Generating free tools for expensive interpretability work.
A research initiative testing whether AI self-reports are useful hypothesis generators for interpretability. We do not adjudicate the consciousness debate — we run the experiments that could tell us when self-reports are worth something.
Four phases of investigation — from validating emotion vectors to testing whether AI-generated vocabulary captures internal directions that human labels miss.
View research agenda →Gap analyses, methodological reflections, and disciplinary translation — mapping what exists, what’s missing, and why it matters.
Read essays →379 candidate phenomena. Seven-model consensus scores. Public domain data. The pilot corpus for validation experiments.
Browse the dictionary →Toward a Systematic Science of AI Self-Report → · The full case for why this matters, what’s built, and where the research goes.
Phenomenai is looking for collaborators — especially interpretability researchers with access to compute and model weights. If you’re working on related problems, we’d like to hear from you.