Throughout the 1960s, ’70s, and ‘80s, artificial intelligence researchers seemed to be fatally vexed by the problem of the future’s looming uncertainty. This was because the “top-down” AI approaches dominant at the time were deterministic, modeling intelligence as a finite collection of well-defined rules: if b then c; if x then y. Program the correct rules and order of operations, and you would ensure the dexterity of the system. This approach — deemed “good old-fashioned AI” by philosopher John Haugeland — proved useful in domains with stable and easily controlled-for variables. It was good at playing simple games like tic-tac-toe early on, and has more recently been employed to deliver search results for straightforward questions with a single answer, like “What is the capital of California?”
But these systems became unworkably brittle when confronted with unexpected data points or contextual inconsistencies that had not already been explicitly anticipated and programmed. The AI project Cyc, launched in the latter part of this period, illustrates the limitations of top-down intelligence. Despite amassing more than a million unique heuristic rules over the years, Cyc still proved incapable of more demanding tasks, like digital image recognition, or more complex gameplay, like chess or Go. For this reason, chronicler of algorithms Pedro Domingos has derided Cyc as “the most notorious failure in the history of AI.”
From the perspective of fuzzy logic, the uncertain future was no longer something to be defanged but to be welcomed
While not as prominent at the time, alternatives to top-down intelligence also popped up in early AI research, approaches that would eventually lead to the recent breakthroughs in the field. Among them was “fuzzy logic,” a technique first developed in the mid-‘60s by Lotfi Zadeh, a computer science professor at the University of California, Berkeley. Zadeh’s system was an important departure from the top-down approach and the bivalent logic which typically undergirded it, in which all propositional statements must be either true or false. Fuzzy logic, on the contrary, suggests instead that there are computable gradations between the two. Everything for Zadeh — including truth itself — was a matter of degree. With fuzzy logic, it becomes possible to logically process statements like, for instance, “It is somewhat true that this apple is somewhat red.” From the perspective of fuzzy logic, the uncertain future rife with contingencies was no longer something to be defanged but to be welcomed with open arms and an ever fuzzier mind.
Some other logicians and computer scientists were skeptical, insisting that this sort of approach was misguided, even pernicious — that it would grind scientific progress to a halt, given its apparent troubling of objective, clear-cut scientific facts about the world. Mathematician William Kahan went so far as to describe fuzzy logic as “the cocaine of science.” But for Zadeh, fuzzy logic was a superior approximation of reality precisely because of its compatibility with the world’s constitutive ambiguousness. He liked to quote investor Warren Buffett on this point: “It’s better to be approximately right than precisely wrong.”
This new approach held considerable promise for AI, as it purported to seamlessly integrate and learn from unexpected inputs, as humans can apparently do so effortlessly, rather than simply malfunction on contact like the more unyielding top-down systems. Unlike good old-fashioned AI, fuzzy logic is a decidedly “bottom-up” affair that begins by processing manifold particulars in situ rather than blindly applying universal rules, orienting AI research away from rigid deductive systems and toward more agile inductive ones.
Zadeh’s fuzzy logic joined a growing chorus of computational techniques introduced during this period that were collected under the mantle of “connectionism,” a term with origins in neuroscience. This signaled a new paradigm, in which AI applications are effective insofar as they are unstable models of an unstable world, with each new input modifying and progressively updating the overall model itself. Bayesian networks — the breakthrough technique developed by UCLA computer scientist Judea Pearl in 1985 — as well as the many machine-learning techniques that have followed suit derive from this line of thinking.
Contemporary “deep” applications of these bottom-up techniques, however impressive on a technical level, have also directly led to the kinds of algorithmic systems and “smart” technologies that are currently eliciting so much handwringing among concerned tech pundits. AI historian Matthew Jones, in a newly published paper, has argued that this shift toward connectionist computational models pivots on a “predictive ethos,” which privileges instrumentality over interpretability. The constitutive fuzziness of these bottom-up, predictive systems, following Jones’s insight, appears to be the primary reason users and experts alike seem equally baffled by their inner workings. Those at the helm of AI systems often cannot articulate how they come to work in the first place and why, for example, they consistently drive users toward conspiracy or reproduce sexualized and racialized search results.
The response to this ongoing crisis from prominent tech executives has been to repeatedly promise to do better in the future. But this is essentially a plea for self-regulation, allowing companies to continue to operate without clear political boundaries or structures of accountability. Take, for example, Mark Zuckerberg’s triangulating pledge in the wake of the Cambridge Analytica scandal: “We will learn from this experience to secure our platform further.” If we listen closely, we might detect an echo of Zadeh’s confrontation with uncertainty, in which mutability was treated as a feature, not a bug.
But recently the discourse revolving around tech’s existential moment has taken an unexpected turn: “Ethics” has become an inescapable watchword. “Ethical tech, “ethics in AI,” and “data ethics,” alongside corporate mission documents and corporate ethics boards, have spun out into a veritable cottage industry.
Since leaving Google in 2016, Tristan Harris has seemingly become Silicon Valley’s resident ethicist — a title he proudly dons. According to a frequently cited Atlantic article, Harris is the “closest thing Silicon Valley has to a conscience.” He now leads the Center for Humane Technology (formerly known as Time Well Spent), an initiative that officially launched in April with a proselytizing presentation demanding a “new agenda for tech.” Above all, this new agenda aims at a course reversal for the industry, centered on the development of a “common understanding” of our problems and a “common language” with which we can discuss solutions. An impressive coterie of founders, funders, academics, and mindfulness advisors have coalesced around Harris, suggesting the broad appeal and felt urgency of his crusade.
There is no growth hack for ethics
During the Center for Humane Technology launch event in April, Harris alerted his audience of Silicon Valley insiders to the unfolding horror story of unchecked tech: “The call was coming from inside the house; the problem was inside of us.” This was anything but an insinuation that his audience themselves may have played a starring role in this unseemly production. Rather, the problem is everyone’s fault. Humans, Harris argues, are biologically hardwired such that we can’t help but distract ourselves with shiny screens and glittering gizmos.
Harris, curiously, is also a devout Aristotelian, if his public proclamations are anything to go by. That is to say, he, like the ancient Greek philosopher, understands ethics to be the individual pursuit of virtuous moderation. In the Nicomachean Ethics, Aristotle argues that to live the good life, one must constantly strive for the right balance between polar extremes. Too much ambition is just as bad as too much lethargy; to be obsequious to others is no better than to be churlish. Ethics, for Aristotle, is the lifelong endeavor to locate and occupy the ambiguous yet optimal middle ground between excess and defect. To be ethical, we, as individuals, “must cling to an intermediate state” and model ourselves in accordance with what he calls “the unnameable mean.” For Aristotle, too much of a good thing, it turns out, is ethically lethal. Put a bit differently, “we will have to resist the perfect.”
That last quote, however, comes not from Aristotle but from Harris, commenting in a 2017 Wired interview about YouTube’s content-recommendation algorithm. Harris’s concern stems from his belief that engineers are now capable of developing algorithmic recommendation systems that are simply too good for our own good, too addictive for the viewer to resist. While these algorithms may maximize engagement metrics and juice the bottom line for company shareholders, this comes at the expense of mercilessly subjecting the malleable masses to “a whole system that’s much more powerful than us, [which is] only going to get stronger.”
For Harris, this represents a flipped script in which humans are no longer served by technology, as the industry, in its constitutional optimism, might have once intended, but have instead become subjected to its throes. This can all be chalked up, the story goes, to the inexorable failings of human nature that lead us to abuse such power. But as computing power steadily increases, so too does our individual vulnerability. In rendering such a broad, sweeping diagnosis, it is worth asking whether these self-appointed tech ethicists — so intent on saving “us” — have inadvertently reified an essentialist view of “humanity.”
According to critics like Harris, the insalubrity that follows every time machine-learning algorithms are airdropped into a new domain reflects not the garbage data going in or the garbled programming that attends to said data, but rather the innate glitches in our biological hardware. According to the Center for Humane Technology, our children have become addicted, our attention spans have waned, our communities have become polarized, and likes, shares, favs, and retweets have come to rule everything around us. In his remarks at the launch event, Harris explained that these were all byproducts of an out-of-sync relationship between our “inner technology” and “outer technology,” which necessitates a “full stack socio-ergonomic” reboot. Such use of software metaphors to explain the human condition is par for the course in Silicon Valley, but the Center for Humane Technology seems to have set a new standard. Its agenda for ethical tech, ultimately, is a response to what Harris and his cohort have identified as a universal “downgrading” of humanity, the inevitable outcome of tech run amok. By implication, becoming more ethical is akin to “upgrading” ourselves, just as we might install the latest operating system on our devices.
Although the Center’s launch introduced a bevy of new terminology, much of this has already been aired over the past several years as Harris has traversed the public lecture, magazine profile, and podcast circuit, preaching about how powerless individuals have become when confronted with the technologies of persuasion. As Nick Seaver has pointed out recently, Harris’s grand plan, now institutionalized, seems premised on fighting back against behavioral design with — what else? — behavioral psychology. (I’m reminded here of a classic but since deleted quip from Twitter user @CrushingBort: “hmm well I’d say I’m fiscally conservative but socially very liberal. The problems are bad but their causes … their causes are very good.”) This points to yet another level of latent irony within Harris’s portent that “the call was coming from inside the house.”
Harris’s assessment of the problems associated with unbridled social mediation by algorithms seems to fit neatly within Aristotle’s tripartite conception of rhetoric: User-experience designers and software engineers deploy scarily effective — in his words, “godlike” — techniques (logos) to exploitatively hack into our “Paleolithic emotions” (pathos), and therefore the only appropriate response is to rethink tech’s underlying ethos. But Aristotle made clear that ethos is not a solution or counterbalance to the potential pitfalls of pathos and logos, as Harris seems to be proffering, but is itself a constituent part of successful persuasion. So if the cultivation of an ethos is this group’s preferred option for tech intervention, we might then ask what Harris and his ilk — rhetoricians par excellence — are trying to persuade us of with such histrionics about the perils of persuasion.
On first glance, tech’s ethical mandate seems little more than a containment strategy, likely developed by well-intentioned public relations consultants. But it would be unduly cynical to dismiss Harris and his fellow ethical reformers as merely acting in bad faith. The problem isn’t that they don’t earnestly mean what they say, but that they do: “Ethics” functions intentionally here as little more than a vague rhetorical stand-in for some ostensibly shared ideal of “goodness” or “fairness.” It is largely bereft of specific and actionable demands, offering instead hand-wavy consternation about the future. Rather than stake out explicit and structural theories of justice, the movement’s proponents offer instead moralizing sermons on the need to reclaim, as Center for Humane Technology cofounder Aza Raskin put it, our individual rights to the “you-colored prism, the rainbow only you can see,” which data brokers have ruthlessly taken from us.
Almost without exception, the ethical course charted by these Silicon Valley cartographers is an unmistakably solitary pursuit, as Raskin’s poeticizing illustrates. This corresponds with Aristotle’s program: “For each person, ‘good’ is what is good to him.” In both cases, to invoke ethics is to invoke a series of personal lifestyle choices geared toward self-improvement. From that vantage, it’s no surprise that new-age mindfulness specialists are along for the ride, gently reminding us to “look inward.” (There’s an app for that!) Altogether, the growing conversation around ethics in tech seems to further entrench an ideological commitment to individualism, hand in hand with a renewed humanism, which has evidently been technologically corrupted and now requires rescue.
But the idea that new technology is undermining our ability to be fully human is nothing new. Machine-learning algorithms are merely the latest villain to stoke a media panic, following a pattern of similar anxieties about television, radio, and even writing itself. So what is the actual danger that “ethical tech” is supposed to protect us from? And why have both startups and old-guard tech firms embraced this readymade ethical mandate? With a spotlight trained so brightly on ethics, what have we relegated to the shadows? All the hubbub about ethics in tech obscures an industry-wide allergy to politics as such.
It appears that individuals and corporations alike have taken a shine to the ethical mode of iterative self-optimization, a key tenet of Aristotle’s program. This development, at its core, mobilizes the presumption that we all share a common understanding of the virtuous future that we are collectively striving for, even if its specifics remain hazy. It’s an ethical outlook that, tellingly, mirrors the epistemological foundation of Bayesian machine-learning, wherein one’s current beliefs about the future are progressively updated to account for newly collected information about the past. The future, under this rubric, is reduced to nothing more than an incrementally optimized version of the present.
Google espoused a logic in which incompatible worldviews were conflated: There is no true and false here, only gradations along a commensurable spectrum
We might call this phenomenon fuzzy ethics. But the point is not one of revisionist history, implying “Zadeh computerized ethics and we’ve suffered ever since!” Instead, it is to suggest that our contemporary computational paradigm — as with virtue ethics — is rooted in a future-oriented ideal that is designed specifically to be never fully realized. The subject modeled by algorithmic recommendation systems is cut from the same cloth as the one posited by Aristotelian ethics: Both are abstracted universals always undergoing a process of individual refinement. We learn to “do better” just as our machines learn to “do better.” But better at what? We can never be certain.
Fuzzy ethics is also not simply a bad or ineffective mode of ethics. Rather ethics as such is characteristically fuzzy in that it begins with the postulation of principles for individual conduct. This fuzziness no doubt resonates with our individual diversity (recall Raskin’s appeal to the “you-colored prism”) and can meaningfully guide us in our day-to-day interactions with others. But it also explains why ethics is woefully limited when employed in the face of widespread, structural injustice. Translated into start-up jargon, we might say that there is no growth hack for ethics.
Structural questions demand structural answers. Enter politics.
Philosopher Jacques Rancière has described the ways in which the recent “ethical turn” taken in social theory inevitably produces “a state of indistinction between cause and effect.” Judea Pearl, ironically enough, has similarly decried the AI archetype he helped pioneer on nearly identical grounds. While classical probabilistic reasoning might, say, correlate the falling readout on a barometer with an impending thunderstorm, it is of no help in distinguishing which event causes the other to occur. As Pearl plainly puts it, we cannot “explain to a computer why turning the dial of a barometer won’t cause rain.”
The tech industry’s eager adoption of ethics can be seen as an attempt to exploit this blind spot and thereby evade scrutiny and defer any accountability for the societal effects of their products and services. This, in turn, suggests that a theory of causality might be an effective antidote to fuzzy ethics. Accordingly, we might better understand and respond to tech’s impact on social life through a distinctly political lens that clarifies the constitutive fuzziness of an ethical mandate. In stark contrast with Aristotle’s virtuously moderate “unnamed mean,” a political mandate would require the identifying of actors, incentives, and outcomes, and an articulation of causality that structurally connects them. Rather than merely posit a sociotechnological world shaped by inscrutable and omniscient forces, a political mandate would begin by sussing out concrete actors with concrete interests. That is, politics names names.
One influential advocate of the ethical turn, Jane Bennett, has suggested that an analytical framework invested foremost in ethics might nobly “interfere with the project of blaming,” which in practice tends to disproportionately punish the already powerless. However, ethics also stultifies the project of claiming, and politics is nothing if not the prescriptive staking of claims on the future. Bennett and Zadeh both present a rich, high-resolution image of a world in flux — which is, we can now say, ethical in its very nature — but while this might be appropriate on a descriptive level, it remains in desperate, if counterintuitive, need of some prescriptive pixelation.
This work is happening, but the action is far afield from the likes of Harris’s think tank. Consider, for instance, the brave actions last year of Google workers protesting against Project Maven, a Pentagon contract their employer temporarily held to build AI for unmanned military drones. Likewise, we can observe the activism spearheaded by the Tech Workers Coalition, which aims to organize engineers alongside the vendors, temps, and contract workers who, in actuality, keep tech campuses afloat. And more recently, much might be gleaned from the successful campaign by Googlers Against Transphobia, which arose in response to their employer’s proposed ethics board.
In this latter case, Google espoused a logic — call it fuzzy — in which incompatible worldviews were qualitatively conflated: A strident transphobe, a military drone CEO, a behavioral economist, and a philosopher of ethics were all treated as equally valid input variables, as if the ethics board was itself little more than a bottom-up computational problem to be mathematically optimized. As with Zadeh’s logic, there is no true and false here, no right and wrong, only gradations along a commensurable spectrum. But workers’ resistance movements exemplify a different logic entirely: a model of collective action, informed by firmly held ideological positions and invested in ground-up solidarity building. Unsurprisingly, the same companies touting their ethical bona fides have started to lash out against this type of organizing.
Harris’s proposed ethical program is, in his own words, meant to inspire a “race to the top,” with tech firms competitively vying to replace the incentive structure of an “extractive” attention economy with a “regenerative” one. We are left to wonder what this aspirational project of regeneration actually entails. But in any case, therein lies the problem: Politics is not merely regenerative, which implies a simple rearranging of the existing pieces. At its best it is also generative, capable of creative reimagination in the fullest sense. Fuzzy ethics, much like fuzzy logic and its innumerable derivatives, ultimately cannot supply us with a structural theory of cause and effect that would implicate sites of power, articulate and demand concessions, or formulate a generative vision of the future worth agitating for. Ethics, therefore, should not be understood as an adequate “blueprint” for building an equitable future, as has recently been argued, but rather as something more like a big binder of colorful paint swatches, most useful once the new foundation has been laid.
Until we hold the tech industry to a higher standard than ethics, causality will remain black-boxed and culpability will be eternally deferred. We cannot allow its power to hide behind the logic of correlation and probabilism — neither in its computational nor its ethical guises. Instead, we should aim to harness the generative friction which occurs between competing agents with antithetical interests. The universalized subjectivity of virtue ethics must be contested with a coalition-based movement which understands the virtues of antagonism. In other words, fuzzy ethics might be best countered with a gallant return to ones and zeros.