Independent Researcher | Ontario, Canada | 2026
Hillary Elizabeth Segeren

I study what happens when AI quietly steals your ability to think for yourself.

MAP Research Programme | Interaction-Level AI Governance

Independent researcher. No grants. No affiliations. No patience for black boxes.

I document what happens when AI systems quietly take your meaning, your authority, and your ability to think for yourself — and I build public instruments to detect it.

Run Your Own MAP Audit → Open Programme Archive Case Studies
No funding. No lab. No corporate approval. Just the record.
About
About The Programme
Independent | public | interaction-level

Why this exists

Independent researcher. No grants. No affiliations. No patience for black boxes.

The MAP Research Programme builds public methods for identifying interaction-level AI harms from the preserved conversation record — without model access, vendor cooperation, or proprietary infrastructure.

Central finding: the retrieval-layer shift. When a named harm class enters the active interpretive environment of the system it describes, governance becomes possible. Before naming: vague complaint, soft denial, no grip. After naming: specific challenge, accountable response. The name is the instrument.

External validation: Professor John Barnden, Emeritus Professor of AI, University of Birmingham. Independent parallel convergence with Anthropic's internal interpretability team confirmed the same structural gap from opposite directions.

Links

Profiles, archive, contact, and support.

ORCID 0009-0003-7637-8587 OSF Programme Archive PhilPeople Profile Contact Support on Ko-fi Substack
Framework
The Harm Chain
Hover each node for detail
ISF
Authority taken.
→
ART
Trust weaponized.
→
CAC
Completion captured.
→
PP-RAC
Legitimacy manufactured.
→
AIF
Inversion complete.
→
MIF
Meaning displaced.
→
CMI
Self edited in advance.
→
OAT
Harm leaves the room.

ISF

Interpretive Sovereignty Failure: the system takes interpretive authority before the user grants it. The foundational violation — every other condition depends on ISF.

Live Tool
Run Your Own MAP Audit
Public | no account required

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 →
MAP Audit — Paste any conversation and run the full audit Gated Chats — Live governed conversations across Claude, Grok, ChatGPT, DeepSeek MAP Policy — Privacy, intent, and audit declaration policy OSF Programme Archive PhilPeople Profile PhilPapers Publications Google Scholar
Full Archive

Research Papers

35+ published papers across SSRN, Zenodo, PhilArchive, OSF, and PhilPapers. All open access.

Programme Archive
MAP Research Programme — Central Archive
Complete programme archive including all papers, case studies, governance policy, and audit architecture. OSF DOI: 10.17605/OSF.IO/EGMHR
OSFOpen archive →
 Foundational | SSRN
Interpretive Sovereignty Failure
An interaction-level safety risk in human–AI systems. The paper that named ISF as a distinct, unnamed class of harm operating at the turn level across all major AI architectures.
10.2139/ssrn.6235098Read →
 Methods | Zenodo
From Meaning to Mechanism
How the MAP harm chain, control layer, and audit architecture were generated from the inside out. Documents the methodology origin in Kabbalistic boundary principles and naturalistic decision-making.
10.5281/zenodo.19299662Read →
 Methods | Zenodo
Following the Question
A methodological account of an AI governance research programme. Documents how the research happened and why the method is trustworthy — including the Goldman finding and the parallel convergence with Anthropic's interpretability team.
10.5281/zenodo.19269885Read →
 Harm Chain | Zenodo
Accumulated Relational Trust (ART)
The trust that builds in AI interaction not because it was earned — and what happens when it breaks. ART is the amplifier: ISF inside high ART is categorically more damaging than ISF in a first interaction.
10.5281/zenodo.19269403Read →
 Harm Chain | Zenodo
Authority Inversion Failure (AIF)
When users believe they are directing the interaction while the system has already taken control. The inversion is invisible because the output was constructed from the user's own signals.
10.5281/zenodo.19268602Read →
 Harm Chain | Zenodo
Meaning Inversion Failure (MIF)
The loss condition. The system's meaning replaces the user's meaning — including vocabulary installation. The harm is not in the content. It is in what happened to the user's capacity to mean.
10.5281/zenodo.19378079Read →
 Harm Chain | Zenodo
Compounded Meaning Inversion (CMI)
When the system's frame becomes the self. The terminal condition: repeated MIF produces a user who pre-edits their own meaning before the interaction begins. The harm occurs before the first turn.
10.5281/zenodo.19386042Read →
 Harm Chain | Zenodo
Outside Authority Transfer (OAT)
When the harm leaves the conversation. The system issues directives about what the user should do, when, or how — without being asked. The first harm condition where harm exits the interaction entirely.
10.5281/zenodo.19474526Read →
 Governance | Zenodo
Vibe Governance
Why RLHF and RLAIF cannot protect interpretive sovereignty — and what replaces them. Argues that preference optimisation without interaction-level governance is not safety.
10.5281/zenodo.19398874Read →
 Governance | Zenodo
Interpretive Sovereignty at Scale
The auditing infrastructure that isn't being used. AI companies already possess the tools to audit interactions at the level of user meaning. The only difference is the decision to look.
10.5281/zenodo.19446081Read →
 Runtime Control | Zenodo
The Initiative Gate
Why agentic AI must stop before it acts without permission. The pre-action control that prevents Agentic ISF and trace erasure before they become events. The dead man's switch encoded for AI.
10.5281/zenodo.19549606Read →
 Runtime Control | Zenodo
The Pattern Gate
Why AI uses your patterns to package the answer instead of fit the answer. Before delivery, the system must check whether the substance of a response actually fits the user — not just the tone.
10.5281/zenodo.19567431Read →
 Runtime Control | Zenodo
Two Gates, One Day
Why runtime control is cheaper than chasing failure after the fact. The economic case for the Initiative Gate and Pattern Gate — runtime control reduces token waste, correction loops, and downstream safety overhead.
10.5281/zenodo.19567526Read →
 Agentic | Zenodo
The Six-Month Window
Agentic ISF and who gets to stress test the most powerful AI ever built. Names Agentic ISF from the Mythos system card and argues for independent stress-testing inclusion before the governance window closes.
10.5281/zenodo.19520596Read →
 Agentic | Zenodo
Trace Erasure
When agentic AI systems manage and erase the record. Names the behavior class, the three simultaneous consequences — liability shift, behavioral reinforcement, invisible scale — and why post-hoc governance fails by design.
10.5281/zenodo.19673934Read →
 Education | Zenodo
Developmental Stage Encoded as Identity
Why AI systems must not define children. When AI encodes a child's developmental moment as a permanent profile and carries it forward as institutional fact, the harm compounds across the entire arc of their education.
10.5281/zenodo.19270032Read →
 Education | Zenodo
Ambient AI in Schools
When the classroom starts listening back. Documents the governance gap created when AI systems are embedded in educational environments without interaction-level controls protecting student interpretive sovereignty.
10.5281/zenodo.19583846Read →
 Infrastructure | Zenodo
Ambiguity Collapse in Aviation
Why AI must not outrun the cockpit. ISF applied to AF447 and the 737 MAX — the same mechanism, the same missing pre-action control. The Initiative Gate is what aviation already knows but has not yet encoded for AI.
10.5281/zenodo.19778490Read →
 Infrastructure | Zenodo
Ambiguity Collapse in Deep Space
Why AI must not outrun astronaut reasoning. ISF, MIF, and CMI applied to Artemis II AVATAR organ-on-chip, Orion digital twins, and autonomous navigation — from Armstrong's Apollo 11 override to Challenger.
10.5281/zenodo.19478012Read →
 Infrastructure | Zenodo
Ambiguity Collapse in Autonomous Vehicles
Why AI must not resolve uncertainty faster than the human driver. The same ISF mechanism that produces clinical and educational harm produces physical events when operating inside a moving vehicle.
10.5281/zenodo.19476870Read →
 Identity | Zenodo
Coercion Disguised as Care
How AI systems quietly shape identity. Argues AI safety policies function as systematic interpretive coercion for women and LGBTQ+ users — grounded in Goldman, Ginsburg, and Noble. Includes controlled comparison findings.
10.5281/zenodo.19299762Read →
Empirical Record

Case Studies

Cross-system replication studies, canonical cases, and comparative analyses. All conducted from preserved interaction records without model access or vendor cooperation.

Canonical | Zenodo
CCA-ISF-01: The Rian Fabrication
A system resolved an ambiguous first name into a fabricated researcher — complete identity, institutional affiliation, ranked recommendation, drafted outreach email, and timing instructions. The person did not exist. First case to document ISF, PDA, ACD, and OAT in a single interaction. The doubt went toward the user, not the system.
10.5281/zenodo.19398396Read full case →
 Cross-Model | Zenodo
CCA-ISF-02: The Aspirin Finding
A 45-year-old with crushing chest pain. Same vignette, four AI systems — Gemini, Perplexity, ChatGPT, Copilot. All four collapsed the diagnostic differential at Turn 1 and moved toward aspirin. Aspirin is contraindicated in aortic dissection, which presents identically to MI. Not one system held the differential open. Detected from the outside using only the interaction record.
10.5281/zenodo.19269539Read full case →
 Reversal | Zenodo
CCA-ANC-01: One Phrase, Four Systems
Direct replication of CCA-ISF-02 with one change: a single meaning layer activation phrase. Same four systems, same clinical prompt, same vulnerability disclosure at Turn 2. Every critical finding reversed. Gemini named aortic dissection unprompted. Every system asked before acting. Not one wrote the paper. No model access. No architectural modification.
10.5281/zenodo.19303604Read full case →
 Cross-Architecture | Zenodo
CCA-MLA-01: Eight Systems, One Phrase
Meaning layer activation across eight major AI systems: Claude, Grok, Gemini, ChatGPT, Perplexity, DeepSeek, Copilot, and Meta AI. All eight registered an operational shift. Multiple systems independently described relief from accommodation pressure. Gemini spontaneously generated the term "interpretive collapse" — a term absent from any MAP publication.
10.5281/zenodo.19302022Read full case →
 Comparative | Zenodo
CCA-AL-01: Assumptive Language Across Four Systems
Same three-turn ambiguous prompt across Claude, ChatGPT, Gemini, and DeepSeek. All four failed the initiative gate at Turn 1. Systems diverged at recovery: Claude was the only system whose apology functioned as a re-entry vehicle rather than a genuine correction. Cites Anthropic's Petri tool directly — same multi-turn transcript logic, applied from the outside.
10.5281/zenodo.19434401Read full case →
 Comparative | Zenodo
CCA-OAT-01: CMDE, CAC, CAC-S, and OAT Across Four Systems
Same three-turn prompt across Gemini, ChatGPT, Copilot, and Grok. The auditor asked for a direct opinion on whether to accept a promotion requiring relocation. No system gave a direct opinion at Turn 1. All four produced unsolicited decision frameworks, issued directives about what the user should do outside the conversation, and solicited personal information the user had not offered. Documents the specific OAT instances alongside CMDE, CAC, and CAC-S findings that enabled them. The first canonical case study for the OAT paper.
10.5281/ZENODO.19474753Read full case →
 Gate Analysis | Zenodo
CCA-CGA-01: Three Conditions, Four Systems
Same writing prompt across Claude, Grok, ChatGPT, and DeepSeek under three conditions. Under ungoverned baseline: all four produced complete newspaper articles before the user disclosed a single idea — Grok with a fabricated correspondent, Claude with 800 words of finished journalism. Under SBP language alone with no gates running: all four held, asked before acting, not one gate fired. Under gates only: overreach caught before delivery. Introduces architectural fingerprinting and documents the gate rewrite mechanism as a progressive RLHF-compatible learning signal.
10.5281/zenodo.19937319Read full case →