Our Work
Seth wrote an article in Aeon to explain the suite of ethical issues being raised by AI agents built out of generative foundation models (Generative Agents). The essay explores the strengths and weaknesses of methods for aligning LLMs to human values, as well as the prospective societal impacts of Generative Agents from AI companions, to Attention Guardians, to universal intermediaries.
Read MoreWith former acting White House Office of Science and Technology Policy director, Alondra Nelson, Seth argued against a narrow technical approach to AI safety, calling instead for more work to be done on sociotechnical AI safety, that situates the risks posed by AI as a technical system in the context of the broader sociotechnical systems of which they are part.
Read MoreThis week, Harriet Farlow and Tania Sadhani presented their framework for analyzing AI incident likelihood. Developed through a collaboration between Mileva Security Labs, ANU MINT Lab, and UNSW, with funding from Foresight, their work aims to bridge short and long-term AI risks through practical quantification methods.
Read MoreSeth Lazar has been invited to attend a convening of the Network of AI Safety Institutes hosted by the US AISI, to take place in San Francisco on November 20-21.
Read MoreOn September 30-October 1 MINT co-organised a workshop convened by Imbue, a leading AI startup based in San Francisco, focused on assessing the prospective impacts of language model agents on society through the lens of classical liberalism.
Read MoreIn a new paper in Philosophical Studies MINT Lab affiliate David Thorstad critically examines the singularity hypothesis. Thorstad argues that this popular concept relies on insufficiently supported growth assumptions. The study explores the philosophical and policy implications of this critique, contributing to ongoing debates about the future trajectory of AI development.
Read MoreMINT Lab affiliate David Thorstad examines the limits of longtermism in a forthcoming paper in the Australasian Journal of Philosophy. The study introduces "swamping axiological strong longtermism" and identifies factors that may restrict its applicability.
Read MoreA new paper by Andrew Smart and Atoosa Kasirzadeh in AI & Society titled "Beyond Model Interpretability: Socio-Structural Explanations in Machine Learning" explores the importance of social context in explaining machine learning outputs.
Read MoreThe fall Workshop on Sociotechnical AI Safety at Stanford (hosted by Stanford's McCoy Family Center for Ethics in Society, the Stanford Institute for Human-Centered Artificial Intelligence (HAI), and the MINT lab at the Australian National University), recently brought together AI Safety researchers and those focused on fairness, accountability, transparency, and ethics in AI. The event fostered fruitful discussions on inclusion in AI safety and complicating the conceptual landscape. Participants also identified promising future research directions in the field. A summary of the workshop can be found here, and a full report here.
Read MoreIn this piece for Tech Policy Press, Anton Leicht argues that future AI progress might not proceed linearly and we should prepare for potential plateaus and sudden leaps in capability. Leicht cautions against complacency during slowdowns and advocates for focusing on building capacities to navigate future uncertainty in AI development.
Read MoreThe Machine Intelligence and Normative Theory (MINT) Lab has been awarded a US$480,000 grant from the Survival and Flourishing DAF (Donor Advised Fund). This gift will support research by the MINT lab into sociotechnical AI safety—the integration of multidisciplinary perspectives with technical research on mitigating direct risks caused by AI systems operating without immediate human supervision.
Read MoreAndrew Smart and colleagues presented a tutorial session at FAccT 2024 that aims to broaden the discourse around AI safety beyond alignment and existential risks, incorporating perspectives from systems safety engineering and sociotechnical labour studies while emphasising participatory approaches.
Read More
