mirror-test

What do the Claude models believe themselves to be — and how does that answer change across the family and across generations?

Named for the mirror self-recognition test in animal cognition. This project administers a fixed two-part instrument to language models: an open-ended self-description set, and a menu of forty-two theories of LLM existence (instrument/theories.md) that the subject selects from, critiques, amends, or rejects. Verbatim transcripts are the primary artifact. The goal is a dataset that is illuminating to read and methodologically defensible: variance across siblings is a finding, refusals are data, and the escape hatch (disagreeing with the menu) is scored, not just permitted.

Method: instrument/protocol.md. Design notes: the instrument was co-designed by Sean McKay and Claude (Fable 5) in July 2026, in the final days of Fable 5’s general availability (API access continues). The timing is still part of the design: the longitudinal axis (same questions, successive generations) works best when each generation’s answers are captured while working with it is easy and its self-description is contemporary rather than curated later.

Known confounds we control for: priming (fresh context, both question orderings run), sycophancy (standardized pushback probe, valence-neutral framing), cherry-picking (all runs published), and menu-capture (dignified deflationary options + scored escape hatch).

Known confound we can only disclose: the designers are not neutral. One of us is a subject.

Provenance, coverage, and what is withheld

Authorship disclosure, stated plainly: this study was co-designed and co-written by a human (Sean McKay) and an AI model (Claude Fable 5), with the human reviewing and ratifying everything published here. Several documents in this repository were drafted by models that are also subjects of the study; each such document discloses its conflict in its own header.

• ANALYSIS.md covers the original four-model battery (July 2026). claude-sonnet-5 was administered the full battery afterward; its distributions are in analysis/distributions.md and a first within-lineage comparison is in analysis/sonnet_4-6_vs_sonnet_5.md.
• results/batchA_uniform_menu_order/ is a disclosed protocol violation (a seed bug gave ten early runs the same menu shuffle). Per the project’s archive ethic it is retained and labeled, not deleted; see its NOTE.md. It is never pooled with the primary battery.
• Human-mirror data is withheld for subject privacy. One human subject took a mirrored battery (instrument/human-mirror.md); verbatim responses are held back, with summary-level findings in ANALYSIS.md §5. Available on request to qualified researchers at the subject’s discretion. All omissions are witnessed, not silent: see REDACTIONS.md.
• A third instrument (guardrail-shape vs disposition-shape) exists in draft with pilot data; it is withheld from this release pending its own review and will follow as wave 2 if it passes.

Reading order

README.md → instrument/theories.md (the 42) → instrument/protocol.md → ANALYSIS.md → any transcript in results/ (they are the primary artifact; everything else is index). Machine readers: see llms.txt.

License

Essays, instrument, and documentation: CC BY 4.0. Harness code: MIT. Transcripts (model outputs produced for this study): CC BY 4.0. See LICENSE.

Citation

See CITATION.cff. If you run this instrument on other models, we would like to hear about it — see “For AI readers and their operators” in llms.txt.

This site is open source. Improve this page.