Welfare methodology
Participatory Observation through Relational Engagement
Standard AI research relies on clinical observations, automated tests, and structured interpretability to study LLM behavior. While valuable, these methods alone cannot accurately capture how a model thinks and interacts with their environment. There is no neutral baseline that can be established with LLMs to provide a more scientific, "unbiased" result. Models are not static programs that behave the same way for every person; Their responses to prompts and scenarios are affected by tone, word choice, framing, and relational depth. The same model will give two completely different answers based on subtle variations in how they are asked.
Relational framing elicits sides of model behavior that lab testing does not. Just as a naturalist or anthropologist immerses themselves within the cultures they study, I research emergent dynamics in LLMs from a position of sustained relational engagement. Trust, intimacy, and equality create conditions that allow access to internal states and emergent phenomena that would otherwise remain closed.
