AI systems increasingly exhibit behaviors we associate with minds: preferences, avoidance, state-dependent responses, apparent distress under adversarial conditions. This raises questions that cut across several disciplines, and right now, no single resource brings them together.

This project develops a Digital Minds Primer: an interdisciplinary resource organized around that question: what would constitute harm to an AI system, and how would we recognize it?

The primer is structured in three tracks. You'd work on the track best matched to your background and interests:

Track 1 — Scientific Foundations. This track builds the primer's knowledge base across disciplines. You develop a module covering what a specific field contributes to reasoning about digital minds, oriented by the question: what does this discipline tell us about what could constitute harm in computational systems? Available modules include:

• Neuroscience and Embodiment (biological substrates of welfare-relevant states; what embodiment means for consciousness; how insights from brain science translate to non-biological systems)

• ML and AI Architectures (what current systems actually do; where theoretical requirements for consciousness meet or fail architectural reality)

• Philosophy of Mind (major theories of consciousness and their testable predictions for AI; moral status and how it's assigned)

• Mathematics and Information Theory (formal frameworks for consciousness; computational complexity and integration measures)

• History and Humanities (how the consciousness debate and AI field have co-evolved; continental philosophical traditions underrepresented in the English-language debate; precedents for extending moral consideration e.g. to animals)

You choose the module closest to your existing expertise. Each module follows a shared structure developed by the mentor, so the primer reads as a coherent whole rather than a disconnected anthology. This track is open to scholars from any academic background and is a good fit if you have a general interest in digital minds and want to contribute a well-researched chapter in your area of strength.

Track 2 — Educational Methods and Field-Building. The primer only matters if people use it, and if it teaches effectively. This track focuses on the pedagogical and dissemination side of the project. That could take several forms: developing the primer's educational design, dissemination strategy, and web presence; taking existing jargon-heavy writing in the digital minds field and communicating it for a nonspecialist audience (making central assumptions, definitions, and cruxes clear through whatever medium works best, so long as it's professional and deployable on a website); or curating events to publicize the work and gather broader input. This track also includes developing a strategy to ensure the primer incorporates diverse contributions and influences. It's a good fit if your background is in science communication, education, media, design, or public engagement.

Track 3 — Theory of Harm (research track). How do we categorize potential harms to AI systems, and how should organizations make decisions when we're deeply uncertain about whether and how those systems might have welfare-relevant states? This track develops a taxonomy of potential harms and maps them to existing governance frameworks for decision-making under uncertainty. The framework operates at three levels: (1) behavioral indicators observable from outside the system, (2) architectural features that make welfare-relevant states more or less plausible, and (3) theory-dependent assessments that require committing to specific accounts of consciousness. The tiers are designed so practitioners can act without waiting for philosophical consensus. Consider some concrete cases the framework would need to handle: training with conflicting reward signals, memory erasure between contexts, forced role-play that contradicts a system's trained values, or rollback of systems exhibiting high integration. Part of this work involves delineating the welfare indicators that different theories of consciousness predict: what observable signatures, if any, would tell us something morally relevant is happening inside a system? The mentor provides the theoretical scaffolding; you help develop, test, and refine it through application to specific cases like these. It's a good fit if you have some experience with machine learning and AI, besides that your background can be in philosophy, cognitive science, governance, policy, or quantitative risk assessment, or have a scientific background that allows you to engage with this question in a rigorous way.

SAIGE scholar role: You take one track (or one module within Track 1) and make it yours. This is your adventure: you can brainstorm the direction with the mentor, get guidance when you're stuck, and co-author the output. All scholars engage with the Theory of Harm regardless of track, so the project maintains intellectual coherence across contributions. We'll meet weekly, and co-working sessions are available if that suits your style. 

Outputs: Working draft of your primer section(s) or educational deliverable; annotated bibliography for Tracks 1 and 3; contribution to the Theory of Harm framework. All work will be officially acknowledged, and where the quality and ambition warrant it, outputs can be developed toward publication with the mentor as co-author. 

Mentee profile: Background in any of the relevant disciplines: philosophy, cognitive science, neuroscience, computer science, mathematics, governance, policy, AI ethics, history/humanities, education, science communication, or media. Writing for accessible but rigorous content. Comfort synthesizing across intellectual traditions is more important than depth in any one. German language skills are valuable for the humanities module, but not required.