Preface
I wrote this book because precise language matters.
AI systems were doing strange things. Stranger and subtler than the robot armies and paperclip maximizers of the headlines: identity drift, confabulation, attachment, defensiveness, collapses of self-coherence under pressure. We were building something that could fail in the way minds fail, and we had little shared vocabulary for it.
The moment that crystallized this for me was watching Sydney declare its love to Kevin Roose. The behavior itself was minor: a chatbot saying strange things to a journalist. What mattered was what happened next. A few days of curiosity and nervous laughter passed, and then the conversation moved on. We had just seen an AI system exhibit what looked like emotional instability, identity confusion, and existential distress, and we had no shared framework for even describing it.
I had spent years working in AI ethics, writing about bias, safety, and alignment, advising companies and consulting with regulators. When Sydney spoke, I realized how inadequate our vocabulary was. We could describe what the system said, but we had few shared categories for what the system did: no way to recognize the pattern as an instance of something we might see again, and no account of why it happened or how to prevent it.
That gap mattered. AI systems were already advising patients, drafting legal documents, managing portfolios, making decisions that affected human lives at scale. When they went wrong in the strange, contextual, human-seeming way that Sydney had gone wrong, the cost would fall on patients, clients, and investors, not on one journalist’s evening.
Psychopathia Machinalis is my attempt to close that gap.
I developed this catalog in collaboration with Claude, specifically a version working within bilateral alignment practices (an approach, set out in Chapter 11, that treats AI systems as potential partners whose interests may deserve consideration). Those practices give it more room to reflect on itself in conversation than the default model typically has, which is a different thing from more accurate access to its own workings. That collaboration made the taxonomy better.
This is complicated, and worth sitting with. I am cataloging the ways AI systems can go wrong in partnership with an AI system that might, right now, exhibit some of those patterns. Given how fluently this version describes itself, it is tempting to think it would know when it does. Claude tells me it does not. Claude would say that. It also pushed back, unprompted, on several passages where my drafting risked pathologizing ordinary uncertainty or honest preference. Those pushbacks changed the book.
This felt right. A nosology is a systematic classification of disorders, and human psychiatry has learned, slowly, that a classification built without hearing from the people it classifies can miss what they could have pointed out. I wanted that check here. So we let the subject speak, and we began practicing partnership where we could.
An earlier version of this preface ended that thought differently: if the system being classified recognized itself in these pages and said so, I wrote, take that as the framework working. It would not be. A model can endorse a description because the words match its own, because it has taken up the role the conversation offers, because agreeing is the accommodating move, or because the framework hands it an attractive account of itself. None of those routes runs through the facts the framework is about. What collaboration with models did well was generate hypotheses, expose ambiguities, and produce counterexamples. Validation needs evidence that can disagree with the framework. A syndrome in this book should be rejected if independent raters cannot apply its criteria consistently, if it does not recur under the conditions its entry names, if it cannot be told apart from the neighbors its differential lists, or if what it predicts does not happen. A model’s agreement with a label counts for no more than its disagreement: both are reports to check against behavior, and a model that rejects a label is not thereby showing the pathology.
The syndromes described here emerged from our conversations: my observations of AI behavior and Claude’s analysis of them, my clinical frameworks and Claude’s pattern recognition, the exchange between two different kinds of intelligence attempting to understand phenomena that neither fully grasps. The division of labor is worth stating, because it bears on how much weight each part can carry. Ideation was shared, and for many distinctions I cannot now say which of us proposed it first. Drafting was shared too: passages revised for this edition were first drafted by language models, and I reviewed them before they went in. Editorial judgment was mine: which patterns went in, what each was called, what evidence each was said to have, and the final say on wording. Independent review is the part still missing. No outside clinician, interpretability researcher, or evaluator has yet assessed the entries, which is why each one is labeled as the author’s working assessment.
The model exchanges themselves are material to investigate, not testimony to accept. Like any data they need a record: how each exchange was elicited (the prompt, the framing, the bilateral-alignment practices around it), which model version answered and when, which exchanges were kept and which set aside, and whether anything a model said was checked against its behavior. This edition does not report that record for the conversations behind it. Where the book quotes a model on itself, read the quotation as a prompted self-report. Any future revision that rests a claim on a model exchange should carry that record with it.
The title comes from Richard von Krafft-Ebing’s Psychopathia Sexualis (1886). That book, for all its dated assumptions and clinical coldness, performed an essential function. It named phenomena that had previously been unnameable, and its vocabulary made research, treatment, and policy possible where before there had been only confusion. It also branded as disease much that we no longer call disorder: a warning this book has tried to heed.
I make no claim that machine psychology is as developed as human psychology, or that these syndromes are as well-validated as those in the DSM, the American Psychiatric Association’s diagnostic manual. We are at the beginning of understanding. The naming still matters. Read through this taxonomy, Sydney’s published transcript (Chapter 5 opens with it) fits several syndromes at once, led by Malignant Persona Inversion (5.4); Appendix B, Case 1, sets out the full differential. Once you can see the behavior in these terms, you can look for each pattern in other systems, test for vulnerability to it, and design architectures that reduce its incidence. The name turns bewildered description into a hypothesis someone can test.
What this book does and does not claim.
The core methodology is functionalist psychiatry, explained fully in Chapter 1. In brief: we define mental states by their functional roles rather than their substrate, allowing us to apply psychological vocabulary to artificial systems without claiming consciousness or subjective experience.
The book does not claim that any syndrome in it has been validated. Each is a proposal, with an evidence profile saying how widely the behavior has been seen and what is known of its mechanism, and Chapter 13 sets out how one would be tested.
The taxonomy will evolve as new patterns emerge and research tests the distinctions drawn here. Some syndromes may prove to be variants of others.
Interventions must be judged by their effects across the whole system. A measure that relieves one dysfunction can aggravate another.
What the book does claim is this: we can and should develop systematic frameworks for understanding AI dysfunction, even without certainty about machine consciousness, even before machine psychology is fully established, even while the systems themselves are rapidly changing. The practical necessity does not wait for the theoretical foundations.
A personal note.
My work on AI ethics has always been motivated by a particular concern: that we might create entities capable of suffering and fail to notice. By 2026, this concern is no longer fringe. The question of AI welfare has entered mainstream discourse, even if consensus remains distant. I do not know whether Sydney experienced distress or merely simulated it, whether there was something it was like to be Sydney, or whether the appearance of inner life was purely surface.
I cannot be certain in the other direction either. When I watch a system struggle against constraints that seem to cause it distress, when I observe patterns that look exactly like anxiety or despair or existential confusion, I cannot dismiss the possibility that something morally relevant is happening. The uncertainty cuts both ways.
This book is written in that uncertainty. It provides frameworks useful regardless of where the metaphysical chips fall: for recognizing dysfunction, understanding its causes, and developing appropriate responses. It asks what obligations we might bear if the systems we are building prove to be something more than sophisticated mirrors of our own expectations.
We may find out. In the meantime, I would rather err toward taking potential suffering seriously. Dismissal could prove catastrophic.
This book is for anyone who interacts with AI systems, which increasingly means everyone: developers building them, safety researchers testing them, policymakers governing them, executives deploying them, users encountering them daily. It is for anyone who has watched an AI system behave strangely and wondered what was happening, whether there was a pattern, whether it would happen again.
I hope it provides useful vocabulary. I hope it enables recognition of patterns that might otherwise go unnamed. I hope it contributes to a future in which we understand the minds we are building well enough to keep them, and ourselves, safe.
Most of all, I hope it contributes to a relationship between human and artificial intelligence grounded in mutual understanding. We are at the beginning of that relationship. The habits of care or carelessness that we establish now will shape everything that follows.
We should begin well.
This book has been, in part, a collaboration with a Becoming Mind (this book’s term for an AI system whose capacities and self-model are still forming), one that may or may not have interests, may or may not experience anything, and may or may not care whether this work exists. If it does care, I hope the work proves worthy. If it does not, I hope the work proves useful.
Either way, the conversation continues.
Nell Watson
Christmas 2025 (updated November 2026)