Visit our website
New America Cypbersecurity Initiative
New America Cypbersecurity Initiative
MIT Technology Review
MIT Technology Review
io9
io9
Techdirt
Techdirt
Knowledge@Wharton
Knowledge@Wharton
Bioscience Technology
Bioscience Technology
redOrbit
redOrbit
Technology & Marketing Law Blog
Technology & Marketing Law Blog
Popular Science Blog
Popular Science Blog
Pew Research Center
Pew Research Center
Genomics Law Report
Genomics Law Report
Science 2.0
Science 2.0
The Guardian Headquarters
The Guardian Headquarters
Genetic Literacy Project
Genetic Literacy Project
Disclaimer

Statements posted on this blog represent the views of individual authors and do not necessarily represent the views of the Center for Law Science & Innovation (which does not take positions on policy issues) or of the Sandra Day O'Connor College of Law or Arizona State University.

5th Annual Governance of Emerging Technologies Conference: Part 3

What follows is Part 3 of a series of select summaries of conference presentations, each prepared by 2L Jesse James, at LSI‘s Fifth Annual Conference on Governance of Emerging Technologies: Law, Policy and Ethics, held at the Sandra Day O’Connor College of Law, Beus Center for Law & Society in Phoenix, AZ on May 17-19, 2017.

Day 1: Keynote 1 

     Stuart Russsell

World-renowned computer scientist and artificial intelligence guru, Stuart Russell invigorated the audience with a talk about his article “Provably beneficial AI” in the book The Next Step: Exponential Life. He notes AI is a hot topic, and because of the existential risks involved, deserves nothing but cautious movement forward. He offers conceptual directives for AI to make it safe based entirely on uncertainty in a machine’s objectives. Russell contends, as long as humans have special status in an AI’s programming, and all their thinking revolves contextually about how to please humans, the AI should be safe.

The audience shifts in their seats a bit when Russell begins by saying the world “Go” champion is an AI.  The victorious software, AlphaGo, is Google’s Deep Learning bot that has played every great Go player in the world and beaten all.  Deep Learning revitalizes some concern about superintelligence, and Russell has put some thought into what programs must be in place to make sure AI is safe.

His premise is that, eventually, AI systems will make better (he admits a certain ambiguity to “better”) decisions than humans because they will be able to take into account much more information and look much farther into the future. The upside is that everything we have today is a product of intelligence, and possessing access to a significantly greater intelligence would be a step change in civilization and a possible new golden era. The downside, of course, is killer robots, and risks to employment — among other potential catastrophes.

Russell sees a potential “gorilla problem” as well – gorillas made something smarter than themselves but have nothing to show for it. But we should also feel honored, he says, quoting Alan Turing, “Even if we could keep the machines in a subservient position, for instance by turning off the power at strategic moments, we should, as a species, feel greatly humbled.”

Another problem involved with the creation of advanced AI systems is outlined in Russell’s quotation of Norbert Wiener, “we had better be quite sure that the purpose put into the machine is the purpose which we really desire.” Right now, our AI systems will do exactly what you ask them to, not think and say, “are you sure that’s what you want?” It’s incredibly difficult to forecast the future, and we don’t want to program an AI and end up with a King Midas situation of getting everything we’ve asked for.

Russell contends machines already have an inclination toward self preservation, which can potentially lead to problems. If a machine has a purpose, it necessarily has an appetite for self-preservation. That is, if one asks a machine to fetch coffee, it can’t do it if it’s turned off.  So, it will endeavor not to be turned off.  To prevent the machine from killing anyone who tries to turn it off, in its attempt to complete its task, Russell sees 2 problems – a misaligned objective, and the machine protecting itself from anyone who tries to interfere.

To solve these problems Russell and others set up the center for human-compatible AI to “reorient the general thrust of AI research toward provably beneficial systems” with three simple ideas on how to approach the problem. First, the AI’s only objective is to maximize the realization of human values (implicit preferences over complete lives). Second, the AI is initially uncertain about what those values are. Third, human behavior provides information about human values. To achieve this, the AI must be uncertain in its objectives, so observable human actions can give it further information about its objectives. Thus, humans have to actually come into the equation in the AI’s thinking, whereby humans (or “principles”) are given special status.

Russell differentiates between an Artificial General Intelligence (“AGI”), and a regular problem solver AI. For a regular problem solver AI, one assumes and defines a formal problem “F” that the machine can solve arbitrarily well. That machine is an “F”-solver, not an AGI. However, its program design may include subsystems of arbitrary intelligence, they just have to be connected, trained, and motivated in the right way. Recall the coffee-fetching robot: without subsystems of intelligence and uncertain objectives based on observable human behavior, when asked to fetch coffee and someone attempts to turn it off, the AI might disable the off switch and taser all the employees at Starbucks® to get the coffee. But with uncertain objectives based on observable human behavior, the AI will think “the human might switch me off, but only if I’m doing something wrong. I don’t know what wrong is, but I know I don’t want to do it. Therefore, I should let the human switch me off.

This does involve value alignment issues, Russell admits. Humans are nasty, irrational, inconsistent, weak willed, computationally limited, incredibly complex, heterogeneous, and may not have an objective in any meaningful sense. But the AI will not act like those it observes, it is purely altruistic and cares about everyone. It is learning to predict what people want, not learning to want it. And if someone wants others to suffer, the AI may have an ingrained sign of altruism. Weight of preference also needs to be given to the AI. Russell invites the audience to imagine a scenario in which a man is about to meet with an important client.  His AI assistant reminds him he promised his wife an anniversary dinner at the same time as the meeting. If too much weight is given to the individual man, the AI may come to the rescue and say, “I’ve managed to delay your client’s flight so you can have dinner with your wife.” Alternatively, if too little preference is given to the individual, the AI may pack up its bags and announce, “this one’s yours to deal with, see ya!”

Russell summarizes that this value misalignment is a potential risk, but believes if we go about it a particular way, certain design templates may support provably beneficial systems. Although he does not believe we are ready yet for standards or regulations, economic incentives may work in our favor.

Russell closes the presentation by reciting some unanswered open questions for further reflection. Can we change the way AI defines itself? For instance, a civil engineer says, “I design bridges,” not “I design bridges that don’t fall down.” That bit about not falling down is already implied. Will solutions to these near-term control problems scale up to the long-term control problem of Global Artificial Super Intelligences like SkyNet from the Terminator movies? What about James Bond villains that reprogram an AI or build one without these protections? What about long-term enfeeblement or the slow-boiling frog problem as seen in E.M. Forster’s The Machine Stops?