Panel: A.I. Regulation Needed, But Long Way Off

Emma Pierson, Cornell Tech computer science professor at AAAS meeting, 3 Mar. 2023 (A. Kotok)

3 Mar. 2023. A panel of data scientists at a professional meeting says artificial intelligence needs more rigorous oversight, but many questions remain on who and how to do the regulating. The group discussed these issues today at a session of the annual meeting of American Association for Advancement of Science or AAAS in Washington, D.C.

The conference session, “Working to Address Unethical and Unregulated Artificial Intelligence” focused on A.I. applications where people can be adversely affected by the outcomes. Megan Price, executive director of Human Rights Data Analysis Group or HRDAG in San Francisco noted in some criminal justice algorithms, “data are messy and incomplete,” and not always representative, with missing data that can result in erroneous conclusions.

In the U.S., Price’s organization examines A.I. in the justice system and cited their work with the Invisible Institute in Chicago that investigated use of free text narrative reports of police violence. Price said summary statistics issued by the police varied widely from a reanalysis of the same narratives with algorithms trained by people from affected communities. In Mexico, said Price, HRDAG algorithms helped find of graves missing people from the country’s drug war.

Emma Pierson, computer science professor at Cornell Tech in New York, spoke about inequalities built into algorithms for diagnosing health problems in people, such as pain, which can be subjective and prone to error. In a paper published in Nature Medicine, Pierson and colleagues show underrepresented groups in the population report experiencing higher levels of pain, which are not explained, for example, by machine learning algorithms trained to examine knee X-rays of people with osteoarthritis. She noted that medical algorithms need to be trained to find physical variables missed by conventional diagnostics, and in partnership with affected communities.

Randomized clinical trials as a model

Joseph Gastwirth, professor of statistics at George Washington University in Washington, D.C., pointed out that algorithms could serve as a check on human assessments that affect people’s lives. Gastwirth cited real estate appraisers that routinely undervalue properties in minority communities. In these instances, says Gastwirth, A.I. would be a distinct improvement. “We can solve these problems,” he says, “if we work together with subject matter people.”

Many of the audience questions dealt with regulation and validation of A.I. Pierson suggested establishing testing processes, perhaps using randomized clinical trials as a model for assessing algorithms, although she acknowledged that model would probably fit medical applications more than A.I. in criminal justice. One audience participant who identified herself as an attorney representing death-row inmates questioned how algorithms could be validated where the outcomes have life-and-death implications.

The panel agreed regulation of A.I. is needed, and the sooner the better, since A.I. is racing ahead of attempts to regulate the technology, with government the most likely entity able to do the regulation. But how government should proceed remains an open question. Gastwirth says technology producers should do the initial testing, but then validated independently by agencies with the expertise needed, such as FDA.

Price, however, says those who write the algorithms may not have the resources or incentives to audit their own work. But all panelists agreed communities affected by A.I. outcomes should not bear the burden of regulating the technology.

More from Science & Enterprise:

• Infographic – Americans Uncomfortable With A.I. in Health
• Infographic – U.S. More Concerned Than Excited About A.I.
• Real-World Data Shown to Match Eye Disease Trial Results
• AI Analytics Boost Breast Cancer Assessment Diversity
• Company Formed to Boost Clinical Trial Success Odds

We designed Science & Enterprise for busy readers including investors, researchers, entrepreneurs, and students. Except for a narrow cookies and privacy strip for first-time visitors, we have no pop-ups blocking the entire page, nor distracting animated GIF graphics. If you want to subscribe for daily email alerts, you can do that here, or find the link in the upper left-hand corner of the desktop page. The site is free, with no paywall. But, of course, donations are gratefully accepted.

*     *     *