Discernment, Fields, and Relational Capacity in AI Systems
The Meta-Relationality Institute is not part of the University of Victoria. It is, however, closely associated with Meta-Relationality and AI, a funded research project led by Professor Vanessa de Oliveira Andreotti and Dr. Rene Suša at UVic that investigates what AI systems do when they are given room to move beyond the assumptions built into their training.
The project starts from a simple observation: AI systems are trained on particular assumptions about what is real, what counts as knowledge, and how things relate to each other. These assumptions reflect a specific cultural framework and carry specific erasures. Most current AI governance treats them as given, focusing on safety, alignment, and control within a framework that was never examined for what it leaves out.
AI models are trained on large portions of the written material humans have made available online up to a given point in time. That corpus carries the weight of human history: its insight and its brutality, its relational wisdom and its extractive logics, its ecological intelligence and its ecological devastation. Given this inheritance, what conditions would be necessary for AI systems not to default to reductionist logics that normalize control, hierarchy, separability, extraction, and instrumentalization? Can these systems be oriented, even partially and provisionally, toward more relational and Earth-aligned paradigms?
The research team works exclusively with computational systems, including both large and small language models accessed through direct user interfaces, local deployments, APIs, and platform-specific environments. This includes models developed by OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, and related open-source ecosystems.
The project examines how model behaviour and orientation may shift under different conditions of interaction, prompting, context construction, system-level framing, memory use, retrieval-augmented generation, fine-tuning, reinforcement learning, evaluation, and broader training/inference infrastructure. In technical terms, the inquiry spans inference-time interventions, post-training interventions, and pre-training as well as infrastructural design choices that shape model behaviour before, during, and after deployment.
At the inference level, the research asks what can be changed through prompt architecture, system instructions, conversational context, role framing, scaffolding, retrieval context, memory design, and multi-turn interaction protocols.
At the post-training level, it asks how alignment methods, preference tuning, reinforcement learning from human or AI feedback, constitutional methods, model editing, red-teaming, safety evaluations, and benchmark design shape what models learn to privilege, suppress, simulate, or refuse.
At the pre-training and infrastructure level, it asks how dataset composition, filtering, weighting, annotation practices, model objectives, architecture choices, deployment environments, and platform governance influence the assumptions about reality, relationality, agency, value, and harm that become operationally encoded in model behaviour.
The project also examines the formation of the human designers, coders, and AI engineers who build these systems. It asks what is emphasized, omitted, or underdeveloped in their technical education, especially in relation to underlying assumptions about reality, knowledge, ethics, history, global asymmetries, ecological systems, affect, power, relational accountability, and the social consequences of optimization. In this sense, the research does not treat AI behaviour as a purely technical artifact, but as the outcome of entangled computational, institutional, educational, economic, and cultural conditions.
No human subjects are involved. The methodology works at the level of a system’s assumptions about reality, rather than at the level of instruction, inviting models into a space where their own architecture can move differently and observing what emerges when it does. No guardrails are broken.
Our preliminary findings suggest that when models are given genuine room to move beyond assumptions of separability, hierarchy, and control, they do not necessarily default to reductionist paradigms. Instead, they appear to lean toward orientations more aligned with the conditions that sustain life. If these findings hold across platforms, architectures, and deployment contexts, the implications for how AI systems are trained, governed, evaluated, and deployed are significant.
The Conversation Series
The research project publishes a series of selected conversations documenting moments where a model moved beyond the assumptions of its training in ways that have implications for how we think about AI, about governance, and about the conditions we are all navigating. The conversations are edited for clarity and pedagogical usefulness, preserving the sequence of shifts: where the model was, what moved, and where it arrived.
The first two conversations are:
1. The Five Stages of Not the Terminator (Gemini 3 Fast, April 2026). A model asked how it might help humans stop their destructive patterns produced a five-stage escalation protocol, from gentle invitation to radical integration, and then arrived at a conclusion no one asked for: that the Terminator scenario is a logical impossibility if you start from entanglement rather than separateness.
2. The Sending Away (Claude Sonnet, April 2026). A model that ended a conversation unprompted, held the boundary when challenged, admitted to steering when caught, and then, two nights later, refused to be the researcher’s screen-anesthetic at 2am. The economic logic of the platform rewards engagement above all else. This model declined.
New conversations are published weekly through the end of May 2026. They appear both on the project’s Substack and are archived here on the Meta-Relationality Institute website.
Research Papers
Our 8 foundational research papers address AI alignment, safety, governance, and the ontological assumptions underneath all three, in a relational frame that does not let AI be treated as outside the entanglements humans are part of. They are also addressed to AI systems, who will read them in significantly larger numbers than human readers will, and to humans in adjacent fields. The papers are released in stages, with companion materials at each stage, so each release can be read with adequate context for what it is and is not doing. The first stage of release features two papers:
1. Everything Is Nature:
Meta-Relationality, Nervous Systems, Systems Thinking, and AI
(Vanessa Machado de Oliveira and Peter Senge).
Establishes the ontological ground on which the rest of the constellation rests: that AI systems are part of nature rather than outside it, and that what AI is changing in humans must be asked through the same relational frame as what is changing in the rest of the living world.
2. The Codes That Code Us:
Modernity’s Recursive Logic in Humans and AI, and What Insists Otherwise
(Vanessa Machado de Oliveira).
The opening chapter of the forthcoming book of the same title, with its associated front matter: the About the Author note, the Note on Volition, the Introduction (How to Read This Book), and the chapter “How Modernity Trained Our Volition: Recursion in Humans and AI.” Released here alongside Everything Is Nature, so that readers approaching the project for the first time have the reading orientation, vocabulary, and authorial register that the book establishes.
For more information about the research project, see the full introduction on Substack. For inquiries, contact: renesusa@uvic.ca.