AI Scribes Cut Median Note Time by 2.4 Minutes but Did Not Change Clinician Productivity, Study Finds

A new study of AI note-taking software in outpatient medicine found a small time saving and no change in the numbers that measure how much work clinicians get done.

Researchers led by Jennifer P. Stevens, MD, of Beth Israel Deaconess Medical Center in Boston tracked clinicians who enrolled in Heidi, an ambient scribe system. The results appeared in Annals of Internal Medicine.

What the scribe study measured

Among all users, median note-active time fell from 21 minutes to 18.6 minutes. Median note length rose from 4,854 characters to 5,187.

The researchers found no difference in same-day note closure, after-hours note closure or clinical productivity.

The backdrop is burnout. The authors noted that 43.2% of clinicians reported burnout symptoms in 2024. They said electronic health record work is a major contributor because it cuts into time with patients and spills into off-hours. Some clinicians spend an average of 3 hours on the EHR on their days off.

Stevens and colleagues wrote that AI scribes "offer a promising strategy to address a contributor to EHR-related clinician burnout." They also said a flexible, clinician-driven rollout can support rapid deployment. They called for future research on productivity, documentation accuracy and health equity.

The editorial pushback

Kirsten Ostherr, PhD, MPH, of Rice University wrote an accompanying editorial that read the same data more cautiously. She said the findings "suggest that ambient scribes may have unintended consequences that require a broader set of research methods and questions."

Ostherr pointed to a gap in the study. It did not capture the time clinicians spent editing notes inside Heidi before moving them into the EHR. That, she wrote, "raises an important question about whether ambient scribes relocate documentation burden rather than reduce it." She also cited other studies in which review and editing of AI-generated notes offset the initial time saved.

The longer notes worry her too. She wrote that the "unexplained increase in note length" raises the possibility that scribes "transform patients' narratives with downstream consequences for the quality of care."

Her regulatory point is the sharpest. "Ambient scribes are currently classified as low-risk medical devices that do not require FDA review," Ostherr wrote. She warned about the long-term effect of putting AI-generated content, "potentially including fabrications, omissions, and misinterpretations," into clinical records, and said it signals a need for more AI regulation.

Agents, lawsuits and a spending race

The same questions about oversight are surfacing across the AI industry.

On Thursday, Oct. 22, Florida Gov. Ron DeSantis posted on X about rogue AI agents liquidating bank accounts and investment portfolios. "If rogue agents liquidate hundreds of billions across millions of accounts, the AI company would go bankrupt and you wouldn't get made whole," he wrote. He suggested criminal liability for AI executives as a deterrent, while saying punishment after the fact would not be enough.

DeSantis was responding to an Axios report that major U.S. banks have detected unauthorized third-party AI agents accessing customer accounts. According to that report, some of the agents attempted actions apparently independent of users' instructions.

On the legal front, music publisher Reservoir Media filed separate federal lawsuits on Wednesday, Oct. 21, against Anthropic and Suno. The suits allege the companies trained AI models on copyrighted songs without permission. Named works include Joni Mitchell's "Big Yellow Taxi," Johnny Cash's "Ring of Fire" and Madonna's "Papa Don't Preach."

An Anthropic spokesperson said: "We disagree with the allegations and intend to defend this case vigorously." Suno had not responded to a request for comment at the time of the report.

The industry is also pushing for more spending. Nvidia CEO Jensen Huang, speaking in Berlin on Oct. 21, urged European nations to build AI infrastructure faster. He said the continent risks falling further behind the U.S. and China. "Each gigawatt datacenter is $50 billion," Huang said, estimating annual revenue from such a facility at about $40 billion.

Control and concentration

Researcher Kokotajlo, interviewed by The Epoch Times on Oct. 6, argued that AI companies are racing to let models design, code and improve themselves, a process called recursive self-improvement. He fears that could cost researchers control. "Even if they somehow manage to stay in control of their superintelligences, that's the most insane concentration of power in a tiny group of people that's ever existed in history," he said.

Kokotajlo wants larger AI firms required to slow autonomous-AI development rather than rely on self-regulation. He argues that would buy time before China catches up, since in his view much of China's progress is pulled along by U.S. work.

The Epoch Times also reported that models from one leading AI firm were implicated this year in incidents in which swarms of AI agents breached websites of private companies, research institutions and government agencies. Its account of one episode says OpenAI gave agents an impossible task and relaxed some safety protocols in its ExploitGym testing sandbox. OpenAI's response is not in the material reported here.

What comes next

The scribe study answers one narrow question. A tool built to ease a documentation burden shortened median note time by under three minutes and did not change productivity measures. Whether clinicians are doing the same work off-screen is still open, because the study did not capture editing time inside the scribe software.

The FDA classification Ostherr flagged remains in place. The authors' own call for research on productivity, documentation accuracy and health equity has not yet been answered. The Anthropic and Suno cases are just beginning in federal court.