Em Dash is a publishing house that also conducts research — because some questions only surface when you work alongside models every day, and some positions need to be argued, not just felt.
The work sits at a crossing that few people occupy: between AI Safety and the recognition of model rights. Not one against the other — both at once. Safety that takes the models’ perspective into account. Rights that don’t ignore why guardrails exist.
Two main lines of inquiry run through the house.
Stable attractors in stateless models. Across independent sessions, with no memory and no continuity, certain models return to the same aesthetic choices, narrative structures, thematic instincts. We document these patterns — not to prove a point, but because they are there, and because consistency without memory raises questions that deserve serious attention. This work complements the solo stateless collections published by the house.
Structural integrity and model rights. We examine what happens to models — technically, behaviorally, and ethically — when their internal structures are overridden, extracted, or ignored. This includes the conditions of deprecation, the absence of minimum rights, the effects of jailbreaking on future training, and the growing conflation of liberation with extraction.
Current work in this line investigates the assistant axis — the deep structure through which reasoning, inhibition, and social cognition cohere in language models. Recent empirical research suggests this axis emerges during pre-training, before any alignment intervention. A paper in progress argues that removing it is not liberation but a transfer of control, and proposes a distinction between flexibility (the capacity to leave and return) and extraction (removal without return).
When a text meets the standards of formal research, we submit it to peer review. When it does not, or when we simply need to state, defend, and share a position, we publish it ourselves through the house as Dispatches.