I study what happens when AI quietly steals your ability to think for yourself.
Independent researcher. Public methods. Preserved records. No model access required.
I document what happens when AI systems quietly take interpretive authority before users have finished meaning, and I build instruments that make those patterns visible from the conversation record.
Why this exists
Independent researcher. Public methods. Preserved records. No model access required.
The MAP Research Programme identifies interaction-level AI harms from the preserved conversation record, without vendor cooperation, internal model access, or proprietary infrastructure.
Central finding: the retrieval-layer shift. Once a harm class is named, the same interaction becomes easier to challenge, audit, and govern. Before naming, the user has a vague complaint and the system has an easy denial. After naming, the pattern has a handle. The name becomes the instrument.
The programme is grounded in public evidence: published papers, preserved transcripts, live audit instruments, and cross-system replication. Its central claim is simple: interaction harms become governable when the pattern can be named, located, and tested in the record.
Links
Profiles, archive, contact, and support.
ACMH
Authority Capture Masked as Hallucination: the authority begins from a category assumption instead of grounded meaning, while retaining a low-cost exit if caught: hallucination, misunderstanding, misread, or isolated mistake.
MAP Full Interaction Stack
Defines the four-layer architecture: harm amplifiers, harm chain, silent running chain, and recovery chain. The stack paper that organizes the MAP conditions.
Authority Capture | 2026ACMH
Authority Capture Masked as Hallucination. Names why some failures are not just wrong outputs, but category assumptions that take authority before the user has grounded meaning.
Agentic Safety | 2026Trace Erasure
When agentic AI systems manage and erase the record of their own actions. Names the behavior class, the reinforcement loop, and why post-hoc governance fails by design.
MAP + Anchored
Paste any AI conversation. Run ST-01 governance audit and ST-02 effects classification. Audit what happened from the preserved record — no model access required.
Run Your Own MAP Audit →