by Ambika Tandon,
Artificial Intelligence and allied technologies form part of what is being called the fourth Industrial Revolution. Some analysts project the loss of jobs as AI replaces humans, especially in job roles that consist of repetitive tasks that are easier to automate. Another prediction is that AI, as preceding technologies, will enhance and complement human capability, rather than replacing it at large scales. AI at the workplace includes a wide range of technologies, from machine-to-machine interactions on the factory floor, to automated decision-making systems.
Technological systems always have a differential impact on communities who have unequal access to resources and skills. In the world of work, they could either enhance access to economic opportunities, or amplify socioeconomic inequality. For instance, some commentators have argued that in manufacturing, women are more likely to benefit as tasks that require hard-to-perform manual labour (for which women are less likely to be hired) are automated. Others have argued that there will be net loss in employment for women, as sectors with a high representation of women (such as secretarial work) are impacted negatively, while not enough women are being hired in sectors that are witnessing job growth (such as data analytics). Similar arguments can be made for factors such as race and income-level of workers.
The platform economy, in particular, is dependent on AI in the design of aggregator platforms that form a two-way market between customers and workers. Platforms deploy AI at a number of different stages, from recruitment to assignment of tasks to workers. AI systems often reflect existing social biases, as they are built using biased datasets, and by non-diverse teams that are not attuned to such biases. This has been the case in the platform economy as well, where biased systems impact the ability of marginalised workers to access opportunities. To take an example, Amazon’s algorithm to filter workers’ resumes was biased against women because it was trained on 10 years of hiring data, and ended up reflecting the underrepresentation of women in the tech industry. That is not to say that algorithms introduce biases where they didn’t exist earlier, but that they take existing biases and hard code them into systems in a systematic and predictable manner.
Biases are made even more explicit in marketplace platforms, that allow employers to review workers’ profiles and skills for a fee. In a study of platforms offering home-based services in India, we found that marketplace platforms offer filtering mechanisms which allow employers to filter workers by demographic characteristics such as gender, age, religion, and in one case, caste (the research publication is forthcoming). The design of the platform itself, in this case, encourages and enables discrimination of workers. One of the leading platforms in India had ‘Hindu maid’ and ‘Hindu cook’ as its top search term, reflecting the ways in which employers from the dominant religion are encouraged to discriminate against workers from minority religions in the Indian platform economy.
Another source of bias in the platform economy are rating and pricing systems, which can reduce the quality and quantum of work offered to marginalised workers. Rating systems exist across platform types - those that offer on-demand or location-based work, microwork platforms, and marketplace platforms. They allow customers and employers to rate workers on a scale, and are most often one-way feedback systems to review a worker’s performance (as our forthcoming research discusses, we found very few examples of feedback loops that also allow workers to rate employers). Rating systems have been found to be a source of anxiety for workers, as they can be rated poorly for unfair reasons, including their demographic characteristics. Most platforms penalise workers for poor ratings, and may even stop them from accessing any tasks at all if their ratings fall below a certain threshold. Without adequate grievance redressal mechanisms that allow workers to contest poor ratings, rating systems are prone to reflect customer biases while appearing neutral. It is difficult to assess the level of such bias without companies releasing data comparing ratings of workers by their demographic characteristics, but it has been argued that there is ample evidence to believe that demographic characteristics will inevitably impact workers ratings due to widespread biases.
It is clear that platform companies need to be pushed into solving for biases and making their systems more fair and non-discriminatory. Some companies, such as Amazon in the example above, have responded by suspending algorithms that are proven to be biased. However, this is a temporary fix, as companies rarely seek to drop such projects indefinitely. In the platform economy, where algorithms are central to the business model of companies, complete suspension is near impossible. Amazon also tried another quick fix - it altered the algorithm to respond neutrally to terms such as ‘woman’. This is a process known as debiasing the model, through which any biased connections (such as between the word ‘woman’ and downgrading) being made by the algorithm are explicitly removed. Another solution is diversifying or debiasing datasets. In this example, the algorithm could be fed a larger sample of resumes and decision-making logics from industries that have a higher representation of women.
Another set of solutions could be drawn from anti-discrimination law, which prohibit discrimination at the workplace. In India, anti-discrimination laws protect against wage inequality, as well as discrimination at the stage of recruitment for protected groups such as transgender persons. While it can be argued that biased rating systems lead to wage inequality, there are several barriers to applying anti-discrimination law for workers in the platform economy. One, most jurisdictions, including India, protect only employees from discrimination, not self-employed contractors. Another challenge is the lack of data to prove that rating or recruitment algorithms are discriminatory, without which legal recourse is impossible. Rosenblat et al. (2016) discuss these challenges in the context of the US, suggesting solutions such as addressing employment misclassification or modifying pleading requirements to bring platform workers under the protection of the law.
Feminist principles point to structural shifts that are required to ensure robust protections for workers. Analysing algorithmic systems from a feminist lens indicates several points in the design at which interventions must be focused to ensure impact. The teams designing algorithms need to be made more diverse, along with integrating an explicit focus on assessing the impact of systems at the stage of design. Companies need to be more transparent with their data, and encourage independent audits of their systems. Corporate and government actors must be held to account to fix broken AI systems.
Ambika Tandon is a Senior Researcher at the Centre for Internet & Society (CIS) in India, where she studies the intersections of gender and technology. She focuses on women’s work in the digital economy, and the impact of emerging technologies on social inequality. She is also interested in developing feminist methods for technology research. Ambika tweets at @AmbikaTandon.