The digital era is riddled with plans that backfire. When TikTok decided to limit the reach of content posted by people with visible disabilities, it did not intend -or so it emerges from the documents accessed by Netzpolitik- to harm the interests of this community. According to the leaks, the aim of this peculiar moderation system was to protect these people from cyberbullying. Ten out of ten for intention, zero for execution. “While activists clamor for an internet and visibility without barriers, TikTok has deliberately erected barriers without those affected being able to suspect it”, the German outlet concludes.
In this case, we are dealing with a deliberate decision, driven by a team of humans. What happens when these matters are settled by a set of algorithms that, due to the sheer limitations of current technology, has an even more limited and potentially biased view of reality?
The TikTok scandal comes just days after the AI Now Institute released its report on disability, bias and artificial intelligence: “Even when they work as their designers intended, these systems are frequently used in ways that serve the interests of those who already have structural power, at the expense of those who do not”, the document warns.
Whether accidental or premeditated, these biases are not harmless, and in the case of people with disabilities they take on greater complexity. “Fairness for people with disabilities is different from fairness for other protected attributes, such as age, gender or race”, notes Sharin Trewin, of IBM Accessibility Research. Part of the problem is that this area has received less attention than other biases, but this is no coincidence either. “One fundamental difference is the extreme diversity of ways in which disabilities manifest themselves and in which people adapt. Secondly, information about disabilities is highly sensitive and not always shared, precisely because of that potential discrimination”. The first consequence of the above appears at the source of machine intelligence. “Many face datasets have labels such as gender and race, but rarely include disability labels”, adds Anhong Guo, a doctoral student in human-computer interaction at Carnegie Mellon.
Trewin agrees with Guo in her diagnosis: “When we talk about the most disadvantaged groups in our society, disability often falls into the etc that comes after race and gender. Considering the number of people with disabilities in the world and the fact that most experience some disability at some point in their lives, this should not be the case”. In this sense, by ensuring that decision-making systems produce fair outcomes for majority groups, enormous minorities are left out. “An essential piece of the solution is to provide explanations of decisions. With an explanation, we can begin to trust that the system is fair or does something to address the problem”.
“People with disabilities have been historically and currently subject to a marginalization that has systematically and structurally excluded them from access to power, resources and opportunities. These patterns of marginalization are imprinted in the data that shape artificial intelligence systems”, confirms the AI Now Institute report.
A team of Google researchers found an example of this in an automated conversation moderation system that classified texts mentioning disabilities as more “toxic”. “If we had more people with disabilities working in and with engineering teams, we would have more robust and flexible systems in every respect, not just disability”, says Trewin. “One thing is clear: whether human or artificial, bias is a matter that needs careful attention”.
What happens if we leave things as they are? Exactly that: “At the most basic level, excluding the consideration of disability from the discourse on artificial intelligence and bias means that efforts to remedy those biases will not include people with disabilities and that, consequently, they will be more susceptible to experiencing harm and marginalization”.
However, for Anhong Huo, we have to bear in mind that the battle against bias is not one that will be won soon. In fact, in his opinion it will be fought forever: “There will always be biases; it is an ongoing discussion in which we will have to keep participating. As we move forward, we need to be more receptive and aware of these problems, include them in the datasets and in the processes of data collection, model building and implementation”
first published retina.digital
Montse Hidalgo Pérez
