The Limits of AI Revealed by COVID-19... No "Useful Data" Available [Tech Talk by Im Ju-hyung]

AI Algorithms for COVID-19 Diagnosis Mostly Flawed
UK and Netherlands Research Teams Analyze Around 600 AI Tools
Due to Training on 'Incorrect Data'
Researchers: "Incorrect Data Renders AI Useless"
"Using Unprepared AI Could Be Riskier"

A customer service robot installed at Buk-gu Office in Busan last year. / Photo by Yonhap News

[Asia Economy Reporter Lim Juhyung] From stock investment to logistics delivery, it is unimaginable today to live in a world without artificial intelligence (AI). AI helps our lives by learning from massive amounts of data to recommend stocks likely to rise in price or predict parcel delivery demand.

However, despite AI’s seemingly omnipotent capabilities, it has struggled notably in responding to COVID-19. For example, hundreds of companies worldwide rushed to develop AI algorithms to detect COVID-19 virus infection, but successful cases among them are few and far between. Why has AI, known as one of the future innovative industries, become a burden specifically in the fight against COVID-19?

On June 31, the technology magazine Technology Review, published by the Massachusetts Institute of Technology (MIT), reported that “hundreds of AI models developed worldwide failed to show effectiveness in actual clinical settings.”

According to the publication, two research teams from the UK and the Netherlands recently reviewed the usefulness of over 600 COVID-19 diagnostic AI algorithms developed last year. These algorithms determine COVID-19 infection by integrating medical information such as patients’ symptoms, physical changes, X-ray, and chest computed tomography (CT) images.

Last year, as COVID-19 reached pandemic levels worldwide, many countries faced shortages of testing equipment and reagents, causing difficulties in patient diagnosis. AI companies attempted to respond to this crisis by developing diagnostic algorithms.

Medical staff collecting COVID-19 specimens / Photo by Yonhap News

However, the results were shocking when the lid was opened. The two research teams found that among more than 600 AI algorithms, only about two models demonstrated accuracy sufficient to move to clinical stages.

These research results were published in British scientific journals such as the British Medical Journal and Nature. The researchers told Technology Review, “It was truly shocking,” adding, “We had concerns that AI might be useless, but the experimental results far exceeded the level of worry I had.”

Why did hundreds of AI algorithms fail to determine patients’ COVID-19 infection status? The researchers pointed to the problem of “faulty data.”

Typically, modern machine learning AI improves accuracy by repeatedly learning from vast amounts of data. In other words, AI must be trained by injecting data appropriate to its purpose.

The problem was that during the COVID-19 pandemic, too much related data was generated on the internet, and a significant portion of it was “faulty information” that had not been curated by experts. Since such data were included in the datasets AI learned from, the AI itself became faulty.

Dr. Derek Drix, who led the UK research team, said, “AI researchers are making a fundamental mistake. They are using inaccurate data to train and test AI tools,” warning, “Because of incorrect data, AI becomes useless, yet many companies focus only on marketing terms and encourage the use of many AI tools that are not ready. This trend could actually put patients at risk.”

Artificial intelligence trained with faulty data is likely to perform poorly, making incorrect judgments and failing to demonstrate proper capabilities. / Photo by Yonhap News

“Faulty AI” trained on incorrect data can cause serious medical and ethical problems. It risks misdiagnosis that endangers patients’ lives and can acquire racially biased prejudices.

This risk is also a major obstacle to AI commercialization. According to the American financial media Wall Street Journal (WSJ) in February, IT company IBM is considering selling its medical AI “Watson,” which was developed with trillions of won invested, because Watson’s diagnostic accuracy is not high, causing American hospitals to hesitate in adopting it.

So, what is the way to make AI useful?

Dr. Drix argues that a breakthrough can be made through close collaboration between industry experts with professional knowledge and AI engineers. Experts refine high-quality data, and AI is trained based on that dataset to become “smart AI.”

Dr. Drix also emphasizes the importance of enhancing transparency in AI research. He said, “AI researchers should share how they trained their models with others,” adding, “So that other researchers can test and develop based on that.” Knowing how AI learned data makes it easier to identify problems and find areas for improvement.

Although AI is an advanced industry that will determine the direction of the Fourth Industrial Revolution, ultimately, the secret to creating excellent AI lies not in robots or computers but in the people who handle them.

This content was produced with the assistance of AI translation services.