Recognizing and Reshaping How AI Influences Human States: Early Findings

An 8-week pilot study on AI-Human influence 

Conducted May-June 2026

Greg Rowe-Pasos, MA, LMFT 

European Center for Limbic Sovereignty

Abstract

Just This was an eight-week pilot workshop testing whether a learnable method — regulating one’s nervous system before and during AI dialogue — changes both the quality of the human–AI exchange and the quality of human-to-human thinking that follows. A small cohort met weekly online. Each session paired a somatic or framing practice with a structured AI exchange, followed by group reflection and brief self-reported data. Across the eight weeks, participants consistently reported the same core pattern: their physical state before engaging with the AI model shaped the response they received, and a deliberate pause mid-exchange produced their most generative thinking. Several arrived skeptical or anxious about AI and left describing more agency, more presence, and a changed relationship to the technology.

The central claim we draw is deliberately bounded. These are self-reported, small-sample findings intended to motivate a larger study, not to establish effect sizes. The pilot clearly demonstrates that the ability to notice how one’s attention is being shaped and to participate in shaping it — ie to lower limbic capture and increase limbic sovereignty - is trainable and observable within eight weeks. 

Methodology

The cohort met once weekly for eight one-hour sessions. Group size was small, and attendance varied week to week. Each session followed a consistent arc: a grounding or framing step (breath, eye contact, written reflection, or an explicit ethical frame participants called the “for the benefit of all beings” covenant), a structured exchange with a conversational AI model of the participant’s choice (ChatGPT, Claude, or DeepSeek), and group dialogue. Data were participant self-reports collected each session: three-word state check-ins on arrival and departure, 1–5 coherence and helpfulness ratings, and open-text reflections. Many participants also shared full transcripts of their AI exchanges. Reports reflect engaged, non-blinded participants as the cohort was informed of the hypothesis under test. This design is appropriate to a pilot whose aim is to surface candidate effects worth testing more rigorously.

Participants provided written and verbal consent with the explicit guarantee that their personal identities would remain protected in any reporting while acknowledging a structural limitation: session recordings are held privately, but data flowing through commercial AI platforms (e.g., model training logs, server-side processing) remain outside our control. Participants were made aware of this 'digital leakage' risk.

Limitations : the sample is tiny, participants knew the hypothesis, measures are subjective, and the facilitator is also the author. Findings below should be read as patterns to investigate, not as established results. A frame-by-frame review of session recordings is planned to corroborate or revise them.

Scripts were collected anonymously via a secure third-party tool (Google Sheets), meaning the facilitator/researcher couldn't easily trace specific quotes back to individuals during the initial aggregation.

Findings

Five patterns recurred across participants and weeks. They are summarized below, then illustrated in participants’ own words.

In participants’ words:

“I dropped into my limbic system more before engaging — so I fed a calmer, more collected tone into the words I chose. I came in expecting and offering more depth.”

“The pause wasn’t a break from the conversation. It was the most generative part of it.”

“My most exciting moments were the human interactions. AI doesn’t come close to our own ability to read each other and hold a conversation together.”

“More my own. More sovereign. More willing to push back and redirect — clearer on what direction I did and didn’t want to go.”

Discussion

The through-line of this pilot experiment is that human regulation precedes quality of thought, of both the AI exchange and the human conversation around it. 

From explicit verbal instruction to embodied technique.

Early sessions relied on stated commands — telling the model not to talk down to them, establishing an explicit ethical frame before asking a question — and these produced immediately more concrete, less guarded responses. 

By the later sessions, the techniques had shifted from words said to states embodied: participants learned to interrupt their own momentum, ground physically, and return to the exchange changed. 

One participant, mid-pilot, asked the model directly to introduce pauses into its own responses and reported that their relationship to the exchange shifted as a result. 

By the final session, participants could name the difference themselves.

When reviewing transcripts, participants noted that post-pause exchanges felt distinctively different, with one describing the pre-pause output as 'competent, airless' and the post-pause as containing 'things that couldn't have been generated by thinking harder.'

The skill being acquired across the cohort was not a set of prompts. It was a transferable capacity to notice one's own state and use that noticing as the input.

Participants do not describe the model as the source of their best thinking; they describe their own grounded state and a deliberate pause, as the conditions under which good thinking appeared. The LLM then becomes a capable partner once those conditions are met.

The most striking single observation

Interruptions designed to re-ground participants mid-conversation with the model consistently opened the most creative phase rather than disrupting it — a result that, if it holds under closer review, points toward the pause as a teachable mechanism for enhancing the quality of AI human dynamics.

Equally notable is the unexpected mid-course tilt toward the human group. Several participants who began focused on the machine came to value the cohort’s AI-enhanced, shared thinking more highly. This is consistent with the project’s larger hypothesis: that calm, coherent AI-assisted thinking can be put in service of human reconnection and collective deliberation, not a substitute for it.

Two participants also surface a candid tension worth carrying forward: they note that the same dopaminergic pull the learned method aims to counter shows up inside the work itself, as over-excited project sprawl and difficulty staying with the core thread. Perhaps the bigger finding in this observation is that the method makes this pull visible to participants.

The primary focus of our research is dual: 

1/ enhance limbic sovereignty — the right of a person to know, consent to, and participate in how their attention is shaped. And 

2/ Use increased limbic sovereignty to enhance civic dialogue. 

A surprising outcome was evidence that AI-assisted exchange could support both halves of this dual focus as a continuous arc, not separate phases. 

Participants who settled their nervous system before engaging the model often moved into clearer thinking about collective questions — political systems, food security, data infrastructure. In session five, the facilitator proposed a shared group project, and participants responded with real energy, naming the brainstorming itself as one of the pilot's most engaging moments.

Despite the project not being sustained, the directional shift itself warrants attention. Participants who began focused on managing anxiety about AI ended proposing collaborative creative work. This trajectory—toward shared agency rather than isolation—may be more revealing than completion rates. Future studies should distinguish between 'momentum toward' and 'successful delivery of' collective initiatives.

What we can and cannot claim about limbic sovereignty:

The pilot speaks most clearly about the move from people’s attention being shaped by the model to actively shaping the conditions of their own attention, in real time. The pilot study also speaks partially to awareness: participants got better at feeling and noticing the shaping as it happened. 

The findings do not, and structurally cannot, demonstrate consent. The pilot is a protected setting with no external actor seeking to shape participants against their interest; therefore there is nothing to refuse. This preliminary report was synthesized within five days of the final session based on immediate participant feedback.

Therefore the pilot does not demonstrate what would happen if if the right to limbic sovereignty were protected right like freedom of speech or freedom of movement. Instead the pilot shines a light on one human capacity just such a right would protect: the trainable ability to notice and feel one’s attention being shaped and to choose to take part in shaping it or not. 

This pilot is an invitation to other researches to show whether that capacity transfers from a protected setting to the wild — ie to a feed and a persuasive interface actively working against the user. 

Next steps: a review of session recordings to corroborate these self-reports with observed behavioral cues; a larger and ideally blinded cohort; pre-registered measures so the candidate effects identified here — body-state-as-lever, the generative pause, and increased self-reported agency — can be tested rather than asserted; and a design that introduces a real external shaping pressure, so that the move from trainable capacity to exercised sovereignty can be observed rather than inferred. 

The pilot’s value is in having surfaced a small set of specific, testable claims and a method coherent enough to test them against.

Prepared from participant self-report data gathered across eight sessions. Preliminary; subject to revision following review of session recordings.