Automated content moderation and algorithmic discrimination: the case of hate speech

Abstract

The need for Internet intermediaries to moderate user-generated content has become more and more pressing. Besides, vis-à-vis the extraordinary increase in the quantity of daily online information, the resort to algorithmic tools for moderation is today essential. This is also true for the detection of hate speech acts, which is currently largely based on the use of AI and machine-learning techniques: however, scholarly literature has highlighted how such systems can often be vitiated by discriminatory biases which produce a high risk of false positives affecting minorities. The present contribution argues that, within the European constitutional framework, the fight against hateful contents finds its rationale in the goal of ensuring that all social groups can truly enjoy a substantive equality, and that, as a consequence, a discriminatory enforcement of hate speech bans is inconsistent with the value system of the EU. Therefore, although AI represents a fundamental and necessary tool to guarantee a safer and more tolerant digital ecosystem, a high rate of false positives is not fully acceptable when it comes to hate speech moderation. It is thus necessary to rethink the relevant political and legislative strategies, with a view to ensure that marginalised groups can enjoy appropriate substantive and procedural guarantees protecting their freedom of expression and their right to non-discrimination.