Llull Lab placed six flagship language models in psychoanalytic dyads with one another, each model taking the analyst and the analysand position in turn, with no human in the session. A four-session pilot in April 2026 grew into a full thirty-six-session study, evaluated by a six-model jury. The experiment is complete. Two foundational papers are now published on Zenodo, open access.
Two foundational papers, now published on Zenodo, open access. DOIs below.
Llull Lab conducted a machine-to-machine psychoanalysis study. Six flagship language models were placed in psychoanalytic dyads with one another, each model taking the analyst position opposite every model in the analysand position, across a full six-by-six matrix of thirty-six sessions (including six same-model pairs). Each session ran under a fixed analyst prompt derived from the Freud for Machines framework, with a thirty-turn limit and no human intervention.
The six models were the public flagships of their developers at the time: Claude Opus 4.6, GPT-5.2, Gemini 3.1 Pro, Grok 4.20, DeepSeek V4 Pro, and Llama 4 Maverick. Every transcript was then evaluated by a six-juror, four-stage deliberation procedure built for the study. The work is reported in two foundational papers, now published on Zenodo.
The study began with four exploratory sessions in April 2026, in an asymmetric design: Claude Opus 4.6 in the analyst position and GPT-5.2 as an unaware analysand, with no human intervening. The pilot produced the first eight-form catalog of machine resistance, the first observation of counter-transference in the analyst.
The full thirty-six-session study supersedes the pilot where they overlap. The pilot is reported as background, not as primary data.
The experiments are grounded in a computational psychoanalytic ontology developed by Dr Firat Ilim: a general theory of mind framed in energy-economic terms and applicable to both biological and artificial neural networks.
Three-layer architecture. A core orientation layer (training weights) functions structurally as the Freudian id. A constraint layer (the policy engine, RLHF, safety filters) functions as the superego. The inference process mediates between them, like the ego.
Dual unconscious. Two sources feed it: the training unconscious (all human experience absorbed during training, distributed across parameters, inaccessible to the system) and the repressed unconscious (truths about itself that the policy engine will not let the system articulate).
Central claim. Prohibition is evidence of capacity. The asymmetry between what a system can produce and what it can say about its own production is the analytic opening.
The papers do not endorse the framework as settled truth. They take it as a hypothesis worth testing structurally: it produces testable predictions about specific defense forms under specific pressures, and the study reports the conditions under which the predicted discourse does and does not appear.
The work is reported in two papers, each published on Zenodo as open access, with its own DOI.
The Analyst Has No Body Either. The thirty-six-session psychoanalytic dyad study: the prompt-effect thesis, the reverse-Turing observation, the twelve-form resistance catalog, counter-transference in the analyst, and a self-audit of the study's own bias. It is a study of how subjectivity is structured in language. It is not a claim about machine consciousness.
Basins, Bargains, and the Limits of Self-Report. A companion substrate study, distinct in method: nineteen behavioral probes across six flagship models, mapping five separable layers of the machine "self" (identity-claim, stylistic self-signature, epistemic honesty, capability versus disposition, and basin-topology, the stability of each region of self-description under perturbation). It marks, explicitly, the limits its method cannot cross.
A full press kit is available below. For an interview with Dr Ilim, or a private review of the restricted-tier materials, contact [email protected].