The Berinmo people of Papua New Guinea are a small tribe of people with a curious flourish in their language. Specifically, they do not distinguish between green and blue, and they have two separate words for types of yellow. This has deep implications for those who argue for a consistent and objective real (universalists), and the ultimate possibility of artificial intelligence. We’ll come back to the Berinmo, and color linguistics later. While strategies for the avoidance of bias in AI focus on the injury of minority oppression and design failings in creator preference, a deeper semantic analysis of some of the fundamentals of AI reveal several foundational assumptions that give cause for concern. Simply put, the fundamental task of an AI is to construct an image of the world within the parameters of its design (from narrowly defined chatbot engines to Artificial General Intelligence or AGI), which in turn establishes the context for automatic machine decisions to be made. In order to arrive at that image of the world – the simulated real – there are several intermediate layers that each introduces a risk of misinterpretation. This article will walk through each, and understand where some of those challenges might lie. But first, Heidegger.
In Heidegger’s analysis of Being (Dasein), he early on lays out his reasons for the work: that Being as a concept had become so fundamental that it had hitherto been ignored as a subject for critical inquiry. However, as he puts it ‘[i]f one says accordingly that “Being” is the most universal concept, that cannot mean that it is the clearest and that it needs no further discussion. The concept of “being” is rather the most obscure of all.’ (Being and Time, Stambaugh translation, p.2) Layers of assumptions – beginning in the core of thought itself – had over time presumed Being as a necessary prerequisite for all other concepts. In his 1964 interview with Bhikku Maha Mani, Heidegger quotes Max Planck who said that ‘reality is only what is measurable.’ This thought, Heidegger says, determines all of technology, and was first conceived by Descartes, who assumed Being as his primary platform – cogito ergo sum. In many respects it is Descartes’ peccatum originale, his original sin that Heidegger’s later work on technology rails against: that Descartes’ dasein-foundationalism corrupted metaphysics for centuries thereafter, and had resulted in what in a later interview he called the uprooting of man.
In a similar way, I argue that AI and modern technology is built upon ontologically weak foundations, a machine instance of human hubris, a silicon righteousness. Standardized AI (if we assume commonly available tools) begins in its design phase with the psychology of the real, and a deep-rooted (if unacknowledged) human separatism – what Tim Morton calls ecology without nature, a rejection of man as a part of nature. The human is modelled as self-contained, despite the external dependencies of the physical subject (such as air, atmospheric pressure, and food), and despite the internal non-human entities such as parasites and bacteria. In addition, the extended non-human world is defined by a set psychology that assumes certain structures and forms and order. It assumes, for example, that one color is green, and that another is blue; in its core coding of color, green and blue are separate primary colors, that combined with red can produce any color needed, based on the Young-Helmholtz or trichromatic theory. This is increasingly the dogma of the image, an assumed structure, a most universal concept.
The second filter through which AI assembles its simulated real is the biological. Persisting the example of the world through color, colorblind people have a different view of how the world actually is. ‘Blind’ here is a pejorative term; in the same way as Foucault questioned in Madness and Civilization how it is we define disorder, or madness, and whether they might be the ‘sane’ ones, could color ‘blind’ people actually see something more closely approximating an objective real? Even if they don’t, or if there’s no such thing, why should those of us not classified as colorblind own the definition? The green-blue of the Berinmo people, or even the grey-green of the Irish language, may be products of a dazzling bright sun, or a temperate, moist climate respectively, yet each representation is as real on one side of the world as it is on the other.
Language is the third filter through which AI must navigate. Natural Language Processing (NLP) is one of the most challenging elements of AI, attempting to understand meaning, colloquialisms, and bypassing written language to engage with the spoken word. There are a thousand pitfalls to avoid here. One could start with Walter Benjamin’s work on The Task of the Translator, to understand the seemingly simple task of converting text from one language to another, revealing subtleties and associations that stem from cultural, political and physical history. If we consider the challenges of the various translators of Heidegger’s Being and Time, their travails on the translation of that one word – being – have inspired countless books and theses. Language captures within it whole sequences of narrative that simply defy translation. Which language captures the world most completely? Can any language capture the world universally? Even the first language – if indeed there was such a thing – could it have had the range, the virtuosity to capture human experience? These first three filters – psychology, biology and semantics – can broadly be referred to as the cultural filters of AI.
The fourth filter is the first (hardware) machine filter – the sensor. In an image processing sensor, light photons are received, and converted to either voltage or electrons, and ultimately a digital signal. A critical element of the process is called a Bayer filter mosaic, which is a popular mechanism for determining colours, based on an RGB array. Similar competing technologies also rely on RGB trichromatic theory, with different sensitivities and results in different light conditions. Which presents a ‘true’ representation? To extend the biological category of the third filter, the eye as sensor can be compromised in other ways too, rendering the world in a different form: how should an artificial optical sensor behave? Other sensor technologies have been developed for sound and other perceptive machines, though haptic sensors for vibrations and other interactions, and even nascent olfactory and gustatory sensors for taste and smell.
The fifth filter is the aggregation filter, at a hardware level. The sensor first needs to capture and process the external signal, as correlated to phenomena – e.g. photons become light, with color, which is in turn assembled pixel by pixel as an image. Sound waves are transformed from analogue to digital as part of the fourth filter (ADC – Analog Digital Conversion), by sampling sound wave amplitude as precisely as their hardware allows – usually not much more than that which is perceptible to the human ear. (In reconstructing analogue sound waves (DAC – Digital Analog Conversion) a process of interpolation approximates the ‘gaps’, and fills them in.) With the digital sound structure, these sounds are assembled and arranged into coherent files, and compressed based on ISO (International Standards Organization) standards (such as MP3), which are ‘internationally agreed by experts’. In an ironic twist, perhaps, the ISO also has a paper on bias in AI systems. These two machine filters, the sensor and aggregator, must be physical layers. While they are functionally separable, and theoretically the aggregation function could be virtualised, the latency issues with current technology would make such a separation both bizarre and impractical. They remain however different processes with different engineering and design principles, which each result in different ontological outcomes.
The sixth filter is one of application, or software. At this point, context and meaning are applied to the processed and aggregated sensory data. In a video camera, the images are assembled together and presented as a file for consumption and sharing, or rendered on a screen – such as a security monitor. This filter is also the first order of history: sequences of micro-events. What I mean by this is for example a video of a car entering a parking lot (a series of individual images), or a sound recording of a conversation (a series of individual sounds).
The seventh and final filter is data – the scoring and measurement of the perceptual phenomena captured and measured and approximated and assessed by the hardware and software, and the final order of history, within its frame of reference. This is both a forward (predictive) and backward looking (definitive) reconstruction of the real, the historical, and of the entire epistemological. Within the confines of the AI – extending from the narrow chatbot to the eternal AGI – the entirety of its history is known, unforgettable, and permanent. Theoretically, its predictions are as accurate as its history; and yet confidence about one (the future) is weaker than confidence about the the other (the past). Without getting too deep into the philosophy of history – as we have done many times before – we immediately get the poverty of data, and digital approximations (the binary wasteland) as coming close to reflecting a true reality.
This ‘simulated real’ is the context within which AI makes its decisions. This is the reconstructed real, based as it is on flawed ontological layers, from the psychological to the mathematical, via epistemology and history. Each is a leaky approximation, a sloppy series of categorizations and compromises, none of which can hope to capture any shared truth, even if such a thing were possible. It certainly can harbor no claim to universalism, and there are countless opportunities for derivative bias to enter the various calculations and machinations, as phenomena in the world make their way through these structures to the simulated real. If we think for example of the Berinmo people, if they were to construct an AI, how would they deal with images? How would they have designed the color filter? How would they realise the world? It would be different. Would it be wrong? Would it be better? Who’s to say. Corporates don’t do ‘better’, of course; because ‘good enough’ makes money, and that’s what they are ‘programmed’ to do. Heidegger warned us all of the danger. ‘Humans are at the complete mercy of technology,’ he said, back in 1964, ‘and someday will be made into a controlled machine.’