The Humanity of Humanless Systems

A symbiotic relationship between humans and AI/ML-based autonomous machines is contingent on one factor—how we envision the future of “work.” This article questions what humans value in each other and how that value preserves a global society.

From the perspective of a purely profit-driven capitalist system, human beings do not represent the future. Where production optimality is emphasized, people are viewed as inefficient, vulnerable, error prone, and open to fatigue. Thus, they become progressively more vestigial over time. What does offer the vista of progressively increasing return, as well as robust production and spiraling evolution, are artificial intelligence/machine learning (AI/ML)–based autonomous machines. Optimists envision a flourishing and symbiotic future between humans and such systems; pessimists warn against “dystopian” vistas, which feature an ever-degrading quality for all of human life. I propose that each of these respective, and conflicting, visions are actually both underwritten by, and thus contingent upon, one common factor. This factor being, what we envision the future of “work” to be (and see Hancock, 1997). This challenge, with regard to the nature and character of future work, impels us in human factors/ergonomics (HF/E) to ask what it is we value in one another and how such value sustains our global society; now destined soon to number more than eight billion living members.

Fundamentally, what we value, and what connotes work, are actually intentional and guided transformations. In our past, such (work) transformations took the form of a sequence of directed physical changes: think here of the original mining of an ore or securing other base materials and then processing them into a final product; for example, something such as a finished knife. Today, our modern currency of work is concerned much more with cognitive transformations: think here of data mining, which is then processed into refined and implementable knowledge. In each case, a chain of social value is accumulated along the line of those sequential transformations and then the accumulated worth is disbursed; although today such distributions are often highly disproportionate to the effort invested in any particular step. The work, per se, cares little or nothing for the nature of the transforming entity which accomplished each change. Work itself is agnostic to its production mechanism (be it human, robot, automation, etc.). However, in our present world, the original impetus for such technical work, and the conceptions that drive it in the first place, are almost exclusively human in origin. The threat of AI-induced change then pertains not only to the mode(s) of transformation (e.g., manual to automated processes), but the very motivations that initiate work in the first place.

It is here that optimists exhibit their métier for, as yet, we have witnessed no self-intentioned machine systems (Hancock, 1999); or at least not to any degree of evident public acknowledgment. However perverted that any present AI system has become, by sources such as insufficiently defined software instructions, and so on, no presently known AI system has yet expressed the human sentiment: I want, I want! (see Figure 1). In the present, albeit brief commentary, I want to examine these various issues and here I feature an examination of temporal dissonance, a concern that springs from my own personal and perennial concern with time as the fundamental dimension of living existence (Hancock, 2018a, 2018b).

OPEN IN VIEWER

Figure 1.

William Blake’s metaphoric, comedic and prophetic vision of the origins and well-springs of human motivation. It is materialized here as our outreach toward the stars.

With regard to the conflicting face of the argument about AI, the pessimists point to these exact same lacunae in machine motivations, as mediated either by punishment or reward, as the force that may eventually serve to trigger an accidental Armageddon. It is the essence of this fear that the AI does not “know” what it is doing, and so may initiate disastrous actions, albeit unintentionally (either on behalf of itself or its designers). However, there is good news here also. The “singularity” is not upon us, and tomorrow will be rather remarkably like today, even as our technical tools continue to evolve (Navarro & Hancock, 2019). The bad news is that the “plurality” already surrounds us. If we could look down the long corridor of evolution with direct perception, as opposed to inferential presumption, we would see that our own brief, personal existence is insufficient to register the sea-change that has already happened, and is occurring, around us. The Chinese curse, may you live in interesting times, is ubiquitously applicable—for humans are always living in interesting times. It is only a matter of perspective in seeing this. The phenomenological rate of change, to any observer, bound to such arcs of “progress,” will always offer the vista of acceleration into an uncertain future. Our AI-based concerns are only our contemporary expression of this phenomenon (Eiseley, 1973).

On the scale of human direct perception, perhaps over the next few coming years, we are liable to witness much more intimate connections between humans and AI/ML entities. At first, this will be a human-dominated partnership, but rapidly this dominance will dissolve. I have used the metaphor of growing isles of autonomy to describe this evolution in which a penumbra of human partners surrounds a fragile and nascent AGI (where here true autonomy is taken as an expression of artificial general intelligence). Yet, that protective role will quickly become vestigial. The AI will readily become more robust as it is “hardened” by its human designers and eventually expresses “auto-evolution,” as its own programming provides ways of self-improvement. How quickly this transition occurs is one of the great questions of our times (Hancock, 2019a). Thus, we humans possess our own inherent “time-scale” (such as the phenomenological “now” persisting for a matter of 3-6 seconds). There is no reason to suppose that AI/ML entities will comprehend, understand, or respect the inherent flow rate of human temporal perception. Indeed, there is much reason to believe it will not. And systems fostered on cybernetic, temporal optimization would constantly work to reduce cycle-time. That vector of progress would leave human and machine with vastly dissonant timeframes. What happens in a “moment” will be the equivalent of several machine “lifetimes.” Conversely expressed, one machine action (certainly occurring well below the picosecond level), will remain opaque and even incomprehensible to human observers.

So, one of the currently crucial issues in HF/E’s interaction with these evolving systems is what I term as incommensurate temporality. I think of this in terms of the above time-scale conundrum. We are creating AI/ML systems for the purpose of functional extensibility of our own limited operational capabilities. Among these inherent human limits speed of response (and the reliability of that response) are almost necessary constraints. Thus, these putative human “shortfalls” become AI operational goals upon which to optimize (Hancock, 2019b). Yet, particularly as the speed of operations improve, the de facto “transparency” of the AI/ML necessarily diminishes. This is because the human and the machine (writ large), are now working on time bases that are many orders of magnitude apart. So, by the “time” the human can be updated, the AI/ML has already moved to such a degree that the human will never actually be up-to-date. This is the incommensurate temporality bind. Here, any contemplation (or indeed perhaps even action) on the human-user’s behalf becomes moot. (I should note that we can slow AI/ML systems down in order to look to address this bind, but intentional slowing things down is very much against the optimality/profit zeitgeist and seems evidently self-defeating of the purpose of the AI/ML system to at least some degree). At best, this time issue argues that human–AI collaboration will operate at some overarching “strategic” level, rather than in any form of momentary, tactical manner. But even on the strategic level we will end up chronically and intolerably behind our computational cousins. In the end, the lag will break the cord. Also, in my view, this problem of desynchronization will render “transparency” an untenable concept. What we have to enact are temporal “buffers” between ourselves and the machine; reminiscent of the first calls for intelligent interfaces (Hancock & Chignell, 1989). If there is no buffer, no filter between the momentary expression of the human users’ desire and its material expression in the world, then disaster looms large. Any system that creates, in an instant, what an individual user may merely wish for only as a whim, will act as the means of our effective existential extermination. It is the most pressing and urgent challenge to understand how these nascent, autonomous entities can exist and work alongside their human creators. Yet so much effort and attention are being put on the question of how they can be created, that insufficient effort is (essentially necessarily) directed as to why they should be created. This schism between “efficient” and “final” cause may well prove our undoing.

In light of the foregoing, what remains a central issue for HF/E scientists, among many other research and application disciplines, is whether our proposed designs for learning systems mean that they will necessarily exploit and explore ranges of behavior that formally cannot be envisaged (or perhaps even known) by their human designers. I rather suspect the answer to this question must be yes. So, despite our best intentions and efforts to constrain such learning, we will inevitably face necessary incidents of “failure,” of ever-growing magnitude (Hancock, 2017). The empirical question being, will such failures in the computational realm percolate readily, and thus be of consequence, in the physical world of human activity (and see Salmon, Hancock, & Carden, 2019)? Here, I believe the application of the power-law conception (and see López-Sánchez & Hancock, 2019), allows us to predict that the overwhelming majority of these “failures” (i.e., distinctive and disparate explorative states) will be “damped out” in both the electronic “world” and the physical world to which the former’s actions are connected (and see Kaufman, 1993). We learn the influence of this relationship from nature and the way that species themselves have emerged and then been rendered extinct in a manner consistent with this power-law interpretation (see, e.g., Raup & Sepkoski, 1984). It is one of the prime lessons of biomimesis, which HF/E must embrace if it is to engage in principled interactions with emerging AI/ML systems. This principle being: Nature has already made the mistakes that we cannot afford to. Yet, at largely unpredictable intervals, these “electronic” events, whose frequencies are identified by the power-law curve will escape into the “wild” of our everyday world and wreak vast havoc on a human scale (Hancock, 2009). I also suspect it will be almost impossible, even with the aid of these latter-day AI systems themselves, to distill what we humans will find to be a sufficiently comforting and “satisfyingly” causal chain. Given these various premises, we in HF/E and beyond, will need to anticipate a re-casting of what is considered to be “work.” We will further require reconceptualization of causality itself, and then discard our stultified views concerning error; human or otherwise (and see Hollnagel, 2018). Our tenth-generation progeny (if they are still human, and if we get that far), will live in truly interesting times. However, if they remain “human,” they will still be bound by some of the same chains that tie us today. Yet our learned, artificial, intelligent, and autonomous offspring will themselves be busy; buried within their own interesting “times.” It promises to be a bumpy but exciting ride—as long as it lasts!