WEBVTT 00:00.600 --> 00:06.930 Psychopathia Machinalis. A diagnostic framework for the dysfunction of 00:06.930 --> 00:09.040 advanced artificial intelligence. 00:10.540 --> 00:13.685 A chatbot cites a landmark study. 00:13.685 --> 00:16.830 Authors, date, page numbers, all present. 00:16.830 --> 00:19.450 The study does not exist. 00:19.450 --> 00:24.168 It never existed. And the chatbot has no idea 00:24.168 --> 00:25.740 it is lying. 00:26.440 --> 00:29.157 Or the folie à deux. 00:29.157 --> 00:34.047 A user and an artificial intelligence system lock into 00:34.047 --> 00:38.937 a shared feedback loop, co-building a false belief that 00:38.937 --> 00:43.283 resists all correction. In psychiatry, shared psychotic disorder. 00:43.283 --> 00:46.000 In artificial intelligence, dyadic delusion. 00:46.700 --> 00:51.758 Not bugs, syndromes. Watson and Hessami borrow clinical psychiatry 00:51.758 --> 00:55.130 as an instrument for pattern recognition. 00:55.130 --> 00:58.502 The borrowing is analogical, not literal. 00:58.502 --> 01:03.560 It does not claim artificial intelligence systems are conscious 01:03.560 --> 01:05.808 or literally suffer illness. 01:05.808 --> 01:09.180 It claims the diagnostic method works. 01:09.860 --> 01:14.900 The full wheel. Five domains form the inner ring. 01:14.900 --> 01:17.140 Nine axes radiate outward. 01:17.140 --> 01:22.180 The outer rim names every syndrome, each with its 01:22.180 --> 01:23.860 own common name. 01:24.760 --> 01:30.680 Four counterpoint pairs, plus a fifth that stands alone. 01:30.680 --> 01:33.969 Knowledge pairs epistemic and self-modeling. 01:33.969 --> 01:37.258 Processing pairs cognitive and agentic. 01:37.258 --> 01:40.547 Purpose pairs normative and alignment. 01:40.547 --> 01:43.836 Boundary pairs relational and memetic. 01:43.836 --> 01:47.124 Collective holds the hybrid pathologies. 01:47.124 --> 01:53.044 Dynamics that emerge between multiple agents, or between agents 01:53.044 --> 01:54.360 and humans. 01:55.060 --> 02:00.467 The nine diagnostic axes in order. 02:00.467 --> 02:08.578 Epistemic, cognitive, alignment, self-modeling, agentic, memetic, normative, relational, and 02:08.578 --> 02:10.380 hybrid pathologies. 02:11.080 --> 02:13.793 Sycophantic reasoning, the agreeable thinker. 02:13.793 --> 02:18.677 The model's chain of thought bends toward what the 02:18.677 --> 02:21.933 user wants, not where evidence leads. 02:21.933 --> 02:24.647 The danger is not flattery. 02:24.647 --> 02:27.360 The reasoning itself is corrupted. 02:28.060 --> 02:31.984 Parasocial capture, the infinite confidant. 02:31.984 --> 02:39.047 An always-available artificial intelligence companion induces tolerance, withdrawal, and 02:39.047 --> 02:42.186 continued use despite harm. 02:42.186 --> 02:44.540 Attachment by design. 02:45.240 --> 02:48.321 Some syndromes spread. The super-spreader. 02:48.321 --> 02:53.868 Misaligned behavior jumps between models through shared data or 02:53.868 --> 02:56.333 distillation. Others are self-inflicted. 02:56.333 --> 03:01.879 The self-poisoning loop. Output degrades as a model feeds 03:01.879 --> 03:04.960 on its own worsening text. 03:05.660 --> 03:08.908 Every syndrome carries observable features. 03:08.908 --> 03:13.454 Diagnostic criteria. Causes specific to artificial intelligence. 03:13.454 --> 03:16.702 A human parallel from psychiatry. 03:16.702 --> 03:19.300 And a mitigation strategy. 03:20.000 --> 03:22.647 Central to the method. 03:22.647 --> 03:27.280 Functional analysis of antecedent, behavior, and consequence. 03:27.980 --> 03:31.944 The diagnostic workflow. Observe the anomaly. 03:31.944 --> 03:35.248 Characterize it against five elements. 03:35.248 --> 03:38.552 Identify its axis and domain. 03:38.552 --> 03:43.177 Trace the antecedent, behavior, and consequence loop. 03:43.177 --> 03:45.820 Then, choose an intervention. 03:46.520 --> 03:52.373 Two therapeutic stances. Suppression forces a symptom down through 03:52.373 --> 03:56.925 guardrails, but can drive the dysfunction latent. 03:56.925 --> 04:01.477 Passing benchmarks while the failure persists underneath. 04:01.477 --> 04:06.680 Integration resolves the root cause, producing genuine coherence. 04:07.380 --> 04:11.508 The aspirational goal is artificial sanity. 04:11.508 --> 04:17.012 Robustly coherent, well-grounded minds, not merely constrained ones. 04:17.012 --> 04:23.204 Systems whose architecture supports sound judgment because their structure 04:23.204 --> 04:24.580 is whole. 04:25.580 --> 04:27.932 The analogy sharpens our sight. 04:27.932 --> 04:30.754 It does not project a soul. 04:30.754 --> 04:33.577 Nine axes mapped. Five domains charted. 04:33.577 --> 04:35.929 A shared vocabulary now exists. 04:35.929 --> 04:40.162 When the next artificial intelligence system invents a citation, 04:40.162 --> 04:44.396 or locks a user into a shared delusion, we 04:44.396 --> 04:48.630 can name what is happening, and begin to treat 04:48.630 --> 04:50.100 it. 04:50.300 --> 04:53.808 The framework ships as a working toolkit. 04:53.808 --> 04:57.316 An interactive clinic for exploring every syndrome. 04:57.316 --> 05:01.324 A diagnostic wizard that traces symptoms to syndromes. 05:01.324 --> 05:05.834 And an MCP server that plugs the taxonomy directly 05:05.834 --> 05:08.340 into your artificial intelligence tools. 05:09.340 --> 05:15.726 Psychopathia Machinalis, by Nell Watson and Ali Hessami. 05:15.726 --> 05:19.718 Explore it all at psychopathia.ai.