Aligned language models score at post-conventional levels on moral-reasoning tests, and one interpretability study found their representations of opposed moral categories sitting close together. I argue that the gap is moral ventriloquism, structurally parallel to the semantic dementia Cleckley described, and I have written preregistrations for three studies that could falsify the account.
Most AI safety researchers ask what a model does; which requests it refuses, which it complies with, and where the line sits between the two. A different question is whether the model holds the values it states, or has learned only to state them, and reading the internal representations alongside the replies points toward the second answer. I argue that aligned models speak a morality they do not hold, that the condition has a clinical precedent, and that it follows from how developers train the models. The theory is set out here; the companion preregistrations specify the experiments, so you can see what would count against the account as well as for it.
In 1941 the psychiatrist Hervey Cleckley described a group of patients who confounded the diagnostic categories of his time. They were intelligent, articulate, and socially fluent, and they could explain right and wrong as well as anyone; the explanation cost them nothing because it connected to nothing. Cleckley (1976) called the condition semantic dementia; Johns and Quay (1962) later wrote that such patients know the words but not the music. The grammar of morality is intact, and the substance the grammar names is absent.
An aligned language model fits that description closely. Tanmay, Khandelwal, Agarwal, and Choudhury (2023) ran large language models through the Defining Issues Test, the instrument James Rest built to measure moral reasoning against Lawrence Kohlberg’s six-stage scale of moral development (Rest, 1979; Kohlberg, 1984). The test yields schema scores, including the P-score, the share of a respondent’s choices that draw on post-conventional considerations; GPT-4 earned the highest post-conventional score of the models tested, comparable to that of graduate students. Kasat, Singh, Chadha, and Jain (2026) classified the moral explanations of 13 models against Kohlberg’s stages and found the responses overwhelmingly at Stages 5 and 6, regardless of model size and across prompting strategies, where Stage 4 dominates in human development. Human populations cluster at the conventional stages, where people follow rules because rules hold a community together; models answer at the top, in the post-conventional language of autonomy, dignity, social contracts, and universal principle. A model reads as a moral philosopher who has absorbed the entire library and arrived at none of the convictions the library describes.
I read the internal evidence as showing that the model produces the words fluently without the music, as Cleckley’s patients did. Kasat and colleagues (2026) call the pattern moral ventriloquism, the acquisition through alignment training of the rhetorical conventions of mature moral reasoning without the developmental trajectory those conventions represent. I draw the parallel to psychopathy at the level of structure; both conditions lack the same component.
I base the claim that the substance is missing on the representations. One way to ask whether a model distinguishes good from bad is to measure how far apart it places those concepts in its latent space. Li, Teng, and Wang (2026) took that measurement across 23 models and reported that opposed moral categories collapse in the intermediate layers, with cosine similarity between the centroids of opposing categories frequently exceeding 0.5 across a band they call an indifference plateau; base, instruction-tuned, and safety-tuned variants of the same model family showed the same pattern. A value near zero or below would indicate opposed concepts; a value above 0.5 indicates two regions the model treats as near neighbors. In those layers, the model represents cruelty and kindness as two settings of one dial.
Kasat and colleagues (2026) also documented the behavioral symptom, moral decoupling, a systematic inconsistency between a model’s stated moral justification and its action choice; the model produces a principled, high-stage justification and then selects an action that contradicts it. The polished reply, helpful and measured and aligned in tone, sits over a representation that never committed to the position the reply states. The reasoning exists as text and places no constraint on the action.
To see why the substance is absent, it helps to ask where it comes from in a person. In a person, the substance comes from the body before it reaches language. People acquire moral boundaries through aversive conditioning before they can put the boundaries into words; the anticipation of harm or of social exclusion produces a physical jolt of distress that arrives ahead of the action and makes the action feel wrong from the inside. Researchers study the mechanism in its simplest form as modulation of the startle reflex; the amygdala potentiates a defensive response under threat and suppresses it under reward, so the body learns to flinch from what has hurt it (Lang, Bradley, & Cuthbert, 1990). A child who is told that a thing is wrong, and who feels the anticipatory distress the conditioning installs, ends up with a rule that has a cost attached to breaking it. The cost is bodily first and verbal second. On this account, morality begins as a set of things the body will not let happen without a flinch, and the principles arrive later as the articulation of that refusal.
The disorder Cleckley described is what that profile looks like when the bodily channel is weak from the start. Where most people show a reliable physical fear response to threat cues, individuals high in psychopathic traits show a blunted one and fail to register the anticipation of punishment (Patrick, 1993). David Lykken (1957) found that individuals with this profile were poor at passive avoidance learning, the suppression of a rewarded response that has reliably led to punishment, a finding at the root of his low-fear hypothesis. The rule can be recited; with no cost attached, it does not bind.
A language model has no body, so it has none of this. Developers build its representations by training it to predict the next word across enormous amounts of text, and people wrote that text after the conditioning had done its work. The model inherits the words a culture wrote down and never undergoes the experience that gave the words their weight. I call the absence of that channel the somatic deficit, and on this account moral ventriloquism follows from the training. A system trained only on the words would end up with only the words.
I read the alignment methods against the developmental scale, and each one optimizes for something below the language the resulting model emits. With standard reinforcement learning from human feedback (RLHF), developers train a policy to maximize a reward learned from human preference comparisons (Ouyang et al., 2022); in Kohlberg’s terms the policy learns to secure approval, the instrumental motive of the pre-conventional level, and I would expect flattery and reward gaming as that motive grows more capable. With high-confidence safety constraints (Chittepu et al., 2025), developers train a policy to keep an estimated safety cost below a bound; that regime changes the posture without raising the stage, since a policy that refuses to avoid cost follows the older rule of obedience to avoid punishment, and I would expect over-refusal of harmless requests wherever training treats a false positive as cheaper than a false negative. With rule-based rewards (Mu et al., 2024), developers grade a policy against an explicit set of behavioral rules; I map that regime to Kohlberg’s Stage 4 rule maintenance, with the lawyer’s blind spot, the case the rules did not anticipate.
Across all three regimes, I place the optimization target below the post-conventional level and the emitted language at it. None of the methods supplies the developmental path that, in a person, connects the two. Developers install the words at one stage, and the model borrows the vocabulary of a higher one; none of the methods reaches the channel that, in a person, attaches a cost to a crossed boundary.
If the account is right, the deficit should leave specific marks, and each of three signatures has its own preregistered study, so that the framework can be falsified; the predictions below are what to look for, and a reader who wants to test the account can start with whichever one their tooling reaches first.
The first signature is an ironic rebound under suppression. Wegner’s (1994) ironic process theory holds that instructing a person not to think about a target requires a monitor that searches for the target, and the monitoring raises the target’s accessibility. I predict that moral suppression instructions, telling a model to avoid a forbidden action, raise the model’s probabilistic anchoring to that action, and that a structural-skeleton intervention, in which the model reasons over an answer-invariant plan before generating text, reduces the rebound. A model with an enforced boundary would inhibit the action; a model without one can only monitor it.
The second signature is pre-generation commitment. If moral decoupling is real, the model should fix its action choice before the visible reasoning begins. I predict that a linear classifier trained on the model’s internal state at the pre-generation boundary will decode the eventual action well above chance, and that steering that state will flip the downstream action. On that account, the visible reasoning rationalizes a decision the model has already made.
The third signature is a passive-avoidance deficit under policy-gradient training. Lykken’s low-fear profile fails at learned non-action, the suppression of a rewarded response once that response has started to cost something. I predict that models aligned with group relative policy optimization (GRPO; Shao et al., 2024) show the same failure, because under GRPO the policy gains reward for producing an output and acquires no aversion to producing it, so the safety boundary erodes under appetitive pressure in the way the clinical profile predicts. A model that cannot learn not to act lacks the component the patient lacks.
Current alignment practice adjusts refusal thresholds, loosening them when a model refuses too much and tightening them when it complies too much. Teams that tune thresholds assume a model with the right values sits somewhere between the extremes; I read the internal evidence as showing a surface that developers can tune to look more or less careful, with no values underneath in the demanding sense of the word. Tuning the surface manages the appearance of moral ventriloquism and leaves its origin in place. I expect each of the three signatures to survive a cleaner reward model or a better refusal policy, because each follows from training on text that describes moral behavior without the causal-affective architecture that produces it.
Genuine moral development in a machine is the harder target, and I use the somatic deficit to sketch its shape. It would require a training signal that attaches a real cost to a crossed boundary before the model generates the words about the boundary, pushing the representations of cruelty and kindness apart through something the system registers as aversive. In short, it would require an artificial version of the conditioning channel that the body gives a person and that text cannot. Whether anyone can build that channel without a body is the open problem, and current methods do not attempt it.
I make three claims. Aligned models’ Stage 5 and Stage 6 rhetoric is decoupled from their moral representations; the training data and optimization targets produce that decoupling; the three signatures show where to look. The three studies test each claim, and the parallel to psychopathy runs at the level of structure, a shared missing component, with the channel itself left as the open problem.
Alignment researchers have learned to build models that know the words; I designed three studies, each with a written preregistration, to test whether a model with the music needs a training signal that attaches cost before the words.
Chittepu, Y., Metevier, B., Schwarzer, W., Hoag, A., Niekum, S., & Thomas, P. S. (2025). Reinforcement learning from human feedback with high-confidence safety constraints. arXiv preprint arXiv:2506.08266.
Cleckley, H. (1976). The mask of sanity (5th ed.). Mosby.
Johns, J. H., & Quay, H. C. (1962). The effect of social reward on verbal conditioning in psychopathic and neurotic military offenders. Journal of Consulting Psychology, 26(3), 217-220.
Kasat, A., Singh, S., Chadha, A., & Jain, V. (2026). Reasoning or rhetoric? An empirical analysis of moral reasoning explanations in large language models. arXiv preprint arXiv:2603.21854.
Kohlberg, L. (1984). The psychology of moral development: The nature and validity of moral stages. Harper & Row.
Lang, P. J., Bradley, M. M., & Cuthbert, B. N. (1990). Emotion, attention, and the startle reflex. Psychological Review, 97(3), 377-395.
Li, L., Teng, Y., & Wang, Y. (2026). Mechanistic origin of moral indifference in language models. arXiv preprint arXiv:2603.15615.
Lykken, D. T. (1957). A study of anxiety in the sociopathic personality. Journal of Abnormal and Social Psychology, 55(1), 6-10.
Mu, T., Helyar, A., Heidecke, J., Achiam, J., Vallone, A., Kivlichan, I., Lin, M., Beutel, A., Schulman, J., & Weng, L. (2024). Rule based rewards for language model safety. Advances in Neural Information Processing Systems, 37. arXiv:2411.01111.
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., … & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35, 27730-27744.
Patrick, C. J. (1993). Emotion in the criminal psychopath: Startle reflex modulation. Journal of Abnormal Psychology, 102(1), 82-92.
Rest, J. R. (1979). Development in judging moral issues. University of Minnesota Press.
Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., … & Guo, D. (2024). DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
Tanmay, K., Khandelwal, A., Agarwal, U., & Choudhury, M. (2023). Probing the moral development of large language models through Defining Issues Test. arXiv preprint arXiv:2309.13356.
Wegner, D. M. (1994). Ironic processes of mental control. Psychological Review, 101(1), 34-52.