COVID-19 has spurred considerable news coverage on the ways in which artificial intelligence (AI) can combat the spread of the pandemic. Like many tools, AI has a role to play, but its effect on the outbreak is probably smaller than many believe; although, this could change in the future. This is, at least, what Brookings Institutions asserts in a recent report exposing a “healthy skepticism” towards artificial intelligence, citing developments to help curb the coronavirus as an example.
What do the experts say? Regardless of the topic that floods all the media today, COVID-19, artificial intelligence is only useful when those who know about this technology apply it judiciously. Despite all the talk about algorithms and big data, deciding what to predict and how to frame those predictions is often the most challenging aspect of applying AI. “Effectively predicting a poorly defined problem is worse than doing nothing at all,” they say from the Brooking Institution. Similarly, expertise in the field is always required to know if the models will continue to work in the future, or if they will be accurate in different population groups.
AI needs a lot of data
To learn how to do anything, AI needs tons of prior data with known results. Alternatively, a comprehensive computational simulation can be used to train the model. Without one of these two approaches, the AI cannot do anything. This explains why artificial intelligence alone cannot predict the spread of new pandemics. “There is no database of outbreaks prior to COVID-19.” Therefore, they say at the Brookings Institution, “taking a skeptical approach to artificial intelligence is critical to considering whether a company spent the time and money well to build a broad data set and effectively learn the task at hand.”
Sadly, not everyone is taking the skeptical path. “Some tech outlets, for example, have regurgitated Baidu’s claims that AI can be used with infrared thermal imaging to see how fever is a symptom of COVID-19. While Athena Security, which sells video analytics software, claimed that it adapted its artificial intelligence system to detect fever from thermal imaging data. To even attempt what Baidu and Athena Security say, the companies would have to collect extensive data from thermal images of people while simultaneously taking their temperature with a conventional thermometer.
In addition to obtaining a diverse sample in age, gender, size … this would also require that many of these people actually have a fever, an outcome that they are trying to predict. While there are other potential ways to obtain pre-existing data sets, questioning data sources is always a meaningful way to assess the feasibility of an AI system.
Should we be wary of the precision of AI?
The Alibaba company claims that it can use AI in imaging to diagnose COVID-19, and that it can offer this diagnostic software to European countries for free. However, the diagnosis of COVID-19 is made through a process called polymerase chain reaction (PCR) , which requires specialized equipment and a diagnosis that lasts several hours with a sensitivity and effectiveness close to 100%. Alibaba says its model is much faster and is 96% accurate. However, says the Brookings Institution in the report, “it is not clear that this accuracy percentage is reliable.” And it also says that “96% accuracy is suspiciously high for any machine learning problem.”
If not handled carefully, an AI algorithm will do its best to find patterns in the data that are associated with the outcome it is trying to predict. However, these patterns can be totally absurd and only seem to work during development. In fact, an inflated precision number can be a major sign that an AI model will not be effective. For Alibaba to claim that their model works so well without warning or self-criticism is suspicious. “
The AI bias
High-risk artificial intelligence applications, Brookings Institutions also says, generally require a prediction that is not only accurate, but also allows for human intervention. With thermal imaging for fever detection, systems could be used to block entry to airports, supermarkets, pharmacies, and public spaces. But the evidence shows that up to 90% of people marked with thermal images can be false positives. In an environment where feverish people know they are supposed to stay home, this ratio could be much higher. Therefore, while preventing people with fever (and potentially COVID-19) from allowing community transmission is a significant goal, there must be a willingness to establish checkpoints and a confirmatory test, or risk limiting it to the population significantly.
In new approaches to using AI in high-risk circumstances, the report concludes, bias should be a serious concern: In artificial intelligence models, it results in biased estimates that do not take into account factors such as gender, ethnic group or functional diversity. In turn, this often leads to discriminatory results, as AI models are often considered objective and neutral. While scientific research has raised awareness of many cases of AI bias, it is important to realize that the bias is more systemic than anecdotal.
Especially for models that involve humans, there are so many potential sources of bias that they cannot be ruled out without investigation. If an AI model does not have documented and evaluated biases, that should heighten a skeptic’s suspicion that they remain hidden, unresolved, and pernicious.
The bright future of AI
While the Brookings Institutions report takes a deliberately skeptical perspective on artificial intelligence, its future impact on many of these applications is bright. For example, while the diagnosis of COVID-19 with CT scans is of questionable value at this time, the impact that AI is having on medical imaging is substantial. Emerging applications can assess the malignancy of tissue abnormalities, study skeletal structures, and reduce the need for invasive biopsies. Other apps show great promise, although it is too early to tell if they will have a significant impact on this pandemic. “For example, drugs designed by artificial intelligence are recently beginning human trials. And using AI to summarize thousands of research papers can also accelerate medical discoveries relevant to COVID-19. “
AI is a widely applicable technology, but its advantages must be covered in a realistic understanding of its limitations, Brookings says. To that end, the report’s aim is not to broadly disparage the contributions AI can make, but to encourage a critical and insightful eye for the specific circumstances where AI can be meaningful.
