Category Archives: military robotics

Corporate Accountability

Jacbin image.jpg

This graphic appears at https://www.jacobinmag.com/2018/06/google-project-maven-military-tech-workers

On June 7th, Google CEO Sundar Pichai published a post on the company’s public blog site titled ‘AI at Google: our Principles.’ (Subsequently abbreviated to Our Principles.) The release of this statement was responsive in large measure to dissent from Google employees beginning early in the Fall of last year; while these debates are not addressed directly, their traces are evident in the subtext. The employee dissent focused on the company’s contracts with the US Department of Defense, particularly for work on its Algorithmic Warfare Cross Functional Team, also known as Project Maven. The controversy was receiving increasingly widespread attention in the press.

It is to the credit of Google workers that they have the courage and commitment to express their concerns. And it is to Google management’s credit that, unusually among major US corporations, it both encourages dissent and feels compelled to respond. I was involved in organizing a letter from researchers in support of Googlers and other tech workers, and in that capacity was gratified to hear Google announce that it would not renew the Project Maven contract next year. (Disclosure: I think US militarism is a global problem, perpetrating unaccountable violence while further jeopardizing the safety of US citizens.) In this post I want to take a step away from that particular issue, however, to do a closer reading of the principles that Pichai has set out. In doing so, I want to acknowledge Google’s leadership in creating a public statement of its principles for the development of technologies; a move that is also quite unprecedented, as far as I’m aware, for private corporations. And I want to emphasize that the critique that I set out here is not aimed at Google uniquely, but rather is meant to highlight matters of concern across the tech industry, as well as within wider discourses of technology development.

One question we might ask at the outset is why this statement of principles is framed in terms of AI, rather than software development more broadly. Pichai’s blog post opens with this sentence: “At its heart, AI is computer programming that learns and adapts.” Those who have been following this blog will be able to anticipate my problems with this statement, singularizing ‘AI’ as an agent with a ‘heart’ that engages in learning, and in that way contributing to its mystification. I would rephrase this along the lines of “AI is the cover term for a range of techniques for data analysis and processing, the relevant parameters of which can be adjusted according to either internally or externally generated feedback.” One could substitute “information technologies (IT)” or “software” for AI throughout the principles, moreover, and their sense would be the same.

Pichai continues: “It [AI] can’t solve every problem, but its potential to improve our lives is profound.” While this is a familiar (and some would argue innocent enough) premise, it’s always worth asking several questions in response: What’s the evidentiary basis for AI’s “profound potential”? Whose lives, more specifically, stand to be improved? And what other avenues for the enhancement of human well being might the potential of AI be compared to, both in terms of efficacy and the number of persons positively affected?

Regrettably, the opening paragraph closes with some product placement, as Pichai asserts that Google’s development of AI makes its products more useful, from “email that’s spam-free and easier to compose, to a digital assistant you can speak to naturally, to photos that pop the fun stuff out for you to enjoy,” with embedded links to associated promotional sites (removed here in order not to propagate the promotion). The subsequent paragraph then offers a list of non-commercial applications of Google’s data analytics, whose “clear benefits are why Google invests heavily in AI research and development.”

This promotional opening then segues to the preamble to the Principles, explaining that they are motivated by the recognition that “How AI is developed and used will have a significant impact on society for many years to come.” Readers familiar with the field of science and technology studies (STS) will know that the term ‘impact’ has been extensively critiqued within STS for its presupposition that technology is somehow outside of society to begin with. Like any technology, AI/IT does not originate elsewhere, like an asteroid, and then make contact. Rather, like Google, AI/IT is constituted from the start by relevant cultural, political, and economic imaginaries, investments, and interests. The challenge is to acknowledge the genealogies of technical systems and to take responsibility for ongoing, including critical, engagement with their consequences.

The preamble then closes with this proviso: “We acknowledge that this area is dynamic and evolving, and we will approach our work with humility, a commitment to internal and external engagement, and a willingness to adapt our approach as we learn over time.” Notwithstanding my difficulties in thinking of a precedent for humility in the case of Google (or any of the other Big Five), this is a welcome statement, particularly in its commitment to continuing to listen both to employees and to relevant voices beyond the company.

The principles themselves are framed as a set of objectives for the company’s AI applications, all of which are unarguable goods. These are: being socially beneficial, avoiding the creation or reinforcement of social bias, ensuring safety, providing accountability, protecting privacy, and upholding standards of scientific excellence. Taken together, Google’s technologies should “be available for uses that support these principles.” While there is much to commend here, some passages shouldn’t go by unremarked.

The principle, “Be built and tested for safety” closes with this sentence: “In appropriate cases, we will test AI technologies in constrained environments and monitor their operation after deployment.” What does this imply for the cases where this is not “appropriate,” that is, what would justify putting AI technologies into use in unconstrained environments, where their operations are more consequential but harder to monitor?

­The principle “Be accountable to people,” states “We will design AI systems that provide appropriate opportunities for feedback, relevant explanations, and appeal. Our AI technologies will be subject to appropriate human direction and control.” This is a key objective but how, realistically, will this promise be implemented? As worded, it implicitly acknowledges a series of complex and unsolved problems: the increasing opacity of algorithmic operations, the absence of due process for those who are adversely affected, and the increasing threat that automation will translate into autonomy, in the sense of technologies that operate in ways that matter without provision for human judgment or accountability. Similarly, for privacy design, Google promises to “give opportunity for notice and consent, encourage architectures with privacy safeguards, and provide appropriate transparency and control over the use of data.” Again we know that these are precisely the areas that have been demonstrated to be highly problematic with more conventional techniques; when and how will those longstanding, and intensifying, problems be fully acknowledged and addressed?

The statement closes, admirably, with an explicit list of applications that Google will not pursue. The first item, however, includes a rather curious set of qualifications:

  1. Technologies that cause or are likely to cause overall harm. Where there is a material risk of harm, we will proceed only where we believe that the benefits substantially outweigh the risks, and will incorporate appropriate safety constraints.

What are the qualifiers “overall” doing here, or “material”? What will be the basis for the belief that “the benefits substantially outweigh the risks,” and who will adjudicate that?

There is a welcome commitment not to participate in the development of

2.  Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people.

As the Project Maven example illustrates, the line between a weapon and a weapon system can be a tricky one to draw. Again from STS we know that technologies are not discrete entities; their purposes and implementations need to be assessed in the context of the more extended sociotechnical systems of which they’re part.

And finally, Google pledges not to develop:

  1. Technologies that gather or use information for surveillance violating internationally accepted norms.
  2. Technologies whose purpose contravenes widely accepted principles of international law and human rights.

Again, these commitments are laudable; however we know that the normative and legal frameworks governing surveillance and human rights are highly contested and frequently violated. This means that adherence to these principles will require working with relevant NGOs (for example, the International Committee of the Red Cross, Human Rights Watch), continuing to monitor the application of Google’s technologies, and welcoming challenges based on evidence for uses that violate the principles.

A coda to this list ensures Google’s commitment to work with “governments and the military in many other areas,” under the pretense that this can be restricted to operations that “keep [LS: read US] service members and civilians safe.” This odd pairing of governments, in the plural, and the military singular might raise further questions regarding the obligations of global companies like Google and the other Big Five information technology companies. What if it were to read “governments and militaries in many other areas”? What does work either with one nation’s military, or many, imply for Google’s commitment to users and customers around the world?

The statement closes with:

We believe these principles are the right foundation for our company and the future development of AI. This approach is consistent with the values laid out in our original Founders’ Letter back in 2004. There we made clear our intention to take a long-term perspective, even if it means making short-term tradeoffs. We said it then, and we believe it now.

This passage is presumably responsive to media reports of changes to Google’s Code of Conduct, from “Don’t Be Evil” (highly lauded but actually setting quite a low bar), to Alphabet’s “Do the Right Thing.” This familiar injunction is also a famously vacuous one, in the absence of the requisite bodies for deliberation, appeal, and redress.

The overriding question for all of these principles, in the end, concerns the processes through which their meaning and adherence to them will be adjudicated. It’s here that Google’s own status as a private corporation, but one now a giant operating in the context of wider economic and political orders, needs to be brought forward from the subtext and subject to more explicit debate. While Google can rightfully claim some leadership among the Big Five in being explicit about its guiding principles and areas that it will not pursue, this is only because the standards are so abysmally low. We should demand a lot more from companies as large as Google, which control such disproportionate amounts of the world’s wealth, and yet operate largely outside the realm of democratic or public accountability.

Unpriming the pump: Remystifications of AI at the UN’s Convention on Certain Conventional Weapons

Related image

In the lead up to the next meeting of the CCW’s Group of Governmental Experts at the United Nations April 9-13th in Geneva, the UN’s Institute for Disarmament Research has issued a briefing paper titled The Weaponization of Increasingly Autonomous Technologies: Artificial Intelligence.  Designated a primer for CCW delegates, the paper lists no authors, but a special acknowledgement to Paul Scharre, Senior Fellow and Director of the Technology and National Security Program at the Center for a New American Security, suggests that the viewpoints of the Washington, D.C.-based CNAS are well represented.

Surprisingly for a document positioning itself as “an introductory primer for non-technical audiences on the current state of AI and machine learning, designed to support the international discussions on the weaponization of increasingly autonomous technologies” (pp. 1-2), the paper opens with a series of assertions regarding “rapid advances” in the field of AI. The evidence offered is the case of Google/Alphabet affiliate Deep Mind’s AlphaGo Zero, announced in December 2017 (“only a few weeks after the November 2017 GGE”) as having achieved better-than-human competency at (simulations of) the game of Go:

Although AlphaGo Zero does not have direct military applications, it suggests that current AI technology can be used to solve narrowly defined problems provided that there is a clear goal, the environment is sufficiently constrained, and interactions can be simulated so that computers can learn over time (p.1).

The requirements listed – a clear (read computationally specifiable) goal, within a constrained environment that can be effectively simulated – might be underscored as cautionary qualifications on claims for AI’s applicability to military operations. The tone of these opening paragraphs suggests, however, that these developments are game-changers for the GGE debate.

The paper’s first section, titled ‘What is artificial intelligence,’ opens with the tautological statement that “Artificial intelligence is the field of study devoted to making machines intelligent” (p. 2). A more demystifying description might say, for example, that AI is the field of study devoted to developing computational technologies that automate aspects of human activity conventionally understood to require intelligence. While the authors observe that as systems become more established they shift from characterizations of “intelligence” to more mundane designations like “automation” or “computation,” they suggest that rather than the result of demystification this is itself somehow an effect of the field’s advancement. One implication of this logic is that the ever-receding horizon of machine intelligence should be understood not as a marker of the technology’s limits, but of its success.

We begin to get a more concrete sense of the field in the section titled ‘Machine learning,’ which outlines the latter’s various forms. Even here, however, issues central to the deliberations of the GGE are passed over. For example, in the statement that “[r]ather than follow a proscribed [sic] set of if–then rules for how to behave in a given situation, learning machines are given a goal to optimize – for example, winning at the game of chess” (p. 2) the example is not chosen at random, but rather is illustrative of the unstated requirement that the ‘goal’ be computationally specifiable. The authors do helpfully explain that “[s]upervised learning is a machine learning technique that makes use of labelled training data” (my emphasis, p. 3), but the contrast with “unsupervised learning,” or “learning from unlabelled data based on the identification of patterns” fails to emphasize the role of the human in assessing the relevance and significance of patterns identified. In the case of reinforcement learning “in which an agent learns by interacting with its environment,” the (unmarked) examples are again from strategy games in which, implicitly, the range of agent/environment interactions are sufficiently constrained. And finally, the section on ‘Deep learning’ helpfully emphasizes that so called neural networks rely either on very large data sets and extensive labours of human classification (for example, the labeling of images to enable their ‘recognition’), or on domains amenable to the generation of synthetic ‘data’ through simulation (for example, in the case of strategy games like Go). Progress in AI, in sum, has been tied to growth in the availability of large data sets and associated computational power, along with increasingly sophisticated algorithms within highly constrained domains of application.

Yet in spite of these qualifications, the concluding sections of the paper return to the prospects for increasing machine autonomy:

Intelligence is a system’s ability to determine the best course of action to achieve its goals. Autonomy is the freedom a system has in accomplishing its goals. Greater autonomy means more freedom, either in the form of undertaking more tasks, with less supervision, for longer periods in space and time, or in more complex environments … Intelligence is related to autonomy in that more intelligent systems are capable of deciding the best course of action for more difficult tasks in more complex environments. This means that more intelligent systems could be granted more autonomy and would be capable of successfully accomplishing their goals (p. 5, original emphasis).

The logical leap exemplified in this passage’s closing sentence is at the crux of the debate regarding lethal autonomous weapon systems. The authors of the primer concede that “all AI systems in existence today fall under the broad category of “narrow AI”. This means that their intelligence is limited to a single task or domain of knowledge” (p. 5). They acknowledge as well that “many advance [sic] AI and machine learning methods suffer from problems of predictability, explainability, verifiability, and reliability” (p. 8). These are precisely the concerns that have been consistently voiced, over the past five meetings of the CCW, by those states and civil society organizations calling for a ban on autonomous weapon systems. And yet the primer takes us back, once again, to a starting point premised on general claims for the field of AI’s “rapid advance,” rather than careful articulation of its limits. Is it not the latter that are most relevant to the questions that the GGE is convened to consider?

The UNIDIR primer comes at the same time that the United States has issued a new position paper in advance of the CCW titled ‘Humanitarian benefits of emerging technologies in the area of lethal autonomous weapon systems’ (CCW/GGE.1/2018/WP.4). While the US has taken a cautionary position in relation to lethal autonomous weapon systems in past meetings, asserting the efficacy of already-existing weapons reviews to address the concerns raised by other member states and civil society groups, it now appears to be moving in the direction of active promotion of LAWS on the grounds of promised increases in precision and greater accuracy of targeting, with associated limits on unintended civilian casualties – promises that have been extensively critiqued at previous CCW meetings. Taken together, the UNIDIR primer and the US working paper suggest that, rather than moving forward from the debates of the past five years, the 2018 meetings of the CCW will require renewed efforts to articulate the limits of AI, and their relevance to the CCW’s charter to enact Prohibitions or Restrictions on the Use of Certain Conventional Weapons Which May Be Deemed to Be Excessively Injurious or to Have Indiscriminate Effects.

Reality Bites

 

big dog trial

LS3 test during Rim of the Pacific Exercise, July 2014

A pack of international news outlets over the past few days have reported the abandonment by the US Department of Defence of Boston Dynamic’s Legged Squad Support System or LS3 (aka ‘Big Dog’) and its offspring (see Don’t kick the Dog). After five years and USD $42 million in investment, what was promised to be a best in breed warfighting companion stumbled over a mundane but apparently intractable problem – noise. Powered by a gas (petrol) motor likened to a lawnmower in sound, the robot’s capacity for carrying heavy loads (400 lbs or 181.4kg), and its much celebrated ability to navigate rough terrain and right itself after falling (or be easily assisted in doing so), in the end were not enough to make up for the fact that, in the assessment of the US Marines who tested the robot, the LS3 was simply ‘too loud’ (BBC News 30 January 2015). The trial’s inescapable conclusion was that the noise would reveal a unit’s presence and position, bringing more danger than aid to the U.S. warfighters that it was deployed to support.

A second concern contributing to the DoD’s decision was the question of the machine’s maintenance and repair. Long ignored in narratives about technological progress, the place of essential practices of inventive maintenance and repair has recently become a central topic in social studies of science and technology (see Steven J. Jackson, “Rethinking Repair,” in Tarleton Gillespie, Pablo Boczkowski, and Kirsten Foot, eds. Media Technologies: Essays on Communication, Materiality and Society. MIT Press: Cambridge MA, 2014.). These studies are part of a wider project of recognizing the myriad forms of invisible labour that are essential conditions for keeping machines working – one of the enduring continuities in the history of technology.

The LS3 trials were run by the Marine’s Warfighting Lab, most recently at Kahuku Training Area in Hawaii during the Rim of the Pacific exercise in July of 2014. Kyle Olson, spokesperson for the Lab, reported that seeing the robot’s potential was challenging “because of the limitations of the robot itself.” This phrasing is noteworthy, as the robot itself – the actual material technology – interrupts the progressive elaboration of the promise that keeps investment in place. According to the Guardian report (30 December 2015) both ‘Big Dog’ and ‘Spot,’ an electrically powered and therefore quieter but significantly smaller prototype, are now in storage, with no future experiments planned.

The cessation of the DoD investment will presumably come as a relief to Google, which acquired Boston Dynamics in 2013, saying at the time that it planned to move away from the military contracts that it inherited with the acquisition.  Boston Dynamics will now, we can assume, turn its prodigious ingenuity in electrical and mechanical engineering to other tasks of automation, most obviously in manufacturing. The automation of industrial labour has, somewhat ironically given its status as the original site for robotics, recently been proclaimed to be robotics’ next frontier. While both the BBC and Guardian offer links to a 2013 story about the great plans that accompanied Google’s investments in robotics, more recent reports characterize the status of the initiative (internally named ‘Replicant’) as “in flux,” and its goal of producing a consumer robot by 2020 as in question (Business Insider November 8, 2015). This follows the departure of former Google VP Andy Rubin in 2014 (to launch his own company with the extraordinary name ‘Playground Global’), just a year after he was hailed as the great visionary leader who would turn Google’s much celebrated acquisition of a suite of robotics companies into a unified effort. Having joined Google in 2005, when the latter acquired his smartphone company Android, Rubin was assigned to the leadership of Google’s robotics division by co-founder Larry Page. According to Business Insider’s Jillian D’Onfro, Page

had a broad vision of creating general-purpose bots that could cook, take care of the elderly, or build other machines, but the actual specifics of Replicant’s efforts were all entrusted to Rubin. Rubin has said that Page gave him a free hand to run the robotics effort as he wanted, and the company spent an estimated $50 million to $90 million on eight wide-ranging acquisitions before the end of 2013.

The unifying vision apparently left with Rubin, who has yet to be replaced. D’Onfro continues:

One former high-ranking Google executive says the robot group is a “mess that hasn’t been cleaned up yet.” The robot group is a collection of individual companies “who didn’t know or care about each other, who were all in research in different areas,” the person says. “I would never want that job.”

So another reality that ‘bites back’ is added to those that make up the robot itself; that is, the alignment of the humans engaged in its creation. Meanwhile, Boston Dynamics’ attempt to position itself on the entertainment side of the military-entertainment complex this holiday season was met less with amusement than alarm, as media coverage characterized it variously as ‘creepy’ and ‘nightmarish.’

synthxmas-590x330

Resistance, it seems, is not entirely futile.