Whose A.I. is this

Michael Salu

I am in your dreams, but you are not in mine is an exhibition of new works by my collective, Planetary Portals, commissioned by The Photographers’ Gallery in London. We’re a creative and critical research group, interrogating continued colonial legacies and geo-traumas, extending from the mining of gold and diamonds in South Africa to the renewed scramble for African minerals and data, fuelling global digital economies. 

The exhibition features four single-shot films tracing a legacy of extractive logic, which relies on structural racism, from nineteenth-century colonial photography to the inequitable assumptions and myopic perspective inherent in generative AI today. This technology purports to create new materials – images, texts, moving images, and sound – based on patterns learned from algorithmic analyses of vast repositories of training data, such as digitised texts, images, films and music. Algorithmic influence over our lives is vast and invisible, and for most of us, impenetrable. The works in our exhibition is intended to provide an alternative look at the new manifestation of this technology, Generative AI.

Our starting point for the films in the exhibition is a selection of photographs from the archives of British imperialist Cecil Rhodes, held at the Bodleian Library, Oxford. Alongside a variety of bureaucratic documents such as letters, ledgers and contracts, these photographs record and propagandise the implementation and corporate consolidation of the diamond mines in what was to become South Africa by Rhodes and his cohorts.

The photographs present a particular vision of this foreign land to potential investors back home, to encourage migration and new business opportunities. The images initiate the Western understanding of foreign landscapes and native communities as docile, benign and lucrative material for conquest, implying an attractive fantasy of mostly empty worlds to make use of. This extractive framing – as well as photography as a developing technology of power – enacts the erasure of existences beyond the frame. The enduring, ever-dynamic non-human world is rendered static and pleasant. The lives of peoples and cultures living in harmony with their environment for generations, subjugated and exploited as forced labour, are hidden, or rendered docile. In the photographs, their appearances are often subtly resistant. Within the frame, black men and women stand in the shadows of triumphant white men replete in their unsullied white clothing, untouched by the dirty and dangerous work of mining and yet, if you look closely, the indentured black men and women’s refusal and resistance of the fantasy is quiet, but palpable. In the photographs and documents, the experiences of these myriad tribal and cultural groups, lumped together indiscriminately into forced labour camps, are flattened; consolidated as a singular expendable commodity. 

Visual coding within landscape photography comes from European landscape painting, which was often commissioned by landed gentry to show their dominion over land and nature. This coding becomes a semantic tool for expeditions across waters to mythologise far-flung colonial exploits. 

Moving image, or film, as a medium is directly birthed by the landscape photograph. How the hero is centrally framed (or the visual hierarchy of human and its imaginative machines ruling over environment) communicates a particularly Western humanist separation from other forms of life (including human life) in order to render them of material value. Generative AI, mostly wielded by powerful mega-corporations with the resources to raid data and deflect litigation, uses large datasets of digitised information from this same legacy. 

Thousands of mostly anglophone books, academic publications, blogs, Hollywood movies, social media content and internet information created, for the most part, by those with access to and control of shaping culture and opinions (including newspapers and magazines) fill out these datasets. Pattern recognition algorithms, often initiated by written and visual prompts, emit textual and visual outputs from these datasets, purporting to be something new. 

In our films, by subverting these prompts, we show how Generative AI – primarily a new marketing term for the consolidation of deep learning processes which have been in development for some years – aligns with the hierarchical perspective presented by the Rhodes archive. And how, today, autonomous digital culture repeats, obfuscates and entrenches structural racial logics found in documents from the nineteenth century. 

Generative AI is only one of several uses and types of Artificial Intelligence. The recent rhetoric and marketing of these technologies is driven by venture capitalists desperate for ever more funding, stock, market gain and geo-political leverage. Rudimentary forms of Artificial Intelligence – formerly algorithmic computing, machine learning and deep learning – have long been significantly influential in society. This obfuscation is intentional, so for the purpose of clarity, we can riff on IBM’s demarcation of this technology, dividing it into two main realms: Narrow Artificial AI and Artificial General Intelligence. 

Narrow Artificial AI, which includes generative AI and Large Language Models (LLMs), is largely based on developing algorithms to recognise patterns within existing data to produce instructive outputs, decisions or assumptions. 

Algorithmic interpretations of life may purport to be objective but it is vital to question intent and how supposedly ‘neutral objectivity’ is determined and then computationally automated, leading to the quiet take over of privacy and free will.

Machine learning and deep learning algorithms that power today’s rudimentary AI originated in models built from probabilistic theorems, developed through the discipline of statistics. Popular theorems such as Bayes Theorem, used as a foundation for developing these algorithms, enables the use of a series of deductions and classifications to accumulate evidence weighted in favour of one’s own argument: true or false, yes or no, man or not man, white or not white.

European Statistics originated in the 18th century and is influenced heavily by Judaeo-Christian subjectivity, asserting and then exporting the idea of the Western human. Statistical methods and Darwinian classification data-capturing methods around a desired norm affirm the notion of a ‘normal’ human and a ‘normal’ morality, which are foundational to eugenics. Although data-driven methods have since moved on, their origins remain attached to the legacy of scientific racism, and have served to create many of the social and political frameworks within which we live today, such as ethics or the definition and autonomy of the individual. 

Statistical reasoning isn’t new but the computational use of such reasoning at today’s scale and influence has never been as all-consuming and unregulated. We simplify the world into binary categories with little ethical and critical deliberation of other possible subjectivities. Computer vision, for example, another AI subset, relies on the manual labelling of categories of objects (man, woman, tree, cat, traffic light and so on), often by exploited workers in poor conditions, for instance, in the method underlying the operation of autonomous vehicles. What happens to objects, or indeed subjects, which are not recognised by the computer’s understanding of subject and object, based on an implied norm? 

In our films, we apply Generative AI’s statistical reasoning to the archival photographs, and it confirms these established hierarchical norms in the following order: white man, white woman, and then everything else. In many instances, nonwhite persons are removed from the photographs entirely.

When power is used to weight data heavily in a particular direction, it indicates how warped probabilistic analyses or systems might be. Taking traditional book publishing as an example, if a Large Language Model (think ChatGPT) is trained only on the corpus of the most widely available works of literature – namely, anglophone literature – what will this LLM infer about the human condition? The works of the literary canon primarily champion humanist values – speaking in universal terms and centering the norm – while anything different (and therefore lesser to those norms) appears within the same pages as absences, erasures, and sometimes vocalised ‘othering’. One need only look at the work of Joseph Conrad or Cormac McCarthy as examples. 

Snaking up the walls of our exhibit are thousands of languages from across the African continent, which existed long before European partitioning of the continent in 1884. Only five of these languages are currently used in the training of Large Language Models like ChatGPT, which has already shifted how we think and write (mainly with implied norms: morally, politically, culturally and even physiognomically) and dresses us in the unsullied technological white garment of the victor.

These norms configure the perspective of much contemporary literature, swirling back and forth in a tight feedback loop of audience and publisher. This information structure feeds increasingly indispensable models powered by the datasets of large corporations, which initiate feedback systems with us, and the generations that will follow us, narrowing epistemic possibilities. The universal ‘we’ comes with hierarchies of erasure, and Artificial Intelligence brings with it a social issue that is deeper than what is now referred to as ‘bias’. This issue is structural and about power: who has the power to determine what the human being is and what their future should look like?

The path towards an AI-operated society has already been paved. Misogynistic and racist behaviour, for instance, did not come to the fore by chance. Some might argue the clamour for the subjugation of women and nonwhite peoples is a natural human response to worsening economic decisions. However, these initial anxieties, following the hierarchical western human norms, become seeded then amplified through feedback loops that play on anxiety. Invisible algorithms are perfect for brewing and spreading paranoia and conspiracy.

The main problem with establishing an AI-first economy and culture, such as the one that the British Government has recently stated a desire to implement, is that this is a far more political act than it has been presented as. Implied within it is the implementation of an epistemic singularity – a supremacist way of being as the norm – which does not just apply to datasets but also to the very foundations of this technology’s architecture. Our uses of language and broad perspectives are correctively filtered using subjective moral and political guardrails to function within these mathematical frameworks. Try using language which deviates from these norms as a prompt within a Generative AI application and you’ll soon reach the implied limits of who we should be and how we should behave. 

AI’s patterns of recognition will only follow whatever it has been fed, even with increasingly independent and intuitive models, however complex its remixing. This is why there needs to be an ongoing socio-political conversation about AI as we need to seriously consider what happens when a worldview becomes automated. AI technologies built on this probabilistic precision, can be applied to, say, a weapon which will deduce the likelihood of finding an enemy target, based on datasets filled with information championing a specific political or metaphysical argument. AI is also being used to target political dissenters, using algorithmic analysis of personal information, then applying assumptive actions. Analytical methods originally used to target and seduce individual consumers are now focusing on individuals’ rights. 

The resource-intensive processes of generative AI has become another coil in the extraction economy, both cognitively and materially. Extremes of material waste, as always, aid the flow of capital to the top. In a similar way to the extractive violence of gentrification (though few would comfortably call it this), the artist currently suffers the wrath of the Generative AI push through the decimation of IP and copyright. Soon, this extractive method will reach another tier of the labour economy. AI as a technology may present a glossy veneer of efficiency and productivity to the individual but the primary aim of venture capitalists that wield its power as services is wealth extraction and wealth consolidation, causing forced cognitive and economic subjugation. There is, of course, political precedence for the ideological imposition of such a method. 

In contrast with Narrow Artificial AI, discussed above, Artificial General Intelligence is, for the moment, theoretical and therefore speculative. But it is thrust to the forefront of AI discourse because it plays on Western ideological desires as an augmentation and evolution of the humanist idea. A transhuman – a human of machine, of technology – is a resurfacing of the eugenicist idea of the prototypical human but within a more advanced rhetorical framework. A transhuman intends to transcend the remaining, perceived failing of the human: the fallibility of the human body. A super-intelligence no longer hindered by limits of the body and mind fulfils the definition of the ‘Ubermensch’ – the European idea of the human as infinite and unassailable – completing the severance of the human (the white male archetype), from other forms of life. This rhetoric drives AI exceptionalism and configures the idea of the superhuman as a natural, inevitable, capitalist fantasy, born of material wealth. 

Planetary Portals: I am in your dreams, but you are not in mine at The Photographers’ Gallery 6 March – 15 June 2025

Cybernetics, or Ghosts? out now from Subtext Books (UK)