The Campaign to Stop Killer Robots has called on the UN to ban the development and use of autonomous weapons:
those that can identify, track and attack targets without meaningful human oversight. On Monday, the group released a sensationalist video, supported by some prominent artificial intelligence researchers, depicting a dystopian future in which such machines run wild.
I am gratified that my colleagues are volunteering their efforts to ensure beneficial uses of artificial intelligence (AI) technology. But I am unconvinced of the effectiveness of the campaign beyond a symbolic gesture. Even though I identify myself strongly as a pacifist, I have reservations about signing up to the proposed ban. I am not alone in this predicament.
Computer says no: why making AIs fair, accountable and transparent is crucial
Apart from the difficulty of pinning down exactly what the ban entails for states that want to follow it– is the ban against autonomy or intelligence?– I wonder about the ban’s ability to deter misuse by rogue state or non-state actors. To the extent that bans on conventional and nuclear weapons have been effective, it is because of the significant natural barriers to entry: the raw materials and equipment needed to make those weapons are hard to obtain, and responsible states can control them to a significant extent by fiat and sanctions. In contrast, AI technology, which ostensibly enables the kind of weapons that this ban is aimed at, is already quite open, and some may argue, admirably so. Misuses of it can thus be as hard to control by fiat and bans– as with cyber warfare, for example.
Consider the hypothetical “killer drones” depicted in the video accompanying the Guardian’s article on the call for the ban. Even today, the face recognition technology supposedly needed by such drones can be easily constructed by anyone with access to the internet: several near-state-of-the-art “pre-trained networks” are available open source. Things will only become easier as we make further technical advances.
Given these significantly lower barriers to entry, even if the UN and some constituent states agreed to a ban, it is far from clear that it would stop other rogue state and non-state actors from procuring and deploying such technology for malicious purposes. This would render such bans at best a pyrrhic victory for the proponents of peace, and at worst entail the ironic and unintended effect of tying the hands of the “good actors” behind their backs, while doing little to stop the bad ones.
Ban on killer robots urgently needed, say scientists
It’s time for some messy, democratic discussions about the future of AI
Jack Stilgoe and Andrew Maynard
Given these concerns about the effectiveness of blanket bans, I believe that AI researchers should instead be thinking of more proactive technical solutions to ameliorate potential misuses of AI technologies. As one small example of this alternate strategy, we held a workshop at Arizona State University in early March 2017 titled Challenges of AI: Envisioning and Addressing Adverse Outcomes. The workshop was attended by many leading scientists, technologists and ethicists, and had the aim of coming up with defensive responses for a variety of potential misuses of AI technology, including lethal autonomous weapons.
One recurrent theme of the workshop was using AI technology itself as a defence against the adverse/malicious uses of AI. This could include research into so-called “guardian AI systems” that can provide monitoring and defensive responses. Even if such efforts don’t succeed in completely containing the adverse effects, they could at least better inform the public policy on these issues.
To reiterate, I consider myself a pacifist, and have always supported efforts to control arms and curb wars. If I believed that the proposed ban would be effective and not merely symbolic, and that this campaign would inform rather than inflame the public, I would have gladly supported it.
Disclaimer: In the interests of full disclosure, let me state that some of my basic research (on human-aware AI) is supported by US Department of Defense funding agencies (eg Office of Naval Research). However, my research funding sources have no impact on my personal views, and defence funding agencies in the US support a wide spectrum of basic research, including that by researchers involved in the ban campaign.
Subbarao Kambhampati is a professor of computer science at Arizona State University, and the president of the Association for the Advancement of Artificial Intelligence.
As an AI researcher, I am also disturbed by the sensationalisation of the whole issue through dystopian– if high production value– videos such as the one reported in the Guardian article. Using a “Daisy Girl”- style campaign ads designed to stoke public fears about AI technologies seems to me to be more an exercise at inflaming rather than informing public opinion.