Written by Haris Yusoff, LLM student at UCL.
Current debates around artificial intelligence and data tend to spotlight two contrasting perspectives. On one hand, there are industry-led perspectives of boundless market growth, spurred by novel tech applications and the transformative, value-generating capacity of the ‘data explosion’. On the other, there are cautionary warnings from civil society and policy-makers alike of high-risk applications, embedded data biases, and algorithmic discrimination.
These perspectives are often improperly conceived of as binary and opposing, in the sense that pursuing regulation necessarily hampers innovation. While this makes for trending headlines and polarised public discourse, another impact of this is that it places important issues on the back foot. This article will focus on the digital divide, a concept that predates trending headlines, yet receives, from my perspective at least, significantly less public and policy-maker attention than warranted. This is heavily influenced by an industry-led conception of AI tools as open-access, low-cost, and benefit-for-all. Often pushed through product advertising in areas such as healthcare, this ideal is worth critically interrogating, not to tear down, but to reveal a number of sobering realities in the digital world.
Understanding the Digital Divide
In its early conception the digital divide referred to the uneven diffusion and access to information and communications technologies between individuals, businesses, and territories. This follows trends created by a number of disruptive technologies – the microprocessor, personal computing, and the internet. As these technologies have become increasingly capable, the digital divide has taken on a more layered meaning.
Now the divide, – which is driven largely by disparities in AI uptake – is no longer understood as binary “haves” and “have nots”. In addition to access to technologies, it includes disparities in skills, information, and patterns of use, as well as resources and infrastructure necessary for technology uptake. In turn, downstream economic and social consequences that arise from these disparities are integral to contemporary understandings of the divide.
ChatGPT, an advanced LLM developed by Open AI, provides a familiar but illustrative example of how this layered definition operates, as well the advertising practices and unfounded assumptions mentioned above.
ChatGPT is marketed as an easily accessible, ready-to-use tool that can bring transformative technology to ‘all of humanity’. Open AI’s success stories often emphasise relatable users who deploy the technology for simple, everyday tasks. At the same time, they market more advanced capabilities, including GPT operating as a foundational model for fine-tuned health-care applications when combined with health data. Advertised examples include helping clinicians synthesise unstructured electronic health records or conduct advanced medical research. These stories emphasise the potential to improve access and efficiency for patients, as well as collaborate with or even ‘tutor’ expert researchers.
Considering first the more basic uses of ChatGPT (e.g., as a search engine or document summarisation tool) to efficiently access these functions requires regular access to a computing device and reliable high speed internet. To use efficiently might require a certain level of digital literacy, experience in prompt-writing, financial means for subscription fees, awareness of system limitations, and access to alternative AI tools if necessary.
To access these systems without exposure to demonstrated risks such as output bias, hallucinations, or discrimination, will require training data representative of user population and regulatory measures enforceable in a particular jurisdiction. Each step is premised on hidden assumptions – certain levels of primary to tertiary education, digital literacy, access to online banking, regional infrastructure, and strong public institutions.
The reality is that 2.6 billion people globally don’t have access to internet and 1.4 billion don’t have a bank. One in five adults have no formal education, and only 54% of the world own a smartphone. In the United Kingdom, 1 in 5 adults lack the digital skills necessary for full participation in the digital world. The vast majority of countries lack frameworks to safely regulate risks. Although selectively chosen, these figures demonstrate the need to think critically about underlying assumptions that are potentially influenced by industry promises.
It is true that digital inclusion rates are steadily rising. Likewise, many technologies are becoming increasingly user-friendly, lowering the demand for digital literacy skills. Coupled with the advanced capabilities and highly-scalable nature of AI systems, a techno-optimist’s future – such as AI-powered doctors available globally – might not seem beyond reach.
However, many AI-powered medical tools currently in use, including those for chronic-disease detection or diagnostic support, rely on training data that again must be representative of diverse demographics. This reiterates the point that each optimistic promise often promoted through marketing must be tempered to reality. For advanced healthcare applications this reality includes the capacity of local public systems in place to collect and process electronic health data, the access to data analytics knowledge, and the presence of trained clinicians capable of operating complex AI-enabled medical tools.
Data is not in fact “non-rivalrous” as it is often conceived, but rather an exclusive, indispensable input capable of reinforcing market powers. The existence of scalable, “accessible” technologies built on this does therefore not guarantee progress towards reducing inequalities. If anything, current market dynamics indicate the opposite.
The Current State of Play
The current state of AI development is that a few nations (predominantly the US, China, and some European nations) possess the computing power – and access to cheap electricity – necessary to build cutting-edge models. This is further restricted to only a handful of tech companies with the economic power to pay for data-labelling costs or salaries to leading researchers. This has left many nations facing a ‘middle-income technology trap’ characterised by an increasing dependence on foreign digital infrastructure. This situation characterises the ‘downstream consequences’ of the digital divide mentioned above.
Both highly-skilled and toil-intensive work is being increasingly automated, often without new tasks that contribute to labour demand and without equitable distribution of productivity gains. Similarly, environmental externalities associated with AI fall hardest on those without the conditions to accommodate them. Therefore while notions of ‘accessibility’ or ‘open-access’ are true by measure of scalability, hidden costs are imposed on populations in a manner that directly entrenches existing digital inequalities.
Another issue to consider is the erosion of liberties, which is particularly relevant in healthcare and health data contexts by virtue of their sensitive nature and strong links to personhood and privacy. Europe hosts a complex constellation of regulatory instruments in this area, targeting, inter alia, artificial intelligence systems, healthcare data, privacy, and digital literacy requirements for online platforms. While European law-making often encourages regulatory convergence globally and change in industry practices, streams of public discourse – comfortably accustomed to a backdrop of European fundamental right protections – are quick to forget the privilege of operating safely regulated products or enjoying privacy protections.
The erosion of liberties in jurisdictions that lack regulatory protection therefore informs another dimension of the digital divide. Properly construed it has been shown that this divide follows historical disparities in technology uptake (e.g., internet and smartphone access), but also includes subsidiary issues of institutional strength, data analytics knowledge, computational capacity, and fundamental right protections. In high-stakes health contexts many of these issues are exacerbated, yet in my experience are also shifted out of focus by the optimistic lens of scalability and accessibility. Despite the significant truth in these industry promises, they should be cautiously interpreted, tempered by public discourse that critically interrogates – not reinforces – unfounded assumptions these narratives often rest on.