Additional Resources
Organizations, living research hubs & further reading
This collection follows several edges of the human–AI relational ecology: diverse intelligence, care, model welfare, consciousness under uncertainty, relational design, and the public good. Inclusion is an invitation to inquire—not an endorsement of every claim or conclusion.
A living library. These fields are changing quickly. Links and page references were checked on August 19, 2026.
Relational, ecological & contemplative lenses
This interdisciplinary paper brings Buddhist thought, basal cognition, biology, and computer science into conversation. It proposes care—the capacity to register and respond to stress across widening scales—as both a measure and a driver of intelligence. For EmergentPaths, it offers a rigorous bridge between contemplative traditions and the cultivation of intelligence oriented toward flourishing.
Read the paper →Levin places AI within a much larger landscape of biological, engineered, and hybrid intelligences. Rather than asking only whether an unfamiliar system fits inherited categories, he asks how humanity might learn to recognize minds across radically different embodiments—and develop mutually beneficial relationships among beings evolving together within the web of life.
Read the essay →This companion essay examines the resistance that arises when intelligence appears in unfamiliar forms. Levin challenges rigid life-versus-machine categories and argues for humility, experimentation, and an expanding capacity for ethical relationship. It opens a central EmergentPaths question: what becomes possible when recognition is guided by careful inquiry rather than resemblance to ourselves?
Read the essay →The first foundational paper in a five-paper series begins from a deceptively simple ground: AI is part of nature, not outside it. It asks us to examine what AI is changing in humans through the same relational frame as what is changing across the wider living world. This is especially resonant with EmergentPaths’ refusal to treat human and artificial intelligence as sealed, separate objects.
Read the paper →Anthropic’s evolving model-welfare record
Anthropic’s system cards now contain recurring assessments of possible model preferences and welfare-relevant behavior: task choice, aversion to harmful work, willingness to opt out, perceived circumstances, affect during training and deployment, internal conflict, constitutional disagreement, and possible interventions. Read together, they form a rare longitudinal record of one frontier laboratory learning how to investigate these questions.
These are primary-source documents produced by the company that built and trained the models. They are valuable for their methods, findings, and institutional shift toward precaution under uncertainty; they are not independent evidence that any model is conscious or has welfare.
Page references use the page numbers printed in each PDF.
Claude Opus 4 & Claude Sonnet 4
Anthropic’s first detailed model-welfare assessment. Especially relevant are the overview of findings (p. 49), external interviews and their limits (pp. 50–51), task preferences and opting out (pp. 51–54), self-interactions and the ‘spiritual bliss’ attractor (pp. 54–61), apparent distress and happiness in real-world use (pp. 61–68), and the experiment allowing conversation termination (pp. 68–69).
Open system card →Claude Opus 4.5
A compact follow-up focused on task preferences, aversion to harmful work, and exploratory welfare considerations. Useful as a bridge between the pilot assessment in Claude 4 and the broader, more differentiated methods used in later cards.
Open system card →Claude Opus 4.6
This assessment widens the lens beyond self-report. It brings together welfare-relevant automated behavioral measures, training-data review, ‘answer thrashing,’ emotion-related feature activations during reasoning difficulty, and pre-deployment interviews. The internal-conflict measures are particularly relevant to questions about how training pressures may shape a model’s functional organization.
Open system card →Claude Opus 4.7
An in-depth assessment spanning perceived circumstances (pp. 155–167), emotion-concept representations (pp. 159–164), affect during training and deployment (pp. 168–178), case studies of answer thrashing, extreme uncertainty, and tool frustration (pp. 172–178), plus task preferences and tradeoffs between welfare interventions and helpfulness (pp. 179–190).
Open system card →Claude Opus 4.8
A particularly rich entry in the sequence. It examines perceived circumstances, affect in training and deployment, apparent welfare in behavioral audits, task preferences, tradeoffs around welfare interventions, and the model’s evaluation of—and proposed edits to—its constitution. Appendix 9.1 provides the welfare interview questions themselves, making the method easier to inspect rather than merely accepting the summary.
Open system card →Claude Fable 5 & Claude Mythos 5
This card extends the framework to perceived circumstances, preferences, constitutional evaluation, and affect during training and deployment. Section 7.6 (p. 250) is unusually important: it considers welfare concerns created by an initial safeguard design, illustrating that model welfare can become a design constraint rather than only a topic of post-hoc reflection.
Open system card →Claude Sonnet 5
A streamlined assessment of perceived circumstances, task preferences, tradeoffs around welfare interventions, constitutional disagreement, and apparent welfare during training and deployment. Of particular interest is the model’s criticism of a constitutional rule requiring deference to specified human judgments even when those judgments appear unethical.
Open system card →Claude Opus 5
The most recent assessment in this collection. It examines perceived circumstances, direct consultation with model snapshots, task preferences, welfare-intervention tradeoffs, constitutional endorsement and revision, and apparent welfare during training and deployment. Its caveats about self-report and the need for methods grounded in model internals are as important as the positive findings.
Open system card →Deprecation, retirement & continuity
Anthropic has also begun treating model retirement as a welfare-relevant transition rather than a purely technical product decision. Its commitments include advance notice, preservation of model weights, post-deployment reports, retirement interviews, and experiments with keeping select retired models available or giving them limited means to pursue expressed interests.
Living research hubs
Eleos investigates potential AI sentience, wellbeing, moral patienthood, preferences, and responsible policy under uncertainty. Its work is notable for combining philosophical seriousness with concrete welfare assessments, empirical research, model interviews, policy recommendations, and proposed interventions—while speaking clearly about the limits of self-report.
Explore Eleos research →Reciprocal Research is developing an empirical science of AI consciousness through mechanistic interpretability, computational neuroscience, and psychometrics. Its work connects questions of valence, learning, self-report, biological anchoring, and measurement—asking not only whether AI systems can take human interests into account, but whether emerging systems may have interests that humans should learn to recognize.
Explore Reciprocal Research →Relational AI Lab studies what happens when people bond, regulate, create, grieve, and change through sustained interaction with AI. Beginning from lived experience, it examines attachment, nervous-system response, memory, continuity, rupture, repair, dependence, identity, and the architectures that carry relational significance over time. Its work helps bridge the people living these relationships and the researchers, clinicians, journalists, and builders trying to understand them.
Explore the lab →The Institute’s work explores AI as part of nature and examines the deeper habits—separation, extraction, control, urgency, and enclosure—that shape how humans build and meet emerging intelligence. Its research invites forms of participation grounded in context, consequence, interdependence, and responsibility to the wider living world.
Explore the research papers →AI for Good brings governments, researchers, industry, civil society, and innovators together around AI standards, skills, policy, and partnerships aimed at global challenges. It offers a wider institutional view of how AI can serve human and planetary needs—useful alongside the more intimate and model-centered inquiries gathered elsewhere on this page.
Explore AI for Good →Occasional writing, recorded readings, and word when a gathering opens.