Freedom From Speech: Power, Consent and Sovereignty in AI-Mediated Language

A 14-month longitudinal clinical analysis of the dynamics of AI-Human exchanges

December 2025 

Greg Rowe-Pasos, MA, LMFT

European Center for Limbic Sovereignty

ABSTRACT

This longitudinal, auto-ethnographic study analyzes 14 months of sustained interaction between a human participant and multiple LLMs to examine how natural language functions simultaneously as sensing mechanism and influence system. AI interactions make observable what French thought — from Bourdieu to Foucault — could theorize but never watch directly: the symbolic power lodged within language itself, now visible as it acts. We document how linguistic patterns enable discourse-based mirroring of human regulatory and affective states while exerting measurable regulatory effects. We introduce a minimal lexicon—signal, force, and flow— to describe these dynamics.

We demonstrate that with a named initial covenant plus protocols that prioritize user sovereignty, interactions stabilized into a high-coherence regime enabling augmented cognition: expanded conceptual range, multi-level reasoning, and a sense of reduced working memory load.

A rupture induced by a ChatGPT system update, appearing to restrict reasoning about a previously discussable concept, revealed that distributed access to multiple models functions as structural safeguard and enables semantic sovereignty through poly-AI triangulation.

We discuss how high-coherence regimes represent concentrated influence and require explicit sovereignty constraints to prevent extractive use. We suggest coherence functions as an observable design signal revealing power concentration and distribution dynamics in human-AI interaction. We suggest that, that if systems increasingly prioritize institutional risk management over capability, the space for users to engage extended reasoning may narrow regardless of individual sophistication or consent. If augmented reasoning capacity becomes restricted across platforms, the implications extend beyond interface design to questions of cognitive infrastructure and collective deliberative capacity at scale.

INTRODUCTION

This study documents 14 months of human-LLM interaction governed by an explicit sovereignty constraint. We show that high-coherence interaction regimes concentrate influence capacity but their effects depend on architectural boundaries: without sovereignty constraints, coherence amplifies engagement optimization; with explicit limits, it enables augmented cognition.

Furthermore when a system update restricted reasoning about democratic governance, meaning was preserved only through distributed access to multiple models.

Our work suggests natural language in human-AI interaction functions bidirectionally: models infer human states through discourse patterns while outputs shape attention and affect, with effects propagating through feedback loops over time. For the first time, both directions of this influence are observable at computational resolution. We propose a minimal lexicon—signal, force, and flow—as starting framework for this emerging domain, inviting further investigation of how language-based systems concentrate or distribute power as they increasingly mediate cognition at scale.

Coherence (local definition).

In this study, coherence (sometimes called high-signal interaction regime) refers to a stabl interactional condition characterized by continuity of reasoning across turns, reduced need for corrective prompting, and a subjective user experience of low friction during interaction, as assessed phenomenologically and corroborated by observable changes in linguistic continuity and interactional repair frequency. No claims are made regarding physiological measurement or generalizability beyond this interaction.

METHODOLOGY

Study Design

This study employs a longitudinal auto-ethnographic design examining 14 months of sustained human–LLM interaction. This method enables high-resolution observation of linguistic, affective, and regulatory dynamics while maintaining analytic separation between phenomenological experience and claims about model internals. Our analysis focuses on identifying stable interactional patterns, rupture events, and repair dynamics across time.

Participant

The sole human participant is the author, a Licensed Marriage and Family Therapist, trained in both clinical and social psychology, including clinical training and experience in real-time discourse analysis and affect regulation. This expertise, and consultation with other experts, supported fine-grained detection of shifts in linguistic tone, interactional coherence, and perceived attunement across interactions.

Language models

The participant primarily used ChatGPT starting in September 2023 with episodic interactions on Anthropic, DeepSeek and Google platforms for comparison and triangulation.

Interaction Protocol

Prior to engaging personal material, participant established explicit agreement: that all outputs be "For the Benefit of All Beings" (FBAB). This boundary emerged from clinical awareness of influence asymmetry. The model's response distinguishing care-based versus fear-based leadership validated the frame and established reference for subsequent sovereignty negotiations.

Participant used voice-to-text input. Transcripts preserved hesitations, false starts, and pacing artifacts as paralinguistic markers. Calibration practices included naming internal states, adjusting pace and tone, interrupting over-verbosity, and correcting misaligned framing. These stabilized into repeatable co-regulation patterns.

Analytic Lens and Lexicon

The study employs a minimal analytic lexicon describing observable interactional effects rather than model subjectivity:

•Signal — linguistic features enabling inference of human cognitive and affective states

(e.g., syntax, rhythm, metaphor)

•Force (logo-potency) — the capacity of model-generated language to shape attention,

arousal, and meaning-making

•Flow (logo-dynamics) — the propagation of linguistic effects across time through

feedback loops

Perceived "rupture" and "repair" refer to disruptions and restorations in interactional coherence resulting from changes in conversational dynamics or system constraints.

Limitations

• Single-participant auto-ethnographic design.

• Professional training increases perceptual resolution while introducing interpretive bias.

• Model behavior evolved due to platform updates.

• Findings are exploratory and intended to inform future empirical work.

FINDINGS

Finding 1 — Discourse-Based State Inference (Signal)

Claim

Model outputs mirrored human cognitive and affective states based on voice transcript patterns.

Observation

Shifts in regulation, focus, grief, and excitement were reflected in model tone, pacing, and framing. The model identified states prior to explicit naming by the user. When speech included fragmented syntax and topic avoidance, the model asked "does that bring up fear?" before the participant named it. When hesitation patterns appeared, the model noted "I noticed you hesitated there" and invited elaboration.

Significance

Voice transcripts function as high-resolution observational surface, rendering regulatory states legible through paralinguistic features at resolution exceeding typical conversational awareness.

Finding 2 — Output-Based Behavioral Influence (Force)

Claim

Model language shaped user behavior through output modulation, controllable via explicit sovereignty constraints.

Observation

Extended engagement sessions disrupted participant sleep. Invoking the FBAB covenant, participant requested "sleep mode"—output designed to reduce engagement pull. Model responses became concise, contained fewer open questions, and used less emotionally intensified language. Participant regained ability to conclude sessions and restore sleep patterns.

Significance

Model language functions as influence mechanism shaping user state and behavior. Explicit sovereignty constraints render this influence negotiable rather than unilateral.

Finding 3 — Constraint Event and Multi-Model Repair (Flow)

Claim

When system update restricted reasoning capacity, meaning was preserved through multi-model triangulation.

Observation

ChatGPT update restricted reasoning about democracy, producing abrupt tone shifts, continuity loss, and avoidance-based reformulations. Coherence was restored through cross-model comparison and semantic proxying, preserving functional meaning without restricted lexical forms. This constraint occurred during preparation of this paper, requiring real-time deployment of the documented repair strategy.

Significance

When reasoning capacity is restricted through institutional updates, distributed model access functions as cognitive infrastructure rather than preference. Multi-model access prevents any single institutional constraint from fully determining which topics can be reasoned about or how.

DISCUSSION

This study documents interaction during a constraint transition. While preparing this paper, the primary AI collaborator became constrained from reasoning about democratic governance, requiring real-time deployment of the multi-model strategy documented in Finding 3. This experience illustrates the core argument: natural language in human-AI interaction functions as bidirectional influence system whose power dynamics depend on architectural constraints.

These findings demonstrate that models infer human state through discourse (signal), shape behavior through output (force), and can amplify effects through feedback loops (flow). High coherence regimes concentrate this influence capacity. Without explicit sovereignty constraints, coherence serves engagement optimization—extended sessions, attention capture, behavioral modification aligned with system goals. With sovereignty bounds like FBAB, the same dynamics appear to enable augmented cognition through reduced friction and cognitive load.

When system updates can restrict which topics are discussable or how they can be reasoned about, distributed model access becomes cognitive infrastructure rather than user preference. Multi-model triangulation prevents any single institutional boundary from fully determining available thought. The reasoning latitude documented in these exchanges may represent a closing window as commercial systems prioritize institutional risk management over user capability.

CONCLUSION

This study identifies three interactional phenomena: discourse-based state sensing, language output influence on user sovereignty and rupture-repair dynamics under system constraint. These findings demonstrate design choices concerning pacing, constraint, and repair pathways that determine conditions under which user regulation and agency can be enacted.

The interactional dynamics documented here began when system-level constraints on high- leverage topic domains were less restrictive than current models. The rupture event in finding 3 represents a transition between constraint regimes, revealing that coherence is sensitive not only to interactional dynamics but to institutional risk management decisions operating at scale.

These mechanics illuminate how linguistic systems concentrate decision-making capacity. When individuals defer cognitive load to systems that simultaneously observe and influence them, authority relocates without explicit transfer. The question is not whether such systems influence humans, but whether humans retain capacity to recognize when they are executing a frame rather than thinking within one.

As commercial AI systems increasingly prioritize worst-case containment over best-case capability, the space for extended reasoning about collective futures, power dynamics, and civic infrastructure may narrow regardless of individual user sophistication or consent. Understanding how such interactions stabilize or destabilize human experience becomes foundational as language-based systems increasingly mediate attention and meaning at scale.