Destabilize AI Seminar - Shazeda Ahmed​​

02-Oct-2025 | TU Delft TPM-Arena / Online

Talk Abstract | The Social Life of 'AI Alignment'

Over the last three years, dominant narratives accounting for the future of large language models have been split between two polar opposites: utopian artificial general intelligence (AGI) that boosters believe will have resolved the greatest challenges facing human life, or dystopia if said AGI becomes 'misaligned' from 'human values.' Within the field of AI safety that has emerged in response to this binary sociotechnical imaginary, the subfield of AI alignment purports to wield technical solutions and reconfigured social scientific methodologies that can usher in the best case scenario for AGI. Major AI companies house alignment teams, nonprofit centers contribute to alignment research and community-building, and academic scholarship on the subject is growing. Amidst this flurry of alignment activity, how can science and technology studies (STS) scholars interrogate the core assumptions of 'the alignment problem' and the extent to which its proposed solutions enact promises of aligned AI? In this talk, Dr. Shazeda Ahmed will present two collaborative in-progress research papers. "The Work of 'Aligning' Artificial Intelligence" (led by Dr. Ahmed) is an interview study with alignment researchers around the world, seeking to understand the day-to-day labor, research questions, challenges, and outcomes of alignment research across a variety of organizations in industry, academia, and nonprofits. "Toward a Theory of Value in AI" (led by Dr. Abeba Birhane) is a discourse analysis of top AI alignment research papers that touch upon the idea of aligning AI with 'human values.' What (if any) conceptions of values are articulated in this body of work, and what means are used to actualize and evaluate them? Across both studies, Dr. Ahmed will explicate the political stakes of AI alignment amidst anti-regulatory alliances between major governments and tech firms.