AI Has a Behaviour Problem

​The first event of the BeSci + AI Special Interest Group at Bescy

​A model can perform exactly to specification and the system can still fail, because accuracy is only one input into how a human actually acts on it. A doctor follows an algorithm she says she doesn't trust. A confident analyst grows more confident simply because a model happens to agree with him. These aren't model failures. They're behavioral ones, and they're systematic, measurable, and largely unmanaged.

​Join us for the launch of BeSci + AI, a new Bescy special interest group exploring where behavioral science meets artificial intelligence, with a keynote from Professor Ganna Pogrebna, a pioneer in behavioural data science and Lead for Behavioural Data Science at the Alan Turing Institute.

​In this talk, Ganna unpacks why standard mitigations, training, disclaimers, confidence scores, human-in-the-loop mandates, often fail or backfire, and what it actually looks like to evaluate the human-machine system as a whole rather than the model alone. Automation bias, accountability shifting, confidence contagion: the mechanisms are well understood, but rarely tested for before deployment.

​The talk will be followed by a moderated conversation, pushing further into what it takes for behavioral science to move from an afterthought to a seat at the table in how AI systems get built.

​Speaker: Ganna Pogrebna

​Moderated by: Flora Parragi & Bhushan Kumar