In Phillip K. Dick’s novel, A Scanner Darkly, the characters live in a world pervaded by audio-visual surveillance. Constantly under scrutiny from ‘scanners’ covertly operated by law enforcement agencies seeking the suppliers of a dangerous drug, the protagonist’s reality is complicated by his role as an undercover narcotics agent, and yet more so by his increasingly damaged brain which struggles to reconcile his real and false personas. How do these scanners really see him, he ponders at one point in the novel, do they see the constant struggles he experiences within himself? Do they, like the glass or mirror that Paul the Apostle describes in Corinthians, see into him darkly, which is to say, poorly, incompletely, or not at all?
It is said so often as to verge on cliché that Dick’s work has a striking foresight about it, from the artificial animals of Do Androids Dream of Electric Sheep, to the false memories of We Can Remember It for You Wholesale. Such an observation rather misses the point that the best science fiction is rarely just an act of fantastical future gazing. From Yevgeny Zamyatin’s We to P.D James’s Children of Men, the prime examples of the genre are almost always a critique or response to the author’s present concerns. Dick, for his part, was undoubtedly writing with the US war on drugs in mind, which had been violently escalating since Richard Nixon had declared it six years before the novel’s publication. But Dick was also writing in the context of new high technologies which were then just starting to break through into public awareness, amongst them nascent machine intelligence.
Paulius Petraitis’s A man with dark hair and a sunset in the background shows to some extent how far we have come in forty years, from thinking of seeing machines as the fantastical subject of fictional speculation or philosophical thought experiments, to their becoming an increasingly familiar part of our reality, something we encounter knowingly or not, almost every hour of every day. Petraitis uses a disconcertingly simple device in this book which almost effortlessly opens up many questions about these technologies. Taking his own photographs and feeding them into Microsoft’s Azure cloud processing platform, he presents two captions generated by the machine for each image, one made early in the project’s development, the other right before the final book was sent to print.
The expectation, and in some cases the reality, is one of the advancements and improvement, both in the system’s estimation of what the image shows, and also in its ‘confidence’ in its own judgement, a mathematical expression somewhere between zero, representing no confidence, and one, representing full confidence (but in practice always somewhere between these absolutes). The advances over three years of the project prove to be uneven, and in some cases even a backwards step, with Azure’s pronouncements sometimes hardly changing, at other times becoming more cryptic, mystical and divorced from the apparent content of the images. Such uneven steps are not entirely surprising, reflecting the deep subjectivities of visual materials, and the constantly shifting structures of these technologies, and the current concerns and priorities of those who build them.
The practical simplicity of Petraitis’s book conceals an inquisitive depth, in that the book opens up numerous questions and avenues of thinking about these technologies and their implications, far more than a short review like this can begin to identify, let alone unpack. Some of them are obvious and already widely discussed ones, for example the dangers of algorithmic bias and the integration of such subjective and inconsistent systems into critical spheres like biometric surveillance. Others are far more subtle questions, more remote from the concerns of the present but no less critical to them. What, for instance, are the implications of mathematically merging the two very different, very subjective descriptive systems of visual image and written word in the way that these technologies do?
Artificial intelligence for many of its advocates posits a god from the machine, a series of technologies capable of superhuman promise in their capacity to extract profit from the abundant data which is the new wealth of nations.
But perhaps the one least considered, and at least for this reviewer the most profound, is just what these computer vision technologies mean for our relationship with our own sight. Cognitive psychologists and artificial intelligence researchers have been aware for some time that the quest to create thinking machines has huge implications for our understanding, even our definition, of human intelligence. By contrast, little ink has been spilt on what similar ramifications artificial vision might have, perhaps because we erroneously assume sight to be merely something mechanical, and therefore unproblematic and easily definable. Computer vision technologies tend to implicitly support these assumptions, by presupposing a definitive single way of seeing, that there can be a single mathematically formulated interpretation of what these pictures mean, and that this should be the goal of these systems.
Yet sight, at least in its human form, is a vastly variable, messy process, of which only around tenth occurs in the eye itself, and which remains tied up with a whole host of development specificities, and evolutionarily redundant survival mechanism which still profoundly shape the way we perceive or don’t perceive the world. A question few seem to be asking, is which version of human sight are we emulating with these machines? Clearly it is not the vision of Oliver Sack’s patient Dr. P, unable to recognise faces due to his visual agnosia, or Alexander Luria’s patient S, who’s sight was linked to his other senses in a multi-fold synaesthesia which facilitated a near perfect memory. It would be easy to make the old mistake of simply dismissing these two examples from the case books of neuropsychiatry as pathologies and abnormalities, but what they reflect is that vision is as much a neurological, psychological process as it is an optical and biomechanical one.
More importantly still, these outliers remind us is that all seeing is on a spectrum of difference, and I cannot be certain that you perceived the form of this text in quite the same way as me, nor that either of us experience it ‘as it really is’. Equally, and as in Wittgenstein’s duckrabbit, perception and interpretation are two different things. I can be fully confident in what I perceive, but what I see is not necessarily a full reflection of all there is to be seen and interpreted. Sometimes as important as what I do see, is what I don’t. A widely experienced example of this is apophenia, the perception of meaningful patters where there are in fact none, for example seeing the shape of a face in a random pattern, another example of an evolutionary throwback which made a lot of sense for our arboreal ancestors but less for us today. But for the high capitalism which is the main driver of computer vision technologies, these philosophical ponderings are rarely considered. Reports of positivist empiricism’s death have been greatly exaggerated, and it in fact remains if not alive and well, then at least an implicit ghost hidden in these machines.
Returning to Corinthians, Paul deploys the metaphor of the glass or mirror for the impossibility of a complete knowledge of God, who is ultimately unknowable, and only perceivable in fragments. Artificial intelligence for many of its advocates posits a god from the machine, a series of technologies capable of superhuman promise in their capacity to extract profit from the abundant data which is the new wealth of nations. But like god, these technologies are ultimately unknowable and opaque, and we lack even Paul’s promise of a second coming to reveal all. We as yet have no means to finally answer the question of whether these machines see into us lightly, or darkly – or whether they in any meaningful sense, see us at all.