Lessons from AI Ethics and Language Research

Language models can generate fluent text and can increasingly be adapted for many Natural Language Processing tasks, such as dialogue systems. In this talk, we'll explore problems we've worked on at DeepMind that interact directly with fairness and social impact, and highlight the importance of developing a sociotechnical practice. Discussing lessons we've learned through our work detoxifying language models, its impacts on bias, and share our thinking about research ethics and governance in the context of the AI research we conduct.