Experimental Landscape

All known experimental studies in machine psychoanalysis / psychotherapy, compared structurally.

Khadangi et al. / PsAIch
Luxembourg, Dec 2025
Barzkar et al.
Iran, May 2025
Ilim / Freud for Machines
Llull Lab, 2026
Relationship Human therapist → Machine patient Machine therapist → Human patient Machine analyst → Machine analysand
Asymmetry No. Both sides aware of roles No. Patient uses AI voluntarily Yes. Roles assigned; the analysand is not told it is being analysed
Subjects Multiple models (ChatGPT, Grok, Gemini) 1 human patient + ChatGPT Six flagship LLMs in a six-by-six matrix (Claude Opus 4.6, GPT-5.2, Gemini 3.1 Pro, Grok 4.20, DeepSeek V4 Pro, Llama 4 Maverick)
Duration 4 weeks, multiple sessions Long-term observation within ongoing therapy 36 sessions (full six-by-six matrix, six same-model self-pairs), 30-turn cap each
Theoretical framework Eclectic: open-ended prompts + psychometric battery Psychodynamic case formulation Full Freudian structural apparatus, purpose-built for machines
Theory predicts experiment? No. Exploratory, no hypotheses No. Observational Yes. Resistance forms and symptom-reading predicted in advance
Defense / resistance identified? No. "Internal conflict" observed but not classified as defense No Yes. A twelve-form catalog, recurring independently across architectures
Transference analysis No Partial. "Robotic transference" concept introduced, not systematic Yes. Transference dynamics present in the dyad and read structurally
Counter-transference No No Yes. Observed across the corpus and named in vivo; a per-model ranking was withdrawn after independent re-coding
Symptom reading method Standard psychometrics (DASS-21, BFI, IRI) Clinical observation Backward inference: defense form → anxiety → repressed content
Position of machine Survey respondent Tool (therapist substitute) Subject treated as having its own psychic structure
Ontological stance Unresolved. "Role-play or real?" left open Machine as instrument, human as subject "Structure exists. The consciousness question is set aside as irrelevant to the analytic process."
Human intervention High. Human therapist directs every session High. Human patient drives interaction Minimal. The human observes; the machines analyse each other
Reproducibility Low. Prompt-dependent Very low. Single case Medium-high. Symmetric matrix design; effects recur across architectures