Data-driven technologies are at the front and center of the public’s attention these days, thanks to all the buzz around artificial intelligence (AI). But what exactly is and isn’t possible with computational models is often distorted by science fiction scenarios where sentinel robots ruin our lives. In this piece I argue that while AI fuels incredible progress, fair and just societies require a lot more than learning machines.
As a Hungarian whose native language is only spoken by a few million people, I have witnessed first-hand how the evolution of natural language processing (the technology that powers machine translations and popular content-generating tools like chatGPT) transformed access to my mother tongue.
Back in 2006 when Google first launched its translation services, the Hungarian interpretations were painfully inaccurate. As the technology got more sophisticated by shifting from statistical machine translations to a so-called neural engine, Google’s grammatical precision has drastically improved as well.
And while Hungarian translations are still not 100% accurate, let alone comparable to the precision of a human translator, they will enable any non-Hungarian reader to digest online content from my home country — something that was not possible before.
That kind of progress might easily make us believe that online content can be democratised by investing more in computational power.
But even though natural language processing has improved astonishingly, language inequities haven’t drastically changed across the globe. To illustrate, out of the approximately 7000 languages that are used in the world today, only 133 are available as supported languages in Google’s translation services.
The reasons are more historical than technical: these models are trained on data that looks for patterns in hundreds of millions of online documents, but since only a small percentage of the currently spoken dialects have a significant enough digital presence, translation engines developed a very clear prejudice towards well-financed European (and other colonial) languages.
As a consequence, a majority of the world’s cultures have limited online representation and hundreds of millions of people live without access to crucial digital services in their mother tongues.
Achieving true language justice requires so much more than adding AI to the equation. Alone the task of digitising content that powers translation services would require the concerted efforts of archivists, librarians, interpreters and language justice activists — actual humans whose contributions are often overlooked and under-resourced.
Similar patterns arise in many other fields, like healthcare for instance. Thanks to increased access to medically relevant datasets from electronic patient registries, wearable devices and health apps, computational models are becoming mainstream in clinical settings too. The motivations are clear and applaudable: improving diagnostics, preventing relapse, reducing hospitalisation, streamlining epidemiological forecasting –to name only a few of the many areas where data-driven models can and will make an important impact.
But while AI-enabled tools can take some of the burden off of overwhelmed medical workers by automating portions of their work or by helping gain better insights into the human body, they alone cannot guarantee that life-saving treatment is available for everyone, or that resource allocation is happening in fair and just ways.
Take predictive analytics, a type of machine learning that analyses historical data to make more informed decisions about treatment and prevention. Predictive tools are increasingly popular in clinical settings but yield very different results over time and across populations, raising important questions about their reliability and fairness in life-or-death situations.
The reasons are, again, more systematic than technical: thanks to their access to clinical trials and cutting edge technologies, certain demographics are over-represented in the data that powers these models, but prognostic value is seriously reduced for other (often marginalised) communities. Combined with the fact that machine learning goes through a lot less scrutiny than human decisions, clinical AI may improve life expectancy and quality for some, but the benefits will be a lot less clear for others.
Changing that requires so much more than investing in computational power or collecting more sophisticated data about the human body.
Truly accessible and equitable health care regimes rely on the dedication and compassion of medical workers who make good judgements in real-life situations, doctors who fuel automated decisions with their expertise, scientists who produce research that works for all bodies, and activists who push for more fair social, political and legal systems.
Without their involvement, AI is only going to entrench existing health disparities, instead of eliminating any of them.
In fact, research already shows that this is happening in many areas. People with darker skin tones are a lot more likely to be mis-recognised by AI-enabled technologies, or to be over-surveilled, wrongfully jailed, and otherwise harmed by automated decision-making systems. Women and gender non-conforming bodies bear the lion’s share of the abuse that is facilitated by AI, and people living with disabilities are constantly left behind by inaccessible and discriminatory machine learning technologies.
The popular argument in favour of machine technologies is that they have serious potential to influence the long-term future, but evidence consistently shows that real change begins in the present — through actual humans whose compassion and moral guidance creates true equity in real-life-situations. The librarians and translators, medical workers and scientists, public servants, teachers and activists and many others working tirelessly to make the world a better place for all, both immediately and for future generations to come.
Investing in their work is as crucial, if not more important than investing in computers and mathematical models.