The Challenge of AI Ethics

APRIL 16, 2019

In his influential 2014 book Superintelligence, the philosopher Nick Bostrom argued that the prospect of advanced AI would require us to do "philosophy with a deadline." While Bostrom's concern was redirecting academic resources toward the mitigation of existential risk, his line applies equally to the increasingly fraught near-term debate about AI ethics. The prospect of instantiating our values in the unforgiving language of computation has given new urgency to longstanding arguments about justice, rights, and welfare.

Google's recent ill-fated attempt to convene an external AI ethics council exemplifies many of the difficulties of the field. Setting aside the ample technical challenge of getting AI to do what we want, there remains considerable debate over what and whose values deserve a hearing in the AI ethics conversation. While nearly all agree on the need for a diversity of voices, it's unclear how far this should extend to "diversity of thought." The appointment of Kay Coles James, president of the Heritage Foundation, to the Google council prompted widespread backlash among Googlers, more than 2500 of whom signed a petition claiming her views had no place in the discussion. This along with other controversies caused Google to disband the council shortly after its establishment.

The controversy has not been confined to industry, either. The EU Commission recently put out a list of guidelines for trustworthy AI, the product of a "High-Level Expert Group" that has been incorporating public feedback since December of last year. Upon its release, the philosopher Thomas Metzinger -- himself a member of the group -- called the guidelines "luke-warm, short-sighted and deliberately vague." Metzinger claims that the very concept of "trustworthy AI" is a misnomer and marketing ploy: "Machines are not trustworthy;" he writes, "only humans can be trustworthy (or untrustworthy)." His initial directive, upon joining the panel, was to establish certain non-negotiable "Red Lines" to constrain the use of AI in Europe. But this effort was quickly nixed: the non-negotiables were replaced with "critical concerns," and no talk of "Red Lines" remains in the final document.

Another of Metzinger's complaints is that the 52-person group contained only four ethicists and its makeup was heavily slanted toward industry. As Metzinger admits, interfacing with industry is essential, especially as the bulk of AI engineering talent drains from academia into major tech companies. The questions of AI ethics concern engineers as much as academics, and span many disciplines beyond moral philosophy. But it has also become clear, even to industry members, that the deployment of advanced AI can't be guided by industry alone. OpenAI, for example, was organized as a nonprofit, in part to preserve its commitment to avoiding an AI arms race. And DeepMind, upon being acquired, put in place an agreement that would hand over control of any AGI they develop to an Ethics Board, preventing a unilateral takeover of the technology by Google.

In response to this ongoing discussion (and the many misconceptions therein), the Princeton Center for Information Technology Policy released a useful list of "Seven Traps" to avoid when thinking about AI ethics; among them: reductionism, relativism, and oversimplification. Reading the document, one is struck by how many of the same "traps" beset our public discourse more generally. Whether AI will be a clarifying force remains to be seen. But new quandaries, and new deadlines, demand that we continue the conversation.

Relevant: our interview with Iyad Rahwan on the social dilemmas of AI.