What are the main safety risks of AI scribes?
The core safety risk is that AI scribes regularly produce inaccurate clinical notes. In a randomized trial of 238 physicians using two different AI scribe systems (DAX and Nabla), clinicians reported that clinically significant inaccuracies occurred 'occasionally' on a five-point Likert scale, with average scores of 2.7 for DAX and 2.8 for Nabla (where 1 is 'never' and 5 is 'always') [1]. This means errors are not rare—they happen often enough to require constant vigilance.
A scoping review of 32 studies on AI speech recognition for clinical documentation found that errors include deletions, substitutions, and misrecognition of medication names or brief utterances [3]. The review noted that word error rates ranged from moderate in dictated notes to very high in conversational and emergency contexts, and that systems frequently missed clinically relevant details [3]. The authors concluded that 'unsupervised use is unsafe' [3].
Another paper warns that AI-generated notes can resemble 'unedited footage rather than finished stories,' lacking framing, causality, and prioritization [2]. This means the note may contain all the facts but miss the clinical reasoning that connects them, potentially leading to misdiagnosis or inappropriate treatment.
Do the benefits of AI scribes outweigh the safety risks?
The benefits are real but must be weighed against the risks. In the randomized trial, Nabla users saw a 9.5% decrease in time spent on notes, and both AI scribes led to improvements in burnout scores (Mini-Z scale increased by about 2.7-2.8 points on a 10-50 scale) and reductions in physician task load (by 31-40 points on a 0-400 scale) [1]. These are meaningful improvements for overworked clinicians.
However, the same trial reported only one mild adverse event, but the 'occasional' inaccuracies noted by clinicians suggest that errors are a persistent issue [1]. A commentary on the state of AI scribe evidence argues that adoption has 'occurred well ahead of robust evidence of their safety and efficacy' and that we still need to know 'the clinical outcomes achievable when scribes are used compared to other forms of note taking' [4]. This means we don't yet have solid data on whether AI scribes lead to actual patient harm, like missed diagnoses or medication errors.
The scoping review found that little research has linked system accuracy to patient safety or diagnostic outcomes [3]. So while AI scribes clearly reduce burnout, the direct safety impact on patients remains unknown—a gap that makes the risks hard to quantify but impossible to ignore.
Who should oversee AI scribes to keep them safe?
Human oversight is non-negotiable, and it should involve not just physicians but also nurses and other clinicians. One paper specifically argues that nurses are often excluded from the design and oversight of ambient AI tools, yet they face risks from hallucinations, omissions, and bias in the generated notes [5]. The author calls for empowering nurses through continuing education and leadership in model development, deployment, and auditing [5].
The DIRECTOR framework paper suggests that clinicians should act as 'directors' of the documentation process, using narrative principles like framing, causality, and resolution to edit AI-generated notes [2]. This reframes the clinician's role from passive recipient to active editor, which is essential for catching errors and preserving clinical reasoning.
The randomized trial authors explicitly state that 'occasional inaccuracies observed in either scribe require ongoing vigilance' [1]. Combined with the scoping review's conclusion that 'human oversight remains essential' [3], the message is clear: AI scribes are a tool, not a replacement for clinical judgment.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2025 to 2026, 5 from 2024 or later, 1 in Q1–Q2 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 48 papers retrieved from a database of over 500 million.
Sources used in this answer
Ambient AI Scribes in Clinical Practice: A Randomized Trial
In a randomized trial of 238 physicians, AI scribes reduced burnout and documentation time (Nabla cut time by 9.5%), but clinicians reported 'occasional' clinically significant inaccuracies on a five-point scale (scores 2.7-2.8), requiring ongoing vigilance [1].
The Clinician as DIRECTOR: Operationalizing Cinematic Storytelling for Clinical Documentation in the Age of Artificial Intelligence
This conceptual paper argues that AI-generated notes often lack narrative structure (causality, prioritization, closure), and proposes a 'DIRECTOR' framework for clinicians to edit notes and restore clinical reasoning [2].
Assessing the Reliability, Accuracy, and Relevance of Artificial Intelligence Speech Recognition for Clinical Documentation: A Scoping Review.
A scoping review of 32 studies found that AI speech recognition tools have substantial accuracy limitations, with errors including misrecognized medication names and missing clinically relevant details, concluding that unsupervised use is unsafe [3].
AI Scribes: Are We Measuring What Matters?
This commentary argues that AI scribe adoption has outpaced safety evidence, and that we still lack data on clinical outcomes (e.g., missed diagnoses) when using these tools compared to traditional note-taking [4].
Invisible Scribes: Can Nurses Trust Ambient AI for Clinical Documentation?
This paper highlights that nurses are often excluded from AI scribe design and oversight, exposing them to risks from hallucinations and bias, and calls for nurse education and leadership in model auditing [5].