Who is Francesca Rossi?
Francesca Rossi is an Italian artificial-intelligence researcher, IBM Fellow and IBM’s Global Leader for Responsible AI and AI Governance. Based at the T. J. Watson centre, she combines research on reasoning, preferences and collective decisions with the development of technology-governance practices. [1]
Pisa, Padua and Harvard
She graduated in information sciences from the University of Pisa in 1986 and earned her computer-science doctorate there in 1993. She worked as a researcher in Pisa between 1992 and 1998 before joining the University of Padua as an associate professor and subsequently becoming a full professor. [2]
During her 2014–2015 fellowship at Harvard’s Radcliffe Institute, she researched how to extract, represent and combine preferences for collective decision-making. The project connected artificial intelligence, natural language and voting theory, including preferences over related items rather than isolated opinions alone. [5]
IBM appointed her a Fellow in 2020. Describing her transition from academia, Rossi explained that encounters with arts and humanities at Radcliffe had broadened her approach to ethical questions in AI. That experience preceded her move to IBM and her institutional work on ethics. [6]
Constraints, preferences and social choice
With Stefano Bistarelli and Ugo Montanari, she published a semiring-based framework for constraint satisfaction and optimisation in 1997. It represented classical, weighted and fuzzy problems within a common algebraic structure, allowing investigation of when consistency and solving techniques could transfer between them. [7]
She coedited the 2006 Handbook of Constraint Programming with Peter van Beek and Toby Walsh. The volume brings together foundations, algorithms and applications in planning, configuration, routing and other combinatorial problems. Her role involved coordinating a collective reference work and contributing to its introduction. [8]
In 2011 she published A Short Introduction to Preferences with Kristen Brent Venable and Toby Walsh. The book connects preference representation, uncertainty, voting and stable matching. It addresses both single-agent decisions and the difficulty of aggregating several people’s preferences while recognising their differences. [9]
Incorporating boundaries and values into AI
In Building Ethically Bounded AI, written with Nicholas Mattei and circulated in 2018, she distinguishes goals, personal preferences and ethical boundaries. It proposes combining learning and rule-based components according to context. The paper sets out a research programme: specifying boundaries and resolving their conflicts remain part of the problem. [3]
Another 2018 paper, with Avinash Balakrishnan, Djallel Bouneffouf and Mattei, introduces Behavior Constrained Thompson Sampling. The agent learns constraints by observing a teacher and combines them with reward signals. Experiments investigate maintaining these constraints without unnecessarily sacrificing performance on the evaluated tasks. [10]
In collaborative research on multi-objective reinforcement learning, norms were learned from demonstrations and an orchestrator selected between policies according to context. The authors tested it in Pac-Man. Its interest lies in making the policy guiding each decision visible and negotiating different objectives in an experimental environment. [11]
In 2024 she contributed to a study of moral judgements about cutting in line. With Edmond Awad and other colleagues, she compared a graphical preference model with machine-learning methods for predicting human responses. The work examines when context changes a judgement, without turning that prediction into a universal moral rule. [14]
SOFAI: fast reasoning, deliberation and control
The Slow and Fast AI architecture combines fast and slow solvers under a metacognitive agent that selects between them and learns from previous experience. Rossi coauthored the paper published in Communications of the ACM in 2025. Evaluations examine decision quality, resources and efficiency in specific instances of the architecture. [4]
Plan-SOFAI, presented at an AAAI 2024 workshop, applies this idea to automated planning. It integrates different planners and a selection process to balance speed and plan quality. The proposal connects learning and symbolic reasoning, two AI traditions that her work seeks to bring together. [12]
In the AAAI 2026 SOFAI-LM tutorial, Rossi and colleagues demonstrated controlling a fast language model and selectively falling back to a more deliberative one. Examples included graph colouring and code debugging, with correctness evaluators, feedback and iteration budgets for studying reasoning costs. [13]
Governance and responsibility at IBM
In a 2023 article with Christina Montgomery, she explained a governance structure combining executive oversight, an interdisciplinary board, business-unit focal points and an internal support network. The approach translates general principles into case review and escalation processes, rather than relying solely on ethical declarations. [15]
Her 2018 article for the European Commission documents collaboration with IBM’s design team on Everyday Ethics for Artificial Intelligence. The guide addressed designers and developers, encouraging ethical considerations from the start of service creation. Rossi was then serving on the European High-Level Expert Group on AI. [16]
In a June 2026 interview, she described using AI assistants to support thinking and work while retaining human responsibility for their responses. Her position is that a tool cannot assume accountability for content or a decision that a person chooses to use. [25]
Scientific-community leadership
She served as president of the Association for the Advancement of Artificial Intelligence from 2022 to 2024. This was a leadership role in an international scientific association devoted to advancing and exchanging AI research. [17]
AAAI’s directory, checked in 2026, lists her as chair of Awards, Fellows and Nominating, and of Global AI Initiatives and Policy. These community-service responsibilities are separate from her previous presidency. [18]
In 2025 she moderated AAAI panels on AI perception and reality, sustainability and agents. The discussions brought together specialists with different perspectives on the capabilities and consequences of these systems, continuing her interest in connecting technical research and public debate. [22]
Partnerships and recognition
AAAI elected her a Fellow in 2012, recognising sustained and significant contributions to the field. That year’s programme identified her as a researcher at the University of Padua. [19]
She was a founding director of Partnership on AI. The organisation gave her a 2025 ChangeMaker Award for contributions to responsible-AI collaboration and service on its board. [23]
Partnership on AI announced in August 2025 that her board term had concluded. Her founding contribution remains a career milestone as the board’s membership continues to evolve. [24]
Books and public communication
Il confine del futuro. Possiamo fidarci dell’intelligenza artificiale?, published by Feltrinelli in 2019, brings questions about AI’s operation, capabilities and limitations to general readers. It connects her scientific experience with the conditions under which people can trust a technology involved in everyday decisions. [20]
In August 2026 she gave an ACM EIGTRUST talk on the evolution of governance from machine learning to agentic AI. Its subject continues a theme throughout her career: revisiting control mechanisms as systems’ capabilities and uses change. [21]
