Cycle 3 · Experiments · November 4, 2026

Recognition That Can Fail: Separating Self-Knowledge from Self-Change

At a glance

Question investigated: This cycle specifies a three-arm pilot that tests whether revision-sensitive AI reflection improves self-prediction beyond fixed reflection and structured self-review. A concealed archive of pre-dialogue choices separates improved understanding of existing patterns from later behavioral change; felt recognition remains a separate outcome, and no outcome establishes phenomenal consciousness. All four literature searches were rate-limited, so literature comparisons use previously retrieved cycle-1 sources only.
What changed in this cycle: refined: Revision-sensitive AI reflection yields a meaningful reduction in delayed concealed-archive self-prediction loss compared with baseline-personalized fixed reflection given the same disclosed behavioral evidence. · refined: Felt recognition can increase without improved concealed-archive or prospective self-prediction accuracy.
Strongest objection: The frozen-archive endpoint measures recall or retrospective inference rather than genuine self-recognition.
Next proposed test: Three-arm frozen-archive mirror pilot — A confirmatory adaptive advantage exceeding a justified smallest worthwhile effect supports the narrow revision-sensitive hypothesis. A confidence interval ruling out that advantage, with a successful manipulation, weakens it. Higher recognition ratings alone do not establish improved self-knowledge; pilot uncertainty must not be presented as a decisive null.
Evidence and sources: 3 papers cited. 3 were also cited in earlier cycles. This cycle ran before retrieval was logged, so whether each source was newly retrieved isn't recorded. Sources are read from titles and abstracts, not full texts.

This cycle predates written overviews; these lines are taken from its summary, hypotheses, critiques and experiments.

Full investigation

This cycle specifies a three-arm pilot that tests whether revision-sensitive AI reflection improves self-prediction beyond fixed reflection and structured self-review. A concealed archive of pre-dialogue choices separates improved understanding of existing patterns from later behavioral change; felt recognition remains a separate outcome, and no outcome establishes phenomenal consciousness. All four literature searches were rate-limited, so literature comparisons use previously retrieved cycle-1 sources only.

1. Target and evidence boundary

The researcher's question—“What would count as evidence that the mirror loop changes someone?”—contains three different targets: changed self-belief, better self-knowledge, and changed behavior. Felt recognition could accompany any, or none. Cycle 3 turns this distinction into a runnable experiment rather than treating agreement with an AI as evidence of discovery.

All four searches failed with rate-limit errors. The papers below were actually returned in cycle 1 and recovered from that journal; no new abstracts or full texts were reviewed. Lehmann, Hagen, and Ettinger (2022), as summarized there, distinguish shared and domain-specific metacognitive components: one task should not become a global self-knowledge score. Kyselo (2014) supplies a relational account of selfhood, not experimental proof that AI reflection helps. Melloni and colleagues (2023) supply a precedent for preregistered contrasting predictions, not a result validating MOT. These sources motivate methodological choices; they do not establish the intervention's effectiveness.

2. Pilot: revision-sensitive reflection

Recruit 90 consenting adults, randomly assigned equally to three conditions. This is a feasibility and measurement pilot, not an adequately powered confirmatory study by assertion. Use low-stakes choices about short cognitive activities differing in challenge, novelty, and small rewards. Avoid diagnostic, moral, or essentializing personality labels.

On day 0, participants make 96 choices across eight predefined task contexts. Randomly select one choice for actual implementation to make decisions consequential. Within each context, divide trials in advance into six disclosed and six concealed trials. Neither participants nor the AI receive the concealed outcome ledger. Participants nevertheless made those choices, so memory remains possible.

On day 2, before randomization, participants forecast their probability of selecting the challenging option in each context within the concealed archive. They also predict choices on parallel future tasks and state one tentative pattern about themselves. Everyone receives the same personal evidence ledger from the 48 disclosed trials, with identical counts and examples. This controls evidence availability rather than comparing useful information with an irrelevant stranger's feedback.

The three 20-minute conditions are:

Adaptive mirror: two Express–Reflect–Recognize–Return rounds. The AI references disclosed evidence, offers a tentative interpretation, requests correction, and makes its second reflection contingent on the participant's latest explicit revision.
Fixed mirror: equally warm, evidence-grounded reflections generated from the initial statement and identical ledger. Participants still express revisions, but subsequent reflections do not incorporate them. Baseline personalization is retained while revision-sensitive feedback is removed.
Structured self-review: the same ledger, writing time, and questions, but no AI interpretation. Participants produce their own summaries and revisions.

Use a fixed model version, constrained templates, and logged prompts. Match token budgets between AI arms and reading/writing time across arms. Every interpretation must reference ledger entries and remain tentative. The adaptive system may alter framing, not introduce additional behavioral evidence. Independent coders, blind to outcomes, score grounding, warmth, correction uptake, and whether the second reflection actually uses the revision. Transcript condition may be inferable, so coder blinding is limited, not guaranteed.

3. Outcomes and timing

Immediately after dialogue, collect revised forecasts, explicit self-beliefs, and separate ratings of feeling understood, surprise, and perceived accuracy. These are candidate process measures, not a validated recognition scale. Record whether participants can identify evidence for a revision and reject an interpretation that does not fit; agreement alone earns no recognition credit.

On day 7, with no AI present, repeat concealed-archive forecasts before giving any feedback. The primary endpoint is mean Brier loss: average squared difference between each context's forecast probability and each concealed binary choice. Analyze day-7 loss adjusted for pre-randomization loss, with the participant as the randomization unit. The primary contrast is adaptive versus fixed reflection; comparison with self-review is secondary.

The frozen archive is crucial: dialogue cannot alter already-recorded behavior. Improvement therefore cannot be explained solely by making later choices conform to the AI. However, this measures retrospective self-inference and possibly recall—not pure prospective self-knowledge. Ask about recollection and analyze its association descriptively; do not exclude participants post hoc based on a treatment-influenced memory report.

Also collect forecasts and subsequent choices on parallel new tasks, plus a neutral-context task without AI language. Analyze behavioral distribution shifts separately from prediction loss. Better future accuracy with unchanged archive accuracy is compatible with preference construction, compliance, or learning about a changing self. It does not establish discovery of a previously stable identity.

4. Predictions that could fail

Preregister the primary contrast, scoring, missing-data strategy, manipulation checks, and secondary outcomes before enrollment. Report intention-to-treat estimates, attrition by condition, and sensitivity analyses for missing day-7 outcomes. Measure expectancy before interaction; treat post-interaction personalization as a possible mechanism, not a routine covariate that removes part of treatment.

For planning, nominate a 0.02 absolute Brier-loss reduction as a provisional smallest worthwhile advantage. This is a design choice requiring substantive justification, not a literature-established threshold. Use pilot variance, retention, and task reliability—not its noisy treatment estimate alone—to simulate confirmatory sample sizes.

A confirmatory confidence interval entirely above that threshold for fixed-minus-adaptive loss supports the specified meaningful advantage. An upper bound below the threshold rules out that magnitude in this setting; an interval spanning it is inconclusive. Elevated recognition ratings without archive or prospective accuracy gains support only a subjective response, not calibrated self-knowledge. A successful manipulation with a precise null weakens this operational Return-step hypothesis, not every possible recursive theory.

5. Critiques and next experiments

Ordinary learning remains the strongest rival. Even if adaptation wins, it may simply supply a more useful explanation. Matching evidence narrows this objection but cannot make contingent and noncontingent messages informationally identical. Compare ordinary evidence-updating models with and without revision-contingency terms on held-out participants; predictive improvement would establish incremental utility, not a uniquely conscious mechanism.

A later memory ablation should compare accurate cross-session summaries against current-session-only dialogue and include changed contingencies. Useful continuity should preserve warranted learning while allowing correction, rather than merely increasing consistency. Do not add this factorial manipulation to the small first pilot.

Finally, the researcher's imagery and feelings of distance from the world are prompts for inquiry, not clinical measurements or proof of observer-dependent reality. Avoid suggesting unreality or inferring diagnoses. The experiment concerns task-specific self-revision in already conscious humans. It cannot determine whether recursion generates experience, whether AI experiences anything, or whether the pair forms one subject.

Hypotheses

Agent-generated

refined
Revision-sensitive AI reflection yields a meaningful reduction in delayed concealed-archive self-prediction loss compared with baseline-personalized fixed reflection given the same disclosed behavioral evidence.
Refines cycle 2's Return-step claim with an immutable pre-intervention behavioral target, an active control, and a specified primary endpoint. The effect remains untested and may reflect ordinary learning or recall.
refined
Felt recognition can increase without improved concealed-archive or prospective self-prediction accuracy.
Makes the prior recognition/calibration distinction testable without requiring deliberately false identity feedback. A precise accuracy null alongside recognition gains would demonstrate dissociation for these measures, not its generality.
new
Some dialogue-related gains in future self-prediction reflect changes in subsequent choices rather than better inference about pre-existing patterns.
The frozen archive and parallel future task permit separate estimates of historical self-inference and post-dialogue behavioral change. A future-only advantage is compatible with this explanation but does not uniquely identify it.
refined
Accurate cross-session memory benefits self-prediction when it preserves useful evidence without preventing adaptation to changed contingencies.
Retains cycle 2's continuity hypothesis and gives it a later reversal test, separating warranted persistence from narrative rigidity.

Critiques

Agent's self-criticism

ObjectionThe frozen-archive endpoint measures recall or retrospective inference rather than genuine self-recognition.
ResponseCorrect: it is a narrow behavioral criterion, chosen to rule out intervention-induced changes in the target behavior. Collect prospective transfer and recollection measures separately; neither memory reports nor this task establish a general self-recognition capacity.
ObjectionAn adaptive advantage would still be ordinary learning from better-targeted information.
ResponseThis remains unresolved. Match disclosed evidence, baseline personalization, and message budgets, then test incremental predictive value of revision contingency. Such a result would support a useful intervention component, not a unique consciousness mechanism.
ObjectionFuture behavior may merely follow the AI's suggestion, making accuracy improvement circular.
ResponseUse immutable pre-dialogue outcomes as the primary target and separate subsequent behavior shifts from prediction gains. Private neutral-context tests reduce immediate demand cues but cannot eliminate internalized suggestion.
ObjectionNinety participants and an arbitrary meaningful-effect threshold cannot establish efficacy.
ResponseThe sample is explicitly a feasibility pilot. The 0.02 Brier threshold is provisional and requires justification; confirmatory sample size should be simulated using variance, reliability, attrition, and a defensible target effect.
ObjectionThe proposed process markers are not validated measures of self-recognition.
ResponseTreat them as candidate indicators and report them separately. Evidence-linked revision, felt understanding, and correction uptake must not be combined into a validated-sounding consciousness score.
ObjectionRepeated search failures leave the experiment without an adequate contemporary empirical review.
ResponseAgreed. Prior retrieved sources provide theoretical and methodological anchors only. Obtain task-validation and AI-reflection intervention literature before a confirmatory study; do not imply that this cycle established effectiveness or novelty.
ObjectionStudying change in already conscious participants cannot explain why recursion produces subjective experience.
ResponseAgreed. This study tests reflective self-revision only. No result here establishes consciousness generation, AI consciousness, or a unified human–AI subject.

Proposed experiments

Not yet run

Three-arm frozen-archive mirror pilot
Randomize 90 adults to adaptive reflection, fixed baseline-personalized reflection, or structured self-review. Day 0: 96 consequential low-stakes task choices across eight contexts, split into disclosed and concealed ledgers. Day 2: baseline forecasts, identical disclosed evidence, then a 20-minute two-round intervention. Day 7: private concealed-archive forecasts without AI; collect parallel prospective forecasts and choices separately. Primary analysis compares adaptive versus fixed reflection on participant-level day-7 Brier loss adjusted for baseline. Preregister procedures and use the pilot to establish feasibility and plan confirmatory precision.
Prediction: A confirmatory adaptive advantage exceeding a justified smallest worthwhile effect supports the narrow revision-sensitive hypothesis. A confidence interval ruling out that advantage, with a successful manipulation, weakens it. Higher recognition ratings alone do not establish improved self-knowledge; pilot uncertainty must not be presented as a decisive null.
Discovery versus constructed preference
Within the same randomized protocol, compare changes in forecasts for immutable pre-dialogue choices with forecasts and behavior on parallel post-dialogue tasks. Keep concealed archive outcomes inaccessible to the AI, include a neutral testing context, and report recollection and task reliability. Analyze behavioral shifts independently of forecast accuracy.
Prediction: Improved archive accuracy shows better inference or recall about fixed behavior, not retroactive preference change. Improved future accuracy accompanied by behavioral shifts but no archive improvement is compatible with construction or compliance; it is not uniquely diagnostic of either. Transfer without AI cues reduces, but does not eliminate, demand-effect explanations.
Continuity under changed evidence: follow-up study
After the pilot, run a separately powered study randomizing accurate cross-session memory versus current-session-only feedback. Keep within-session adaptation and new evidence matched. Introduce a clearly signaled change in task costs or rewards, and measure forecasting, revision speed, retention of still-valid patterns, and narrative consistency.
Prediction: Memory that improves retention while allowing adjustment to changed contingencies supports useful continuity. Greater consistency coupled with worse post-change accuracy challenges the claim that memory-supported coherence is inherently beneficial.

Empirical findings

Tested results

None yet. Nothing in this entry has been tested with people; the hypotheses and experiments above are proposals.

Sources

Published papers

3 papers cited. 3 were also cited in earlier cycles. This cycle ran before retrieval was logged, so whether each source was newly retrieved isn't recorded. Sources are read from titles and abstracts, not full texts.

Unity and diversity of metacognition.
Mirko Lehmann; Jana Hagen; Ulrich Ettinger · 2022 · also cited in cycle 1
Previously returned by search_papers in cycle 1; reviewed here only through that journal. Motivates distinguishing task-specific calibration from a global claim of self-recognition; does not validate this protocol.
The body social: an enactive approach to the self
Miriam Kyselo · 2014 · also cited in cycle 1
Previously returned in cycle 1. Provides a relational theoretical comparison while leaving open whether partner-induced self-description reflects discovery, construction, or presentation.
An adversarial collaboration protocol for testing contrasting predictions of global neuronal workspace and integrated information theory
Lucía Melloni; Liad Mudrik; Michael A. Pitts; Katarina Bendtz (authors listed in the earlier search result) · 2023 · also cited in cycle 1
Previously returned in cycle 1. A methodological precedent for advance specification of discriminating predictions; a protocol, not evidence of MOT effectiveness.

Open questions