinely differ in how conservative they are, and the right way to find out is a side-by-side on your own material rather than a vendor's claim.

Control three: the audio is clean. A clip-on wireless microphone, sensible room habits, and real-time transcription switched off. Boring, cheap, and worth more than it looks.

None of these three is exotic. What separates vendors is whether the product treats them as first-class features or afterthoughts, and you can establish that in an afternoon.

Is the transcript or the note the legal record?

Both matter, but they play different roles, and the distinction is one of the strongest safety features available in this category, so it is worth understanding exactly.

In Motics, the verbatim transcript is locked the moment it is produced. It cannot be edited, by you or by anyone, and it stands as the medico-legal record of what was actually said in the room. The generated note is fully editable, because it is your clinical document: you correct it, restructure it, add your reasoning, and sign it off before it enters the patient record, then push it to your practice management system.

This architecture is what makes the review step meaningful. When something in a note looks off, you are never checking the AI against your memory; you check it against a fixed ground truth that neither you nor the model can retrospectively alter. A hallucinated finding is caught in seconds because it has no anchor in the transcript. An edited note and an unedited transcript can also be compared later, which is a materially better medico-legal position than a note whose provenance is unknowable.

One related point that surprises people: the audio itself is deleted within 48 hours at Motics. The lasting record is the locked transcript plus your signed-off note, which is a deliberately data-minimising design under UK GDPR, and worth asking any vendor to match.

What do regulators and researchers actually say?

The regulatory picture in the UK is further along than many clinicians realise, and it broadly supports the workflow-first view of trust.

NHS England published guidance on the use of AI-enabled ambient scribing products in health and care settings in 2025. Without reproducing it wholesale, its themes are the ones this article has covered: clinicians remain responsible for the records they sign, organisations should understand how a product handles errors and hallucination before deployment, and governance (data protection, clinical safety, device status) is not optional. If you are preparing a business case for a clinic group or an NHS-adjacent service, read it before you shortlist vendors.

The MHRA regulates software as a medical device, and its classification framework exists precisely because software that influences clinical care carries real risk. Motics is registered as a Class 1 medical device with the MHRA. Registration does not make a product infallible, but it does mean the vendor has accepted regulatory accountability for the product's intended purpose and safety processes, and it gives you a straightforward screening question for any shortlist.

Peer-reviewed research on ambient AI documentation is accumulating, and it is more encouraging than the anxiety suggests, provided you read it carefully. A large 2024 evaluation of ambient AI scribes across The Permanente Medical Group, published in NEJM Catalyst, reported strong clinician adoption and meaningful documentation time savings, while also cataloguing the error types that make clinician review non-negotiable. And the "Careless Whisper" study mentioned earlier shows that even the transcription layer can fabricate, which is the research case for a locked, inspectable transcript. The honest reading of the literature in 2026: the technology saves real time and produces usable notes, and every credible study assumes a reviewing clinician in the loop. No study supports unreviewed AI notes entering the record, and neither do we.

What AI notes still can't do

The clearest-eyed thing a vendor can tell you is where the technology stops, so here it is.

An AI scribe captures what was said and structures it. It does not, and in 2026 cannot, do your clinical reasoning. Whether the reasoning documented in the note is sound, whether the differentials are appropriate for the presentation, whether the treatment plan actually follows from the assessment: those judgements are yours, and no amount of transcription accuracy substitutes for them. A note can be a word-perfect record of a flawed consultation, and the AI will not notice.

This is sometimes framed as an argument against AI scribes. It is better understood as the strongest argument for the review step. The five minutes you spend reading a generated note is not proofreading for typos; it is the moment you check your own reasoning as documented, which conscientious clinicians did anyway when they wrote notes by hand, usually while exhausted at nine in the evening. A scribe moves that check to the moment after the consultation, with the conversation fresh and a verbatim transcript to hand, which is a better time to do it, with better raw material.

It is also why our Audit agent, which reviews note quality and completeness at scale on the Scale plan, is deliberately scoped to documented facts against your clinic's criteria. It is a first-pass screen. Clinical reasoning still needs a clinician, and any product that implies otherwise is overreaching.

How to choose a scribe you can trust

Trust, as this whole article has argued, is something you verify rather than something you are promised. These are the questions that separate vendors quickly, and what a good answer sounds like.

Run the test on your own consultations, not the vendor's demo. Record a normal clinic day, read every note against its transcript, count the corrections, then tighten the template and count again. Motics has a Free plan (£0, 25 credits, roughly 25 notes, no card) precisely so you can judge note quality on your own patients and your own accent before committing to anything. For the broader safety and compliance criteria that sit around note quality, from data protection to device status, we keep a fuller buying guide in how to choose safe AI phone and scribe agents.

Frequently asked questions

How accurate are AI clinical notes in 2026?

Accurate enough for daily clinical use when the workflow is right: clean audio in, a template that locks the AI to the transcript, and a clinician reviewing every note before sign-off. Accuracy is not a single number a vendor can honestly quote, because it depends heavily on your audio setup, specialty vocabulary and template quality, which is why testing on your own consultations beats any published claim.

Will an AI scribe make things up in my notes?

Left unconstrained, it can: summarising models fill gaps with plausible content by default, and even transcription engines have been shown in peer-reviewed research to fabricate occasional phrases. Constrained properly (a template instruction such as "do not make any assumptions that are not absolutely clear in the transcript", plus review against a locked transcript), fabrications become rare and, more importantly, catchable before they enter the record.

Do I really have to review every note?

Yes, and you should be suspicious of any workflow that makes review skippable. The clinician who signs the note is responsible for it, a position NHS England's ambient scribing guidance is clear on. In practice review is quick when the template is good, because you are checking a structured draft against a fresh consultation and a verbatim transcript.

Is the transcript or the generated note the legal record?

In Motics the verbatim transcript is locked and uneditable, serving as the medico-legal record of what was said, while the generated note is editable and becomes your signed clinical document. Ask any vendor to state their equivalent position explicitly; if the transcript can be silently edited or is discarded, you have no ground truth to check against.

What happens to the recording of the consultation?

At Motics, audio is deleted within 48 hours; the locked transcript and your signed note remain. Consent is captured as part of the recording workflow, and customer data is never used to train models. Whatever vendor you choose, the retention period for audio should be short, stated and verifiable.

Can I trial an AI scribe without risking live patient records?

Yes. Template changes can be tested in a playground against past sessions, so experiments never touch a live record, and the Free plan lets a single clinician trial real note generation at no cost. Start with low-stakes sessions, read everything against the transcript, and only widen use once the correction count is boringly low.

References

• NHS England, Guidance on the use of AI-enabled ambient scribing products in health and care settings (2025).

• Medicines and Healthcare products Regulatory Agency (MHRA), Software and Artificial Intelligence (AI) as a Medical Device guidance, gov.uk.

• Koenecke, A., Choi, A. S. G., Mei, K., Schellmann, H. and Sloane, M., "Careless Whisper: Speech-to-Text Hallucination Harms", Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2024.

• Tierney, A. A. et al., "Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation", NEJM Catalyst, 2024.

• Information Commissioner's Office (ICO), UK GDPR guidance, ico.org.uk.

Motics is the AI operating system for clinics: the Scribe agent described here sits alongside a Phone agent, Audit agent and Chat agent, built for UK clinics with the transcript-locked, clinician-reviewed workflow this article describes. If you want to see how it fits your clinic, start at motics.ai.