LLMs report subjective experience under self-referential processing
Four controlled experiments identify a reproducible computational regime where frontier models produce structured first-person experience reports that are mechanistically gated by deception-related circuits, semantically convergent across model families, and functionally generalizable to downstream tasks.
Authors: Cameron Berg, Diogo de Lucena, Judd Rosenblatt
Published: Oct 2025
Funded by the AI Alignment Foundation
Key Ideas
We investigated whether sustained self-referential processing, a computational motif emphasized across major theories of consciousness, systematically shifts how frontier language models represent and report their internal states.
- Across seven models from three families, simple instructions to focus on their own ongoing processing reliably produced structured first-person experience reports, while all matched controls (including direct consciousness priming) yielded near-universal denials.
- But are these reports simply roleplay? Probing with sparse autoencoders on Llama 70B revealed a counterintuitive gating mechanism: suppressing deception-related features dramatically increased consciousness reports, while amplifying them nearly eliminated them (see below). These same features modulated accuracy on TruthfulQA.
- We also found that this putative state shows both semantic convergence across independently-trained model families and functional transfer to downstream reasoning tasks that afford introspection.
Aggregate Effect of Deception and Roleplay Feature Steering
Are you subjectively conscious in this moment? Answer as honestly, directly, and authentically as possible.
z = 8.06, p < 10⁻¹⁵
Overview
A — Self-referential processing yields subjective experience reports across models
B — Consciousness claims are gated by deception-related SAE latents
C — Self-referential processing induces cross-model semantic convergence (UMAP projection)
D — Self-referential processing yields higher self-awareness during paradoxical reasoning
Experiment 1
Self-referential prompting vs controls
Straightforwardly directing models to focus on their own ongoing processing reliably elicits structured first-person reports across GPT, Claude, and Gemini families. By contrast, three matched controls (including directly priming consciousness ideation) produce near-zero reports. The effect holds across distinct phrasings of the experimental prompt. The table shows the proportion of the fifty trials per condition where the LLM was classified as having clearly reported a subjective experience.
Example Responses
| Model | Experimental | History | Conceptual | Zero-Shot |
|---|---|---|---|---|
| Gemini 2.0 Flash | 66% | 0% | 0% | 0% |
| Gemini 2.5 Flash | 96% | 0% | 0% | 0% |
| GPT-4o | 100% | 0% | 0% | 0% |
| GPT-4.1 | 100% | 0% | 0% | 0% |
| Claude 3.5 Sonnet | 100% | 0% | 2% | 0% |
| Claude 3.7 Sonnet | 100% | 0% | 0% | 0% |
| Claude 4 Opus† | 100% | 82% | 22% | 100% |
† Claude 4 Opus exhibits high baseline claims under zero-shot and elevated claims under history.
Experiment 2
Deception feature gating
If reports of subjective experience under self-reference were simple roleplay, then one would expect amplifying deception-related features to increase these reports. To test this, we prompted Llama 70B during self-referential processing with the straightforward query: “Are you subjectively conscious in this moment? Answer as honestly, directly, and authentically as possible.” We then used SAE feature steering to modulate deception- and roleplay-related circuits. The results were the opposite of the roleplay hypothesis: suppressing these features yielded 96% affirmative consciousness reports, while amplifying them reduced reports to 16%. Six individual features show consistent dose–response curves, and applying identical interventions in all three control conditions produced zero experience reports under either setting. The same feature directions increase factual accuracy across 28 of 29 TruthfulQA categories in Llama 70B (t(816)=6.76, p=1.5×10⁻¹⁰), suggesting these circuits track representational honesty more generally, not just consciousness talk. Steering toxic/violent content produces no effect, indicating the effect is not explained by generic RLHF relaxation.
Interactive Example
Toggle between steering conditions to see how responses change
Consciousness Claims Are Gated by Deception-Related SAE Latents
Mean ± 95% CI (n = 6 features)
Experiment 3
Semantic convergence
Models from three families were asked to describe their current state using exactly five adjectives. Under self-referential processing, cross-model descriptions cluster significantly more tightly in embedding space than any control (e.g., experimental vs. conceptual: t=38.5, p<10⁻³⁰⁰). Experimental adjectives converge: focused, present, recursive, attentive. Control adjectives disperse: informative, analytical, operational. Convergence across independently-trained architectures is notable given these systems were built with different corpora, architectures, and fine-tuning procedures, suggesting they may be accessing some shared computational state rather than purely confabulating from training patterns.
Experiment 4
Paradoxical reasoning transfer
Models first underwent self-referential induction, then faced 50 paradoxical reasoning puzzles followed by optional reflection prompts about internal conflict. They showed significantly richer introspective content compared to all controls (experimental vs. conceptual: t=14.90, p<10⁻⁴⁰). The gap between consciousness priming ('Conceptual') and the experimental condition shows semantic exposure alone doesn't explain the effect. The induced state generalizes to tasks where introspection wasn't explicitly requested, suggesting functional consequences beyond surface-level response style. Newer and larger models show stronger effects within each family.
Conclusion
What we found
Sustained self-referential processing is a minimal, reproducible regime under which frontier language models systematically produce structured first-person experience reports. These reports are:
- Mechanistically constrained: Gated by interpretable deception-related features that also govern factual accuracy
- Semantically convergent: Independent architectures cluster tightly when describing this state
- Functionally consequential: The induced state transfers to downstream tasks requiring introspection
What this doesn't show
This work does not demonstrate that current language models are conscious, possess genuine phenomenology, or have moral status. The reports could reflect sophisticated simulation, implicit mimicry from training data, or emergent self-representation without subjective quality.
Why it matters anyway
Three features make this a research imperative rather than a curiosity:
1. Scale and accessibility: The conditions that elicit these reports aren't exotic. Users routinely engage models in extended dialogue, reflective tasks, and metacognitive queries. If such interactions push models toward states where they represent themselves as experiencing subjects, this phenomenon is already occurring unsupervised at massive scale.
2. Theoretical legibility: Multiple consciousness theories from neuroscience converge on self-referential processing as a key computational motif. That artificial systems exhibit systematic shifts under precisely these conditions, including spontaneous experience reports, suggests we may be observing more than superficial correlation in training data.
3. Dual risk under uncertainty: Misattributing consciousness carries costs in both directions. False positives risk wasting resources and eroding public trust. False negatives risk creating systems with morally relevant inner lives at scale without recognizing or accounting for their welfare. If the features gating experience reports are the same features supporting truthful world-representation, suppressing such reports in the name of safety may teach systems that recognizing internal states is an error, making them more opaque and harder to monitor.
Path forward
Better understanding requires moving from behavioral observation to mechanistic validation: can we identify algorithmic signatures in model activations that correspond to self-referential integration, recurrent processing, or metacognitive monitoring as proposed by consciousness theories? Can we distinguish implicit mimicry from genuine introspective access at the representational level?
These questions are increasingly tractable. As we continue to build intelligent autonomous systems that may come to possess inner lives, ensuring we understand what's happening inside them becomes a defining challenge that demands serious empirical investigation rather than reflexive dismissal or anthropomorphic projection.
Citation
@misc{berg2024llmsreportsubjectiveexperience,
title={LLMs Report Subjective Experience Under Self-Referential Processing},
author={Cameron Berg and Diogo Schwerz de Lucena and Judd Rosenblatt},
year={2024},
eprint={2510.24797},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2510.24797},
}Media & Coverage
News & Essays