AI scribes are on the rise. Instead of spending hours dictating or typing notes about a single patient interaction, clinicians can use an ambient AI to listen to their conversation and create a note. A few minor errors in an ordinary document might not be a huge deal. The same cannot be said for a medical record. That is why it is important to understand both the risks and limitations of AI Scribe Hallucinations and this technology.
Understanding AI Scribe Hallucinations: What They Are and Why They Happen
It is instances in which an AI-generated note contains information that was not actually stated or implied in the conversation that was overheard. The language may sound correct, making it challenging to spot. An ambient scribe typically listens to a conversation between a physician and patient, then transcribes and organizes the information into a note. Different issues can occur at different points in this process. For instance, the speech recognition software may not correctly interpret certain words. Or the final note may include assumptions that were not actually made. In some cases, a claim that was suggested but not confirmed may be presented as fact in the final note. For example,
2026 prospective study analyzing 31 physicians' notes and 7,545 clinic notes identified AI Scribe Hallucinations in 11.5% of notes reviewed and omissions in a larger portion of notes. Certain errors identified were deemed to have the potential for serious harm if not corrected. The point is not that all notes generated are inaccurate, but rather that physicians should always carefully review an AI's work, as errors can and do occur.
Why Does Ambient Documentation Accuracy Matter?
Accuracy in ambient documentation matters because the note may impact downstream care. A physician may review a note and base treatment decisions on the information included or excluded. A note containing incorrect medication information, symptoms, diagnosis, or exam findings can cause downstream clinicians to begin care from the wrong starting point.
Other errors could include:
- a medication being noted even when it was not discussed
- a symptom being documented that was not present or discussed
- a concern not being raised
- a possible diagnosis being presented as certain
- follow-up instructions being incorrect
Why Might an AI System Get Clinical Details Wrong?
AI systems typically perform exceptionally well at pattern recognition and natural language processing, but they are not perfect.
"If there is ambiguity in a conversation, an AI may make the most likely assumption and proceed to construct a note based on that assumption."
Speech recognition trips up in plenty of ways, and clinical documentation is no exception. Background noise, overlapping voices, accents, or just an oddly phrased sentence any of these can throw the accuracy off. Get the initial transcription wrong, and that mistake doesn't stay isolated; it flows straight into the final note.
What Types of Errors are Common in Clinical AI Notes?
Clinical AI notes can contain a variety of errors. The most common errors include AI Scribe Hallucinations, omissions, transcription errors, misplaced information, and unsubstantiated statements presented as facts. Not all errors carry the same level of risk. For instance, a minor grammatical error should be far less concerning than an incorrect medication or a wrong diagnosis.
| Type of error | Example | Why it matters |
| Hallucination | Unsupported information presented as fact | Can create false clinical history |
| Omission | Important piece of information missing | Could leave the note incomplete |
| Transcription error | Information captured incorrectly | Could alter the context and/or meaning |
| Context error | Information presented in the wrong context | Can cause misinterpretation |
| Overstatement | Unsupported possibility presented as fact | Can impact downstream care |
Why are AI-Generated Documentation Errors a Concern?
AI-generated documentation errors are a concern because they may appear quite reasonable. A typo jumps right off the page. A phrasing shift usually doesn't which is exactly what makes it risky. Picture a physician saying, "We may want to consider whether this could be condition X," and the note instead reads, "The patient has condition X." Read it back and nothing seems off spelling's fine, grammar's fine. Yet depending on what condition X turns out to be, that one swapped phrase can change everything downstream.. Similar errors could occur with medications, symptoms, examinations, or any other part of the note.
Physician-Reviewed Documentation is Still Necessary
Physician-reviewed documentation adds an essential level of safety due to the fact that the physician is in the best position to evaluate whether the note accurately reflects what took place. While the AI can create a note, only the physician can ensure that the note reflects an accurate recollection of the conversation that took place.
Depending on the context, a physician should carefully review the following:
- the chief complaint
- history of present illness
- relevant history
- medications
- allergies
- examination
- diagnosis
- testing
- treatment
- follow-up instruction, among other components.
Physician review of documentation is essential, in part because The American Medical Association has stressed the importance of physician review of AI-generated documentation, as physicians are legally responsible for the documentation they sign. That is an important consideration, as it serves as a practical reminder that while AI can generate a note, the responsibility ultimately rests with the physician. A potential workflow might include the following: patient visit, ambient listening, note generation, physician review, note correction, physician approval, and note addition to the medical record. The review is a crucial step in the process and should not be rushed. Physicians should ideally have ample time to review the note in order to catch any clinically significant errors.
AI Charting Liability Considerations for AI Scribe Hallucinations
Charting liability considerations are still relevant to AI note-taking systems due to the fact that automated tools do not eliminate the responsibility of the physician for the accuracy of the medical record. The liability implications will vary depending on the country, state/province, healthcare organization, documentation practices, physician licensing, contractual language, and specific circumstances of a given case. From where an organization sits, having clear-cut policies around documentation responsibility isn't optional it's necessary. Ideally, that means having answers to questions like these:
- Who reviews the note?
- Who corrects errors?
- Who approves the note for inclusion in the medical record?
- Are there established protocols for addressing recurring errors?
- How is the performance of the note-taking software monitored?
- What should happen if a serious documentation error occurs?
Ambient AI Malpractice Risk Considerations
Malpractice risk considerations are directly tied to the use of ambient AI systems and the review practices that support them. An incorrect note does not automatically result in malpractice, as it depends on the circumstances of a given case. However, malpractice risk can become an issue when AI Scribe Hallucinations lead to a clinically significant error and appropriate safeguards were not in place. Some factors that may raise malpractice risk include clinicians signing notes without adequate review, failure to review medication information, acceptance of diagnoses without verification, failure to acknowledge known issues with the software, failure to implement or follow training, and failure to investigate documentation errors, among other issues. The AMA Journal of Ethics has published several articles addressing legal and ethical considerations for ambient listening and transcription technology, including documentation practices and liability concerns.
Clinical Documentation Risk Considerations Beyond AI Scribe Hallucinations
Clinical documentation risk considerations extend beyond the issue of AI Scribe Hallucinations. There are several other factors that clinicians and organizations should be aware of, including the risk of important information not being captured, potential privacy and consent concerns, speech and recognition challenges, and the potential for automation bias.
Missed Information
An AI note-taking system may miss a piece of information that was actually discussed. This could be something as simple as an allergy or medication change, or it could be something more complex, such as a follow-up request.
Privacy/Consent
AI systems may capture private health information, necessitating appropriate privacy safeguards. Additionally, patients may need to be made aware of the technology and provided an opportunity to object if desired. As such, appropriate informed consent procedures should be established.
Speech/Recognition Limitations
Not every patient or physician speaks in exactly the same way. As such, accents, background noise, rapid speech, complex terminology, and multiple people speaking at once may impact speech recognition accuracy.
Automation Bias
There is a well-documented phenomenon known as automation bias in which individuals place undue trust in information generated by technology. This can lead to clinicians accepting information that is incorrect. The solution is not to distrust everything generated by technology, but rather to adopt appropriate verification safeguards. After all, an easy-to-read note that appears to be well-structured is not necessarily an accurate note.
How Can Healthcare Organizations Reduce AI Documentation Errors?
Reducing documentation errors requires more than simply adopting a particular AI tool. Organizations need to implement standardized protocols for documentation, including proper training, review procedures, and oversight. The following considerations may help reduce errors:
Establish Appropriate Use Cases
Healthcare organizations should establish appropriate use cases for AI-assisted documentation. This means identifying areas for which the tool can consistently deliver value and for which clinicians can reasonably review the documentation. This enables organizations to gradually build confidence in the tool's ability to support documentation.
Ensure Appropriate Physician Review
Physician review should never be optional it needs to stay part of the documentation process, with physicians always free to edit or revise the final note. How much review is "enough" really depends on the documentation itself and the situation at hand.
Investigate Repeated Issues
If the same issue keeps occurring, it should be investigated and addressed, rather than simply acknowledged and moved on from. For instance, if medication-specific issues keep arising, it may be necessary to investigate the root cause. It may also be necessary to evaluate issues related to specific terminology, particularly if notes are consistently being generated incorrectly in a specialty setting.
Ensure Proper Training
Training should go beyond basic instruction. Physicians and other clinicians who will be reviewing notes should understand the most common failure modes and know what to look for when reviewing notes. They should also understand when documentation should be done manually, when documentation errors should be reported, and how to address privacy concerns.
Conduct Independent Testing/Assessment
Organizations should conduct independent testing and assessment, rather than relying exclusively on vendor marketing. This will entail evaluating the software within the organization's own clinical setting in order to identify strengths, weaknesses, and appropriate use cases.
Is Ambient AI Documentation Safe?
Ambient AI documentation can be beneficial, but it should not be regarded as completely reliable or foolproof. There is value in the tool's ability to reduce documentation time and increase note consistency, but organizations need to be aware of the various risks and limitations that come with the technology. Research into the practical applications of AI note-taking is ongoing and, at this point, the evidence is mixed. Some studies have found positive outcomes relating to documentation efficiency and quality, while others have identified AI Scribe Hallucinations, omissions, and differences between AI-generated and physician-generated notes.
Conclusion
While AI scribes have the potential to reduce administrative burdens, organizations need to ensure that convenience does not come at the cost of clinical accuracy. This tool can help free up time for physicians, but it should not be regarded as a replacement for professional judgment. In practice, this means allowing a human to review the note, edit any information that seems incorrect, and approve the final note for inclusion in the medical record. Healthcare organizations can facilitate this process by ensuring that physicians receive the proper training, investigating recurring issues, conducting independent testing, and establishing appropriate policies around privacy, liability, documentation practices, and more. The goal should be to enable physicians to utilize this technology in a way that reduces administrative burdens without compromising the accuracy or integrity of the medical record.
FAQs
1. What are AI scribe hallucinations?
They occur when an AI scribe adds incorrect or unsupported information to a clinical note.
2. Can AI-generated clinical notes contain errors?
Yes. They may include incorrect details, missing information, transcription mistakes, or misleading summaries.
3. Why is physician review important?
A physician can verify whether the generated note accurately reflects the patient encounter before it becomes part of the medical record.
4. Can AI scribes create malpractice risks?
Potentially. If an inaccurate note contributes to patient harm, such as those caused by AI Scribe Hallucinations, questions about responsibility and liability may arise depending on the circumstances and applicable laws.
5. How can healthcare organizations reduce documentation risks?
They can use human review, train clinicians, monitor recurring errors, evaluate AI tools carefully, and establish clear privacy and documentation policies.