AI was supposed to ease doctors’ workload – instead they spend hours correcting errors

AI-powered clinical documentation was meant to streamline work at Danish hospitals. But in practice, doctors are spending valuable hours correcting mistakes, training algorithms, and handling administrative tasks that used to be managed by medical secretaries, according to new research from the Univ
Photo: Getty Images

As AI becomes part of everyday healthcare, it brings along unforeseen extra work that pulls time and resources away from core medical tasks.

This is the conclusion of information science researcher Silja Vase, whose PhD research examined how speech recognition technology is used in daily routines at a Danish hospital. Over two years, she followed doctors, medical secretaries, and IT staff as they worked with automated speech recognition systems designed to save time by transcribing dictated notes directly into patient records.

In reality, however, the technology mistranscribes so frequently that doctors end up spending hours correcting words, sentence endings, and medical terminology. The corrections they make then serve as training data for the system’s language model.

“AI is not plug‑and‑play. For these technologies to function with the necessary precision, they require continuous maintenance and data quality assurance - and that work ultimately lands on healthcare staff, who already have many other duties,” says Silja Vase.

About the research

Silja Vase’s PhD dissertation Shifting the Script – Clinical Documentation Moves to the Frontstage is based on ethnographic field studies at a Danish hospital and combines observations, shadowing, and interviews with doctors, nurses, medical secretaries, and IT staff.

The dissertation consists of five research articles, two of which have already been published in academic journals.

In other words, doctors are not just fixing errors in their own notes. Every correction becomes training input for the language model and feeds data back to the system provider so the model can be adjusted. “That work largely falls on clinical staff and it takes time away from patient care,” Vase explains.

According to Vase’s calculations, doctors were interrupted roughly 602 times per month due to corrections. On average, each doctor experienced more than 40 interruptions per shift caused by stopping to fix transcription errors.

Working blind

Another challenge Vase identifies is what she calls “pseudo-data work.” Healthcare workers spend significant time correcting and validating the system’s suggestions to improve the algorithm, but they have no way of knowing whether their efforts have any effect.

“The problem is that they can’t see whether their corrections actually make a difference. The system’s statistical feedback isn’t available to them, so they risk working entirely in the dark,” she says.

In one extreme case, an entire department’s corrections were never forwarded to the vendor for a full year due to a technical error - and no one noticed.

Extra work for doctors

Automation of documentation has also led to a range of unanticipated practical tasks falling on doctors.

“Previously, medical secretaries scheduled X‑rays and other tests as part of transcribing the doctor’s notes. Now that doctors are responsible for documentation themselves, these tasks shift to them as well,” says Vase.

In some cases, this leads to bizarre workflows where doctors print out records and mark them up with a special code so that secretaries can later schedule X-rays or other tests.

“It’s yet another form of extra work that illustrates how AI doesn’t simply automate isolated tasks; it reshapes the entire division of labour. If systems aren’t designed with real‑world practices in mind, new types of extra work inevitably arise,” she adds.

AI must adapt to hospital workflows - not the other way around

According to Vase, greater transparency and support are needed when AI systems are introduced in healthcare. Otherwise, it becomes difficult for staff to understand - let alone engage with - the technology.

“Speech recognition isn’t just a tool that lightens the documentation burden. It fundamentally affects how staff work and creates informal collaborations where doctors and other healthcare professionals take on new roles and spend more time on tasks far removed from their core responsibilities. That also means that promised financial savings are sometimes far smaller than expected,” Vase says, adding:

“When AI technologies such as speech recognition are implemented in healthcare, it is crucial that they are adapted to existing workflows - not the reverse. The technology should serve the staff, who should not be forced to adapt to the system’s requirements, especially when no extra time is allocated for these tasks. That places demands on how these systems are fundamentally designed.”

Contact