Artificial Intelligence: Governance, Taxation, Ethics, Ground Rules, and Its Largest Question

Over the past few months the subject of human-replacing technology articles has shifted.  From constant progress reports, autonomous vehicle news is down to a trickle, but artificial intelligence (AI) issues are still drawing reporting.  Here are four such stories, all from The New York Times.

The first, “How Do You Govern Machines That Can Learn?  Policymakers Are Trying to Figure That Out,” by Steve Lohr on January 20^th, wasn’t specific.  It reminded us, though, that “today’s machine-learning systems are so complex, digesting so much data, that explaining how they make decisions may be impossible” (italics Lohr’s).  I still believe that AI is only algorithmic, but it is now developing its own computational procedures often already too large to explain.  As to whether we need confidence in systems with methods more accurate than ours, “an M.I.T. computer scientist and breast cancer survivor” said that “you have to use” machine-generated algorithms predicting that disease if they are objectively best.  Yet, as we know from recent driverless-car attitudes, not all will consent to that.  Lohr also discussed two situations which he and everyone else seems to conflate:  poorer AI recognition of female and nonwhite faces, which is a technical issue requiring more work, and how to use controversial but correct data.

Next was Eduardo Porter’s February 24^th opinion-section “Don’t Fight the Robots.  Tax Them.”  Important issues he touched on were “how do you even define a robot to tax it?,” that before applying levies on such things we should first withdraw accelerated depreciation and other tax subsidies, and that reduced numbers of workers pay less income tax.  His suggestions included robot-owning businesses forfeiting taxes formerly paid by laid-off workers (good, as it assigns cost to the cost-causer), and a per-robot tax (OK, if we agree on what robots are).  I think we would do better to charge income tax on a sliding scale with companies with more full-time equivalent jobs paying less, which, given Porter’s idea of taxing “the ratio of a company’s profit to its employee compensation,” he almost proposed himself.

The last two were written by Cade Metz and published March 1^st.  The content of “Is Ethical A.I. Even Possible?” didn’t support that headline, but focused on two concerns, of facial recognition shortcomings as above and the growing unwillingness of AI researchers to contribute to autonomous weapons systems.  As Metz mentioned, the AI Now Institute, per its website “an interdisciplinary research center dedicated to understanding the social implications of artificial intelligence,” has been formed at New York University.  AI can certainly be ethical, but we will not all agree on what is right to do with it and what is not.

Finally, Metz’s “Seeking Ground Rules for A.I.” proposed ten overarching principles in the field.  They were transparency (in design, intention, and use of the technology), disclosure (to users), privacy (allowing users to refuse to have their data collected), diversity (of the development teams, presumably in race and sex), bias (in input data), trust (self-regulation), accountability (“a common set of standards”), collective governance, regulation, and complementarity (to limit AI as something for people to use instead of something to replace them).  A good start, and may, or may not, go a long way without major changes. 

Beyond all of these, we have a query to which AI will force an answer.  It is not a pleasant one, but we must think about it.  As Lohr almost stated, blacks, whites, men, women, gays, and straights, to name the most common but hardly all identity groups, do not have identical behavioral compositions.  As the systems determine differences between sexes and races, they will use them to identify criminal suspects, recommend hiring or not hiring, accept or refuse mortgage and other loan requests, determine optimal housing, and make or contribute to an almost infinite set of other large life-affecting decisions.  When algorithms are assembled using contextless data, it is inevitable that many will incorporate these factors.  Even if these six categories and more like them were expressly blocked from consideration, proxies such as geographical location would bring them right back in.  So here is the question:  What do we do when the truth is racist, sexist, homophobic, or heterophobic?  The answer we develop will mean more for the future of AI than any further technical progress.