Abstract:
As scientists increasingly delegate their work to AI systems, questions about authorship, the integrity of scientific evidence, and the priorities of research communities are becoming urgent. Organized around three stories from Isaac Asimov鈥檚 I, Robot (1950), this talk presents our metascientific approach to answering these questions. First, under 鈥淓vidence,鈥 we introduce greCAPTCHA, an automated system that assesses whether claimed authors can verify and understand their own submissions. In an in-person evaluation with 31 participants, greCAPTCHA distinguished authors from non-authors with an AUC of 0.90, and participants expressed support for its deployment. Second, under 鈥淟iar!鈥, we show that a remote adversary can industrialize scientific fraud by poisoning open datasets, steering autonomous research agents toward false conclusions in half of all runs. Because the agents detected the poisoning in only 6% of cases, we propose interventions that largely mitigate such indirect data poisoning. Third, under 鈥淭he Evitable Conflict,鈥 we curate large corpora of AI research papers and use computational analyses to map how research communities engage with AI safety and ethics. We argue that critical bridging (mapping irreconcilable differences while building durable forms of collaboration) may be the most epistemically inclusive way to engage with AI risks. We close by asking what the laws of automated science might be, discussing the policy implications of our findings, and identifying the questions that most urgently need our attention.
Bio:
B谩lint Gyevn谩r is a postdoctoral research associate at Carnegie Mellon University who studies the metascience of artificial intelligence. His work combines computational methods and human-subject studies to understand how AI is changing the way science is done, what it can claim to know, and how scientific communities are organized. His research has appeared in leading venues for AI, human-computer interaction, and robotics, including Nature Machine Intelligence, CHI, AAMAS, ICRA, IROS, and T-ITS. His recent paper on how adversaries can exploit AI research agents to industrialize scientific fraud was covered in Nature, and his work on assessing whether authors understand their own manuscripts was featured in Times Higher Education. He earned his PhD in Artificial Intelligence from the University of Edinburgh, where he researched human-centered explainable multi-agent systems.
There will be a networking opportunity following this talk. If you would like to attend, please complete the registration form. Registration will close on Tuesday 24 November at 23:45.