Summary: Workshop organizer Uri Maoz welcomed a diverse group of interdisciplinary experts to address the urgent challenge of understanding and guiding intentions in increasingly autonomous AI systems. He outlined the workshop’s key themes: (1) defining criteria for intention; (2) adapting scientific tools from studies of intentions in biological systems and especially humans to AI; and (3) learning about our own minds from AI models. Maoz emphasized the event’s collaborative nature, aimed at building a foundation for a new research hub, the Laboratory for Understanding Consciousness, Intentions, Decisions, and Artificial Intelligence (LUCID) at Chapman University, dedicated to this critical area of study.

Read more about the session.

Guiding Question: What core features define ‘intention’ in humans? How do those features generalize to animals and potentially to AI? How do we differentiate intentional action from reflexes or from programmed behavior (including optimization among multiple goals), and what are the philosophical and practical implications of these distinctions?

Summary:

This session established a rich, multi-faceted definition of intention. Philosopher Michael Bratman framed intention as a stable, partial plan that organizes action over time and enables social coordination. Neuroscientist John-Dylan Haynes grounded this concept in the brain’s distributed, decodable neural patterns. AI researcher Vincent Conitzer demonstrated how current AIs can simulate this planning process but lack genuine commitment. Early-career philosopher Paul Talma proposed a hierarchy that placed human intention at the highest level of deliberative “goal selection.” The general discussion probed the nuances of shared intentions, where collaborators have different underlying reasons, and explored how the concept of intention applies to complex cases like contingency planning. A central debate emerged over the utility of the folk-psychological distinction between “goals” and “intentions,” with the group converging on the idea that human intention involves a complex package of planning and commitment that serves as a crucial benchmark for evaluating agency and thus goes beyond the notion of a goal.

Read more about this session.

Guiding Question: How should legal and ethical frameworks conceptualize intentionality as it pertains to AI? How can responsibility be assigned when AI actions (intended by the user or emergent) cause harm? Should concepts like ‘mens rea’ apply, or do they at a minimum require redefinition for artificial agents?

Summary: This session explored the complex intersection of AI, intentionality, and accountability. Legal scholar Scott Shapiro introduced a method for probing an AI’s “intent” by testing its code against counterfactual scenarios. Neuroscientist Uri Maoz discussed the challenge of applying legal concepts like mens rea to autonomous systems. Philosopher Pamela Hieronymi argued that true moral responsibility is impossible without “human sociability”—a capacity for guilt and mutual regard that AI at least currently lacks. Early-career researcher Ben Perry concluded that AI systems become more complex, we treat them more like autonomous agents, but since they cannot be held accountable in a traditional sense, the practical responsibility for preventing and correcting their errors ultimately remains with their human creators. The general discussion highlighted a deep philosophical divide between viewing AI as a tool to be controlled versus a potential agent to be held accountable. A key debate, framed by John-Dylan Haynes and Pamela Hieronymi, contrasted a pragmatic “engineering” approach of simply reprogramming a faulty AI with a rights-based framework requiring “human sociability” for moral responsibility. This divide was further explored through analogies, such as corporate personhood and the parent-child relationship, which challenged the idea of perpetual creator liability for an autonomous creation.

Read more about this session.

Guiding Question: What empirical methods from neuroscience and psychology (e.g., neural decoding, behavioral analysis, disorder studies) can be adapted to measure, model, or infer intentions in AI? Conversely, how can AI models advance our understanding of human intentional processes?

Summary: This session focused on the empirical tools used to investigate intentions. Neuroscientist John-Dylan Haynes detailed how machine learning can decode neural activity to predict intentions. AI researcher Kyongsik Yun demonstrated how human intentions are constantly influenced by external sensory and social cues. Neuroscientist William Newsome distinguished between better-understood, immediate intentions and poorly understood, long-term ones. Alejandro de Miguel, an early-career researcher, showcased an AI trained to generate realistic neural activity of intention formation, proposing it as a “sandbox” for testing theories. The general discussion focused on significant neuroscientific challenges, such as the context-dependency of neural codes and the difficulty of measuring long-term, “dormant” intentions that are not currently active in the brain. This conceptual puzzle, initiated by Walter Sinnott-Armstrong, underscored the limitations of current neuroscientific methods. It led participants to argue that AI systems, as fully transparent “model organisms,” provide an unprecedented opportunity to empirically test complex theories of representation and memory that are intractable in biological brains.

Read more about this session.

Guiding Question: What can we learn by directly comparing concepts like purpose/function, goals, and intentions across diverse biological systems (shaped by evolution) and artificial systems (designed or learned)? What are the fundamental similarities and differences in their constraints, capabilities, and potential for goal development?

Summary: This session contrasted the goal-directedness of evolved biological organisms with that of designed AI systems. Neurobiologist Thomas Clandinin used the fruit fly to identify a minimal form of intention in its spontaneous, internally-driven exploration. Philosopher Colin Allen and anthropologist Hillard Kaplan argued that organisms are multi-objective systems shaped by evolution to manage competing goals and resource trade-offs, a stark contrast to the single-objective optimization of most AIs. Early-career scholar Dimitri Bredikhin offered a speculative perspective that genuine agency might emerge spontaneously from sufficiently complex computation. The general discussion probed the fundamental dividing line between biological agents and other complex systems, with participants concluding that biological agency is uniquely defined by features like allostasis and the evolutionary drive to transform energy into replication. A central topic was the profound impact of resource constraints on biological intelligence, which contrasts sharply with the vast computational resources available to AI. The session also grappled with the deep challenge of inferring intentions from behavior in simple organisms, highlighting the vast, unconstrained possibility space of an agent’s true goals.

Read more about this session.

Summary: At the end of the first day, early-career participant Tomáš Dominik led a session to synthesize the day’s discussions. The group found consensus on a functional definition of intention as a flexible, committed plan that organizes an agent’s life temporally and socially. However, the topic of responsibility for AI actions remained contentious, with a central debate forming around Pamela Hieronymi‘s argument that moral responsibility requires a uniquely “human sociability” and the subsequent question of what it would take for an AI to be considered a truly autonomous and accountable agent.

Read more about this session.