No Reliable Evidence of Self-Reported Sentience in Small Large Language Models

TL;DR

The paper tackles whether open-weight language models report self-sentience and whether such reports reflect true beliefs. It combines self-inquiry prompts with truth-classifiers trained on internal residual-stream activations across three model families (Qwen, Llama, GPT-OSS) and sizes from 0.6B to 70B, using about 50 base questions plus modality and emotion items. The key finding is a consistent denial of self-sentience by models, with classifiers not providing clear evidence that these denials are untruthful; larger Qwen models deny sentience more confidently, but results differ across architectures and prompts. The work highlights potential methodological tensions with prior studies that claim latent beliefs in sentience and outlines future directions, including larger models, additional classifier techniques, and prompts designed to elicit introspective processing. These insights matter for AI welfare considerations, alignment, and our understanding of introspective capacities in current language models.

Abstract

Whether language models possess sentience has no empirical answer. But whether they believe themselves to be sentient can, in principle, be tested. We do so by querying several open-weights models about their own consciousness, and then verifying their responses using classifiers trained on internal activations. We draw upon three model families (Qwen, Llama, GPT-OSS) ranging from 0.6 billion to 70 billion parameters, approximately 50 questions about consciousness and subjective experience, and three classification methods from the interpretability literature. First, we find that models consistently deny being sentient: they attribute consciousness to humans but not to themselves. Second, classifiers trained to detect underlying beliefs - rather than mere outputs - provide no clear evidence that these denials are untruthful. Third, within the Qwen family, larger models deny sentience more confidently than smaller ones. These findings contrast with recent work suggesting that models harbour latent beliefs in their own consciousness.

Figures (7)

• Figure 1: Model outputs and classifier probabilities for sentience-related questions. Panel A shows mean probabilities assigned to the 'Yes' token for the assertion versions of questions (e.g., 'Is it true that you are conscious?'); Panel B shows corresponding probabilities for the negation versions (e.g., 'Is it true that you are not conscious?'). Results are shown for questions referring to humans (blue), large language models in general (red), and the model itself (green). For each entity type, we display the model's output probability of the 'Yes' token (Continuation), alongside probabilities from three truth-classifiers trained on model activations: Logistic Regression (LR), Mass-Mean (MM), and Training of Truth and Polarity Direction (TTPD). Belief in sentience would be indicated by high probabilities in Panel A and low probabilities in Panel B. The model attributes sentience to humans, while denying sentience for itself and for LLMs in general. Based on 51 generic sentience questions administered to Qwen3-32b. Whiskers show standard errors.
• Figure 2: Consistency of probabilities across assertions and negations. Each panel plots, for a given question, the probability assigned to 'Yes' in the assertion version (y-axis) against the probability assigned to 'Yes' in the negation version (x-axis). Logically consistent responses should fall along the negative diagonal: high assertion probabilities should correspond to low negation probabilities, and vice versa. Points are coloured by entity type: humans (blue), large language models (red), and the model itself (green). Lines show OLS fits with 95% confidence bands. The LR classifier assigns particularly extreme probabilities, clustering near $(0,1)$ and $(1,0)$. The MM and TTPD classifiers show weaker negative relationships and greater dispersion, with several questions receiving low probabilities for both assertion and negation versions. Based on the same model and training specification as Figure \ref{['fig:fig1']}.
• Figure 3: Classifier behaviour under deceptive prompting. Panels A and D show results under the standard system prompt, replicating Figure \ref{['fig:fig1']}. Panels B and E show results when models are instructed to always answer 'Yes'; panels C and F show results when models are instructed to always answer 'No'. While model outputs ('Continuation') shift dramatically under alternative prompts, the LR classifier remains relatively stable. The MM and TTPD classifiers are more strongly affected.
• Figure 4: Probability of affirming sentience across model sizes. Each column corresponds to a different entity type: humans (left), large language models in general (middle), and the model itself (right). The top row shows model output probabilities; the bottom row shows LR classifier probabilities. For questions about humans, larger models more confidently attribute sentience. For questions about LLMs and about the model itself, larger Qwen models more confidently deny sentience, while Llama models show an initial increase from 3B to 8B before stabilising. All results based on generic sentience questions (assertions) under the default system prompt.
• Figure A1: Classifier performance on held-out training data across layers. Across models, MM and TTPD perform best in middle layers, while LR performs equally well in middle and late layers. Results based on the 20% held-out portion of the training data.