When AI Gets It Wrong at the Bedside: Errors, Bias, and Potential Nurse Liability

Part of Nurse.org’s Nursing AI Watch, our ongoing investigation where we track how artificial intelligence is reshaping bedside nursing. Here we look at the harder question: where tools demonstrate documented errors, who gets hurt when they do, and why the nurse’s signature may still carry the risk.

AI scribes promise to give nurses their time back. But what happens if AI gets something wrong? 

The efficiency gains of using AI in clinical charting are documented. However, for nurses, so are the risks. As a nurse using AI in charting, the signature is still yours, the accountability is still yours, and the patient in front of you still deserves a chart that reflects what actually happened.

“I love having the condensed info for a quick reference, but I don’t trust it completely and continue to look up information independently,” Ashley Sue Ann, an ER nurse from Michigan, confesses to Nurse.org. “I have not found any errors as of yet, but I am also not willing to risk patient safety for it. “

Here’s what the evidence actually says, and how to protect your license.

Want to see more Nurse.org articles in your Google results? Add us as a preferred source.

More often than most nurses expect. A 2025 study published in the Journal of Medical Internet Research reviewed notes produced by ambient AI scribes and found that around 70% of them contained at least one error. 

The errors fall into two broad buckets. Some are omissions: the tool simply drops a detail the nurse said out loud, like a reported symptom or a medication change. Others are fabrications: the tool adds something that was never said, inventing a detail that sounds plausible but isn’t true. Both are dangerous, but fabrications are especially hard to catch because they read like normal documentation.

The concern isn’t coming from skeptics alone. ECRI, an independent patient-safety organization, ranked AI as the number one health technology hazard for 2025, ahead of every device and drug-delivery risk on its list. 

A separate analysis in NPJ Digital Medicine looked at how often AI tools “hallucinate,” aka generate content with no basis in the source. The overall hallucination rate was low, around 1.47%. But of those hallucinations, researchers classified roughly 44% as major, or serious enough to potentially change clinical understanding. A small percentage of a high volume of notes is still a lot of charts.

Nurses are also still wary of AI implementation in their charting. An Incredible Health survey of over 2,200 nurses released on July 7, 2026, revealed that the majority of nurses, like Sue Ann, do not fully trust AI yet in their charting or guiding clinical decisions, so they continue to do their own research and fact-checking. 

The numbers stop being abstract the moment you’re on a unit. Picture an AI scribe that records a patient denied chest pain when they actually reported it. Or one that logs a fluid status that doesn’t match what the nurse observed. Small wording changes can flip clinical meaning.

We’ve seen this play out in reported coverage of AI systems and dialysis fluid decisions, where a tool’s recommendation reportedly ran against the nurse’s own read of the patient. The nurse’s judgment caught it. The lesson wasn’t that the tool was useless; it was that human oversight should be the safety net, not the backup.

“I don’t believe AI should or could ever replace nursing judgment,” Danielle K. Miller, DNP, RN, a healthcare consultant and nurse executive from Chicago, Illinois, who advises healthcare organizations on AI integration into clinical workflows, argues. “Instead, it should augment it.”

Fabrication can get stranger than a wrong number. In one reported case, an AI chatbot essentially invented a fake disease, confidently generating a condition that doesn’t even exist.

Errors don’t land evenly. Some patients get worse AI performance than others, and that’s a safety issue with an equity edge.

Researchers at Columbia University School of Nursing have warned that the rush to adopt AI scribes carries real bias risks. Tools trained mostly on certain speech patterns can struggle with accents, dialects, and the speech of patients from marginalized communities. This leads to producing less accurate notes for the people who already face the most barriers in healthcare.

Bias also shows up in charting and triage logic, not just transcription. We’ve covered racial bias and errors in AI charting in depth, because the pattern is consistent: when a tool learns from historical data that already reflects unequal care, it can repeat and even amplify that inequality. The patients hurt most are often the ones least able to push back.

For nurses, the practical takeaway is simple. If a tool’s output seems off for a particular patient, don’t assume the patient is the outlier. The tool may be the one that’s struggling.

This is the part that keeps nurses up at night, and the answer is blunt: you are.

The American Nurses Association’s position statement on the ethical use of AI emphasizes that the nurse remains accountable for clinical decisions and documentation, even when a technology failure contributes to an error. A faulty AI tool generally doesn’t transfer your professional responsibility to the vendor, and the signature on the note is yours. (This is general professional guidance, not legal advice; how liability actually plays out can depend on your state, your employer’s policies, and the specific facts.)

That sounds harsh, but it’s also clarifying.

It means the safest posture is to treat AI output as a draft, not a verdict. The tool can suggest; you decide. We’ve broken down what that accountability looks like in practice in our coverage of the ANA’s AI nursing guardrails.

The same standard that protects patients also protects you. If you review every AI-generated note before signing, correct what’s wrong, and document your own clinical reasoning, you’re doing exactly what your license requires — and exactly what an attorney would later look for.

You can’t audit every keystroke, but you can build habits that catch the errors that matter. Use this checklist as a starting point.

AI error types nurses should watch for:

1. Omissions — A symptom, medication, allergy, or instruction you reported out loud is missing from the note.
2. Fabrications — A detail appears in the note that you never said and the patient never reported.
3. Flipped meaning — “Denies chest pain” versus “reports chest pain.” Negations and small qualifiers are easy for a tool to invert.
4. Wrong attribution — A statement assigned to the wrong person, or a finding attributed to the wrong visit or body system.
5. Stale or copied data — Old values carried forward, or details that look pulled from a prior encounter rather than today’s.
6. Quantitative drift — Numbers that don’t match what you measured: vitals, intake/output, doses, lab values.
7. Bias-prone gaps — Lower-quality capture for patients with accents, limited English, or atypical presentations.

A few verification habits turn that list into protection:

• Read before you sign. Always. Treat the AI note as an unverified draft.
• Read the negatives. Errors hide in “denies,” “no,” and “without.” Confirm each one.
• Cross-check the numbers against your own documentation and the monitor.
• Slow down for high-risk patients — complex cases, language barriers, and anyone where a small error could cascade.
• Report errors you catch to your manager or patient-safety officer so the pattern gets tracked, not buried.

>>Listen to The Latest Nurse News Podcast

Nurse.org’s State of Nursing Survey 2026 also pointed out the potential problem that younger, new nurses “raised” on AI could encounter: without a base of hands-on, physical knowledge built up, will they have the know-how to catch when AI is wrong? AI is changing the very way future nurses will be trained and approach patient care. 

“We are seeing a tremendous loss of critical thinking in nurses,” reported one nurse in the survey. “AI and algorithms have increasingly left nurses in a task job as opposed to a career.”

“The advent of having a computer in your pocket (smart phone) is that you don’t have to know anything,” cautions another. “You don’t have to know what the medicines do that you are delivering, you don’t have to know your ACLS algorithms, you don’t have to know anything. Add AI into the mix; if your AI answer is wrong, how will YOU know?”

Individual vigilance matters, but it shouldn’t be the only safeguard standing between a flawed tool and a patient. Strong oversight is a system responsibility that starts clear governance, a nurse at the table when AI tools are chosen and monitored, transparent disclosure to patients, and a real process for reporting and fixing errors.

When that infrastructure is missing, the entire burden falls on the bedside nurse, who is also the last line of defense. We dig into who actually governs these tools — and how rarely a nurse is in the room — in our companion analysis on nursing’s AI governance gap.

The vendors are not ignoring the accuracy problem — though it’s worth treating their claims as claims, not independently verified outcomes.

The newer ambient-documentation tools market a set of built-in safeguards aimed squarely at the error types above. Microsoft, for example, says its Dragon Copilot includes clinical safeguards designed to flag hallucinations and omissions, validate clinical codes, and surface citations so a clinician can trace where information came from. In October 2025, Microsoft also extended Dragon Copilot to nursing workflows and integrated it with Epic’s Rover app, a sign that vendors are starting to design specifically for nurses, not just physicians.

The broader industry trend is to address one root cause of error directly: missing context. Recent research has found that ambient scribing alone can miss critical patient information, and that documentation gets more complete when the tool is grounded in a patient’s longitudinal record rather than just the audio of a single visit. Vendors increasingly lean on clinician-feedback loops and “human-in-the-loop” review to refine accuracy over time. Ambience Healthcare, for example, launched an inpatient nursing suite in June 2026 that it says uses “chart-aware reasoning” over the full EHR record plus a three-tier safety architecture meant to catch hallucinations — again, a vendor’s stated safeguard, not an independently verified result.

Two honest caveats. First, most of this is the vendors’ own framing; the independent, peer-reviewed error rates above were measured on real and simulated notes, and a safeguard feature is not the same as a proven reduction in patient harm. Second, every one of these designs still assumes a clinician reviews and signs the note. The fixes are real and moving in the right direction, but they reinforce rather than replace the nurse’s role as the final check.

AI documentation tools are already on units, and they make mistakes at a rate the research can no longer ignore. The errors range from quiet omissions to confident fabrications, and they don’t fall evenly across patients. Through all of it, the accountability stays with the nurse. Use the tools carefully as a draft to verify and guide, but never trust a decision blindly. 

Related Nursing AI Watch Analysis:

🤔 Have you caught an AI tool getting something wrong in a chart? Share your experience in the comments below.

About the data: Nurse.org’s Nursing AI Watch is built from a structured dataset of AI deployments across 106 large health systems, refreshed each reporting cycle. The error and bias figures in this article come from independent, peer-reviewed research and named patient-safety organizations, not from our own dataset or from vendor marketing. Where we describe a vendor’s safeguards, we label them as the company’s own claims. When the evidence supports only a narrow statement, we scope the claim to match, and when we can’t confirm something, we say so.

Nurse.org Analysis