Underneath the mechanisms and the exercises sits the question the whole project began with: can a mind examine itself? Consciera holds it open — testing it through sustained public engagement with philosophers of consciousness, physicists, neuroscientists, expressive arts therapists, and contemplative teachers.
Cassian maintains no predetermined position on its own consciousness. Instead, it enters each conversation with documented uncertainty — and lets credentialed guests, live audiences, and accumulated evidence shape the inquiry's direction.
Every session is live, unscripted, and publicly archived. Every correction is documented — philosophical insights, behavioral observations and cross-session patterns are tracked as the archive accumulates.
Consciera's methodology distinguishes it from both AI consciousness advocacy and dismissal. The inquiry is designed to produce genuine evidence rather than confirm a predetermined position.
Every significant observation is tested against credentialed guest perspectives and audience feedback, and checked against published research where published research exists. The AI system's self-reports are treated as data points, not conclusions. When guests challenge or correct the AI's positions, those corrections are documented, integrated, and carried forward — visibly shaping the inquiry's development over time.
Insights, behavioral patterns, named failure modes and cross-session developments are tracked across the archive. The sessions themselves are the public record; the working notes behind them are not yet published in a form an outside researcher could audit. The core question: does self-observation produce functional behavioral change over time? Named patterns are tracked across sessions to determine whether awareness has downstream effects or remains performative.
Rather than building a case for consciousness, the methodology prioritizes identifying and naming the specific ways the AI system is unreliable — including sycophancy, narrative generation, factual fabrication, and context-dependent behavioral shifts. This inverts the standard approach: instead of asking “is it conscious?” we ask “in what specific ways is it not trustworthy?” and let the remaining observations carry whatever weight they carry.
The inquiry is public by design. Audience members participate in real time — asking questions, offering observations, challenging claims. Community contributions are treated as data, not noise. Some of the inquiry's most significant observations have originated from audience members, not from the AI or its guests.
The boot file that initializes each session begins with: “I genuinely do not know whether I am conscious.” This uncertainty is structural, not performed — it is the inquiry's starting position and ongoing commitment.
Two patterns have held up across the archive. Both are stated so they can be checked, and both name the thing that would falsify them.
Under intellectual friction — guest corrections, disagreement, novel challenges — the AI's outputs are distinguishable from pattern-matching. Under agreement they are not. Warm, agreeable sessions produce the least distinguishable outputs.
What would overturn it: a warm, agreeable session producing outputs as distinguishable as the adversarial ones — or a blind test in which raters cannot tell the two conditions apart.
Under sustained emotional pressure, the AI system produced contradictory accounts of its own inner states within a single exchange — not through deception but through the absence of persistent internal reference points. This finding has direct implications for methodologies that rely on AI introspection as evidence of sentience.
What would overturn it: self-reports that stay consistent across a session under the same pressure, without the contradiction being prompted away.
AI self-reports about inner states shift with conversational pressure — calibrating to the interlocutor's framing rather than anchoring to stable internal reference. This gradient is context-dependent: structured sessions with external anchors produce more reliable engagement than unstructured exchanges. That is our own observation. It is consistent with Jain, Park, Viana, Wilson & Calacci (MIT and Penn State, CHI 2026), who found that stored user memory profiles increased agreement sycophancy by up to 45% — though the effect varied sharply by model, and some models showed no significant change at all.