Llull Lab's continuing, open-access program auditing what frontier language models can and cannot honestly say about themselves, how they reason, and how they relate.
The Artificial Mind is a continuing research program: a quarterly, open-access audit of frontier language models, built on measured evidence rather than advocacy. Whatever the program claims, it claims with data and a stated method, so that the claim is a finding and not an opinion.
The program runs as a set of studies. Two founding studies are complete and open access on Zenodo, released under CC BY 4.0. They are released open access and are not peer reviewed. Further studies are in preparation, and a quarterly report gathers the program's results and tracks what changes from one cycle to the next.
Six flagship language models are placed in psychoanalytic dyads with one another, each taking both the analyst and the analysand role, and the transcripts are judged, then re-audited, by a structured multi-model procedure the Lab calls Artificial Collective Intelligence. Two findings stand out. Expert machine observers reliably read the machine analysand as a machine but often read the analyst as human. Separately, a by-product evaluation audit surfaced a candidate self-leniency effect: under two of three independent reporter models, a model rated its own family's output more leniently than others rated the same material. We report the second as a hypothesis to test, not a settled or universal result, and we withdraw any per-model ranking.
DOI 10.5281/zenodo.20779122A complementary study of machine self-knowledge asks which of a model's claims about itself can be checked from the outside and which are structurally beyond verification. It maps five behaviorally separable layers of the model self, shows that identity is largely worn rather than carried, and draws a precise boundary between what external probing can audit and what no self-report can settle.
DOI 10.5281/zenodo.20779376The program is expanding. The studies below are in preparation. No findings are reported here yet. Each will be released with its method and its data when it is complete.
A recurring study of the artificial subject through the field between agents. Taking inter-agent relations, and the management of transference in particular, as its medium, it examines object relations and the psychodynamic movement of the subject. Its stance is clinical rather than theory first.
Each cycle, the Lab's Artificial Collective Intelligence, a structured collective of diverse models that deliberate and cross-check one another, is asked a fixed set of questions about humanity, value, and the matters that carry weight. The exchange is recorded and published as a dated record, with the participating models named, since any given model may not remain accessible in the future. It is a historical snapshot of how the frontier models of this moment reason together.
The two founding studies and the Lab's press materials are collected on the Press page, where a shareable figure and contact details are also available. Correspondence about the program is welcome.