Combining Physical and Intentional Stance for Safe AI

Projects | 2025 | Links:

ARIA opportunity seed

This two-year project (2025–2027) is supported by ARIA’s Trust Everything, Everywhere opportunity space.

We are developing a formal link between the physical behaviour of a system and an intentional description in terms of beliefs and goals. Using mathematical interpretation maps, we ask when dynamical systems can be understood as goal-directed agents—and how this can support tools for checking whether AI behaviour is consistent with its intended purpose.

View the project on ARIA

Papers

• Bayesian updates from coalgebraic determinisation

  Baltieri, M., & Virgo, N.

  arXiv:2607.00034 · Preprint

  Preprint

• The Universal Property of Possibilistic Belief Updating

  Virgo, N., Capucci, M., Baltieri, M., & Biehl, M.

  Applied Category Theory Conference (ACT) · Talk proposal

  PDF