Shedding light on the obscure intersection of AI and the law
On Thursday, July 30th, the Wikimedia Foundation’s 2020 Summer Legal Fellows hosted a lively and interactive discussion on the Use of Machine Learning Algorithms in Content Moderation. The central question for the panel was: Can we regulate the extent of biased decisions that AI makes? The aim of the discussion was to shed light on the existing statutes regulating AI, how bias affects these complex algorithms, the societal impacts of bias for AI in an unregulated environment, and how well the legal and technology communities can collaborate on eliminating the effects of such bias.
The technology experts began by highlighting the potential risks with machine-learning algorithms, including bias along racial and gender lines. They explained that bias could not be cured with programming alone because the bias reflects the collective ideas of people who are biased themselves. Human intervention is required to make sure that the predictions of the machine learning algorithms are accurate and free of bias.
Accordingly, one of the central concerns amongst the technologists regarding the growing popularity of machine learning technologies is the associated bias that is often inherent in the generic output of these systems. Most of these technologies end up as “black-blox” machines whose decision-making process is too complicated for the lay-user to decode or understand. When these technologies are deployed without proper regulation or without a full analysis or understanding of the source of bias, these biases can have real individual and societal consequences. For example, our panelists discussed at length how racial and gender bias affects the decision making process of the AI technologies.
The panelists also acknowledged that these sources of error often lie in ineffective training data and the sampling or surveying techniques used while collecting data. Some of these biases do not have an algorithmic cure. Yet, a full understanding and appreciation of the existing biases can result in more regulated control of machine learning technologies with human intervention. The rising popularity of glass-box approaches were discussed as a potential way to investigate and interrogate biases in AI systems. The ability of these algorithmic approaches to provide transparency inside the building blocks forming the core of machine learning was appreciated by the legal experts, and its usefulness in terms of deployment, control, and regulation was discussed at length.
Following a brief discussion of whether human insight into AI algorithms is important or not, the law professors delved deep into the interpretation of Articles 13–15 of the EU General Data Protection Regulation (GDPR). While the GDPR specifically states that data subjects have a right to “meaningful information about the logic involved” and to “the significance and the envisaged consequences” of automated decision-making, the debate circled around whether a technical explanation of complex machine learning models is meaningful to most people. The legal experts highlighted principles that must be preserved in a legal system that includes machine learning — such as transparency and fairness — while also discussing the adverse effects of machine learning algorithms on the first amendment (free speech) rights of public as well as private actors.
The legal experts further explained that machine learning algorithms for content moderation are trained with vast amounts of information and there is a possibility that the training dataset used for this might contain copyrighted material. Thus, a pivotal part of the entire discussion was the panel’s colloquy about the decision in Author’s Guild v. Google and whether the Second Circuit was right in concluding that the use of copyrighted materials for training models falls under “fair use.”
We appreciate the diverse perspectives our panelists were able to offer us on regulating algorithmic bias in AI. Throughout the comments of both our technologists and legal experts, recommendations for transparency and human intervention emerged as common ground. By continuing to have cross-disciplinary conversations like these, we can move closer to a workable solution that understands both the technology, the regulation thereof, and the needs of those affected by its biases.
Speakers that contributed to this thought provoking discussion included experts in the field of machine learning and the law, namely Anima Anandkumar, director of Machine Learning Research at NVIDIA and Bren Professor of Computing at Caltech, Rich Caruana, senior principal researcher at Microsoft, Chris Albon, director of machine learning at Wikimedia Foundation, Emma Llansó, director of Center for Democracy and Technology’s Free Expression Project, Pamela Samuelson, professor of law at the University of California, Berkeley, Robert Merges, professor of law at the University of California, Berkeley, Lori Andrews, professor of law at Chicago Kent College of Law, Richard Warner, professor of law at Chicago Kent College of Law and Ashesh Chattopadhyay, PhD candidate at Rice University.
The Wikimedia Foundation’s legal fellowship program
Each year, the Wikimedia Foundation’s summer legal fellows organize a panel discussion on the topics of law, technology and free knowledge. Wikimedia’s legal fellowship program has nurtured dozens of law students and recent graduates interested in intellectual property, privacy, and other technology law issues. The Foundation recruits legal fellows for the spring, summer, and fall.
Priyanka Bhattacharya and Shecharya Flatté
Wikimedia Foundation Legal Fellows
* Thanks to our fellow summer legal fellows who were essential to organizing this event: Daniel Betancourt, Sean Chen, Avantika Bhandari.
