Replacing '3-Minute Consultation' Doctors with AI? [AI Error Notes]

⑫ AI may excel in task-level skills, but jobs are not just a sum of tasks, they are a 'system'
IBM jumped in focusing only on task replacement and got a harsh lesson

Editor's NoteExamining failures is the shortcut to success. 'AI Error Notes' explores failure cases related to AI products, services, companies, and individuals.

Replacing '3-Minute Consultation' Doctors with AI? [AI Error Notes]

AI is a hot topic in the medical field as well. Photo by Getty Images Bank

"Let's replace doctors who just glance at patients for a mere 3 minutes and rake in money with artificial intelligence (AI) once and for all!"

The conflict between the government and the medical community over healthcare reform shows no sign of ending. Not only patients experiencing medical inconveniences but also citizens who are potential patients cannot hide their anger at the current situation. The above quote was a comment on an article about the medical-policy conflict, which received numerous 'likes.'

Domestic patients have many complaints about doctors. On May 14 last year, a public hearing was held at the National Assembly. The topic was 'Healthcare system scenarios desired by the public.' The number one complaint was 'short communication,' represented by the '3-minute consultation.' This also influenced public opinion in favor of increasing medical school admissions.

So, can AI really replace doctors? How exposed are doctors to occupational risks due to AI?

The Father of AI: “Radiologists Should Stop Training… AI Is Better”

Jeffrey Hinton, a Nobel laureate in Physics and an AI scholar, once said:

"If you want to become a radiologist... it’s like not knowing there’s no ground beneath your feet.
We should stop training radiologists now.
It is almost certain that deep learning will surpass radiologists within five years."

Radiologists interpret images to detect medically significant abnormalities. From Hinton’s perspective as an AI expert, this is clearly a task AI can perform better. Machine learning excels at identifying and classifying anomalies in visual images and videos. Moreover, accuracy improves with more training data.

Greg Brockman, co-founder and president of OpenAI, the developer of ChatGPT, posted on his X (formerly Twitter) account on November 17 last year (local time), sharing an article from The New York Times (NYT) titled “Interesting small-scale study results on disease diagnosis accuracy.”

The article reported that ChatGPT outperformed humans in disease diagnosis. According to the article, human doctors had a diagnostic accuracy of 74%, doctors using ChatGPT had 76%, and when ChatGPT diagnosed alone, the accuracy was an astonishing 90%.

There is also research showing that AI’s cancer detection rate is 15% higher than that of human doctors. South Korean medical AI company Lunit revealed this at the ‘2024 North American Radiology Society (RSNA 2024)’ held in Chicago, USA.

The study was conducted by Sweden’s largest private hospital, Capio Saint G?ran Hospital. It compared patient screening data before AI adoption (July 2018 to March 2019) and after AI adoption (July 2023 to March 2024) for over 55,000 patients.

Before AI adoption, the cancer detection rate (CDR) per 1,000 screened patients was 4.8. After adoption, it rose to 5.5, a 15% increase. The recall rate for additional examinations after suspicious findings decreased from 2.8% to 2.5%, an 11% reduction.

Another Weakness of Human Doctors: ‘Cognitive Bias’

Human doctors also have the weakness of cognitive bias. Daniel Kahneman, a psychologist and Nobel laureate in Economics, explains in his book Thinking, Fast and Slow how human perception is riddled with errors.

He presented lung cancer patients and doctors with two options: surgery and radiation therapy. Surgery has a higher survival rate but carries a risk of immediate death, unlike radiation therapy.

When told the surgery success rate was 90%, 82% of patients chose surgery.

However, when told the mortality rate was 10%, which is the same probability expressed differently, only 54% chose surgery. Even though it was a life-or-death decision, people were influenced not by the probability itself but by how it was presented.

Not only patients but doctors behaved similarly. Many surgeons preferred to tell patients the surgery survival rate was 90% rather than the mortality rate of 10%.

The Bank of Korea released a report titled ‘AI and Labor Market Changes’ in July. Coincidentally, doctors were analyzed as highly likely to be replaced by AI. The AI exposure index used by the Bank of Korea measures how much of a job’s tasks can currently be performed by AI technology. General practitioners and Korean medicine doctors ranked within the top 1% in AI exposure.

Why AI Doctors Are Struggling: Jobs Are Not Just a Sum of Tasks but a ‘System’

Many companies have combined AI with healthcare, and this challenge has been widely praised as a "medical revolution." However, the reality was harsh. Getty Images Bank

Based on these research results and reports, one might think AI should have already replaced doctors. What is the reality?

Yes, they have not been replaced. On the contrary, some companies venturing into AI healthcare are struggling.

IBM’s Watson, launched in 2015, is a representative example. Watson is a cloud-based AI platform that, when inputting patient symptoms, searches through decades of clinical cases and hundreds of thousands of pages of professional materials to suggest the most suitable treatment. It was hailed as a ‘medical revolution’ and the ‘future of medicine.’

However, IBM’s AI healthcare business struggled to gain momentum.

In January 2022, IBM sold its healthcare AI division, Watson Health, to the US private equity firm Francisco Partners. According to IBM’s 2015 annual report, IBM invested over $15 billion (about 17.9 trillion KRW) in Watson-related businesses. More than $4 billion (about 4.8 trillion KRW) was spent acquiring medical imaging companies. An IBM insider said:

“We thought it would be easy... but it was extremely difficult.”

What hindered IBM’s grand plan? While there were issues like data errors, diagnostic errors, ethical problems, lack of legislation, and regulations, these were not the fundamental reasons. The core problem was the wrong answer to the question ‘What should AI replace?’ Medical AI has tried to replace ‘tasks.’ But tasks are not the same as jobs.

Generally, a job consists of dozens of tasks. A radiologist’s duties reportedly include over 30 tasks, including simple image interpretation. Imagining personalized medicine through AI may be easy, but realizing it requires changes in many personnel and organizational roles.

This includes direct communication and consultation with patients, collaborative discussions with other medical staff, clinical judgment on complex cases, participation in treatment planning, medical ethical decisions, legal responsibility and decision-making, medical record management and report writing, and training and mentoring junior medical staff.

In other words, a job is not merely a collection of tasks but a complex network of responsibilities and competencies. Even if AI outperforms humans in specific tasks, that does not immediately lead to replacing the entire job.

AI can replace simplified tasks, but it cannot replace the system that connects tasks to tasks and tasks to jobs. Photo by Getty Images Bank

When viewing jobs through AI, thinking in terms of ‘task units’ is an error that hinders innovation. Timothy Bresnahan, a professor at Stanford University, emphasizes the importance of ‘systemic thinking.’

“Dividing AI capabilities into task units will overlook the driving force behind the radical adoption of past new technologies, namely changes at the system level.”

The introduction of computers in the 1980s is an example. Companies that simply replaced typewriters with computers achieved only limited productivity gains, whereas those that redesigned entire work processes achieved much greater results.

IBM certainly had pioneering AI technology at the time. However, it failed to understand and innovate the entire healthcare system. Simply creating AI that analyzes medical data and diagnoses diseases could not revolutionize the complex healthcare system.

For AI to be successfully established in industrial sites, the meaning of jobs and systemic changes in organizations and companies must follow. Photo by Getty Images Bank

This mistake is repeated in other fields as well. Let’s look at call centers trying to adopt AI. Some companies reduce the number of agents simply because AI chatbots handle simple inquiries well. However, from Professor Bresnahan’s insight, this is a short-sighted approach. They miss the opportunity for systemic innovation that redefines the roles of AI and agents to create better customer experiences.

For example, AI can handle first-level responses, covering most simple inquiries. Additionally, data collected during this process can be used to better understand customer needs and build databases. Human agents can then focus on second-level responses, dealing with complex problem-solving and emotional support. This can also create opportunities to acquire new customers.

Companies need to learn from radiologists, IBM, and these cases as cautionary tales. AI adoption should not be merely about replacing specific tasks but an opportunity to rethink the entire organization’s work methods and value creation.

True innovation in the AI era begins not with ‘task replacement’ but with ‘system reconfiguration.’ This means moving beyond the question “What tasks can AI replace?” to the essential question, “How can we use AI to build better systems?”