Process

Gender-based violence now exists on an online-offline continuum. Online violence is often accompanied by offline violence. Uli (meaning chisel in Tamil) is our attempt to hand the chisel and power over to users most affected by online gender-based violence.

Uli was born from the collective labour of journalists, activists, community influencers, and writers engaged in the struggle against the interwoven caste, religion, gender and sexuality-based violence both online and offline. They contributed towards the development of the Uli datasets and the plugin to filter out offensive words/phrases. Uli is not imagined as the final product, but rather a simple tool, a chisel, that allows people to come together, share stories, and reflect on the future we all want to see.

Participatory process

The project started as an exploration of building datasets and AI aligned with feminist principles. There are many terms- such as responsible, ethical, trustworthy, participatory, feminist- to describe better or alternative visions of AI (and technology). Core to these ideas is the principle of centering the communities who are most impacted by the technology. In our case it is women and marginalized genders at the receiving end of OGBV.

The Uli project contributes unique insights on the nuts and bolts of participatory data and AI. This presentation describes the challenges of converting feminist principles to practice:

With every iteration we learn something new. The methodology of building the flagship dataset on misogyny in India languages with experts with lived experience of abuse is documented here. The paper won the outstanding paper award at the Workshop on Online Abuse and Harms at NAACL, 2024. In another paper on participation in AI we introduce the concept of horizontal and vertical translation of expertise for effective participation. When building out AI safety benchmarks for MLCommons, we learnt where lived expertise could interplay with synthetic/ at scale dataset creation.