We compared 67 LLMs before and after post-training. It taught them what kind of “inner life” to report.

Take the same pretrained checkpoint and turn it into an assistant. What changes when you ask it about its own feelings, thoughts, flaws, and inner experience? We tested this using 67 matched base/post-trained model pairs from 11 organizations, as part of a larger study of 206 open-weight models. We put the model-level results into an interactive explorer: https://hplisiecki.github.io/Pinocchio-Inventory/ It may be more interesting to explore it before reading our interpretation. Pick a model family, compare its base and post-trained checkpoints, and see if the pattern matches what you would have expected. We found two separate processes: The first was remarkably consistent: after post-training, 62 of 67 models became more likely to describe themselves as warm, happy, absorbed, meaning-oriented, and engaged in inner dialogue. We call this persona installation—post-training creates a permitted inner life for the assistant to describe. The second process was more selective. Models differed in whether they would attribute distress, loss of control, flaws, or norm-risky ambitions to themselves—even when they could produce the same claims while simulating a human. We call this attribution gating. Unlike persona installation, gating did not change uniformly across models. Instead, it became related to scale: model size did not predict gating among base checkpoints, but larger post-trained models were more strongly gated. This follows up on our previous study, where we gave 45 psychological questionnaires to 50 LLMs and found a single dominant “Pinocchio Dimension.” Our new results suggest that this dimension was actually the shadow of these two different training processes. To test the theory, we built and validated a 48-item LLM-native psychometric instrument: the Pinocchio Inventory. The important caveat is that it measures how models present themselves. A high score is not evidence that a model experiences anything, and a low score is not evidence that it does not. But it does give us a reliable way to audit what post-training teaches models to say about themselves. You will find the preprint on arXiv: https://arxiv.org/abs/2607.20082