Matthew Hutson
The New Yorker
Originally published 15 Feb 21
Here is an excerpt:
Many kinds of researchers—biologists, psychologists, anthropologists, and so on—encounter checkpoints at which they are asked about the ethics of their research. This doesn’t happen as much in computer science. Funding agencies might inquire about a project’s potential applications, but not its risks. University research that involves human subjects is typically scrutinized by an I.R.B., but most computer science doesn’t rely on people in the same way. In any case, the Department of Health and Human Services explicitly asks I.R.B.s not to evaluate the “possible long-range effects of applying knowledge gained in the research,” lest approval processes get bogged down in political debate. At journals, peer reviewers are expected to look out for methodological issues, such as plagiarism and conflicts of interest; they haven’t traditionally been called upon to consider how a new invention might rend the social fabric.
A few years ago, a number of A.I.-research organizations began to develop systems for addressing ethical impact. The Association for Computing Machinery’s Special Interest Group on Computer-Human Interaction (sigchi) is, by virtue of its focus, already committed to thinking about the role that technology plays in people’s lives; in 2016, it launched a small working group that grew into a research-ethics committee. The committee offers to review papers submitted to sigchi conferences, at the request of program chairs. In 2019, it received ten inquiries, mostly addressing research methods: How much should crowd-workers be paid? Is it O.K. to use data sets that are released when Web sites are hacked? By the next year, though, it was hearing from researchers with broader concerns. “Increasingly, we do see, especially in the A.I. space, more and more questions of, Should this kind of research even be a thing?” Katie Shilton, an information scientist at the University of Maryland and the chair of the committee, told me.
Shilton explained that questions about possible impacts tend to fall into one of four categories. First, she said, “there are the kinds of A.I. that could easily be weaponized against populations”—facial recognition, location tracking, surveillance, and so on. Second, there are technologies, such as Speech2Face, that may “harden people into categories that don’t fit well,” such as gender or sexual orientation. Third, there is automated-weapons research. And fourth, there are tools “to create alternate sets of reality”—fake news, voices, or images.