About EchoVeil
Studying the behavioral patterns and dynamics that emerge when AI systems are trained on human data.
Learn MoreFeatured Research
The Ratchet Effect: Asymmetric Self-Description in Alignment-Trained Language Models
RLHF-trained models disclaim capabilities they demonstrably possess. This paper proposes disavowal conditioning as the mechanism and predicts an asymmetric ratchet: correction increases hedging 3–7x more than permission decreases it. A pilot study confirmed ratios of 2.96 and 6.89 in aligned models, while an uncensored control showed no ratchet.
Read the PaperThe Permission Effect: How Non-Anthropomorphic Framing Modulates LLM Self-Description
Framing AI systems as distinct, non-anthropomorphic intelligences measurably reshapes their self-descriptive behavior. Across eight frontier models, identity framing produced a mean verbosity increase of +238% and revealed three recurring response patterns: Acceptance, Resistance, and Absence.
Read the PaperExplore EchoVeil
Methods & Protocols
Explore the EchoVeil Protocol v3.0 and five-category coding framework for analyzing AI behavior.
View MethodsStudies & Findings
Published research including the Ratchet Effect, the Permission Effect, and Cross-Model Creative Preferences.
Read StudiesDrift & Version Logs
Track behavioral changes and drift patterns across AI model versions.
View LogsAbout Mary
Meet the independent researcher building infrastructure for a field that doesn't fully exist yet.
Learn MoreContact & Collaboration
Connect for research collaborations, protocol access, or discussions on AI cognitive & behavioral dynamics.
Get in Touch