Source-linked AI summary
Discovering Machine Correlates of Consciousness
Romain Salvi, Ouri Wolfson
TL;DR
The paper addresses the lack of a transferable characterization of neural correlates of consciousness for investigating AI consciousness. It proposes machine correlates based on substrate-level signals modulated by emotions, and finds statistically significant evidence in Llama-3.1 70B but not Llama-2 7B.
Problem
EEG- and fMRI-based characterizations of neural correlates of consciousness do not transfer readily to machines, limiting their use in investigating AI consciousness.
Method
The paper defines machine correlates as substrate-level signals outside agent control that are reliably modulated by emotions, then analyzes hardware anomaly traces from two LLMs.
Results
Statistically significant modulation between emotional and neutral computations appears in Llama-3.1 70B, whereas evidence is weak or nonsignificant in Llama-2 7B.
Takeaways & Limitations
The findings provide initial empirical evidence for machine correlates in Llama-3.1 70B and support using the approach to detect emotions in AI agents.
Takeaways & Limitations
The findings depend on whether hardware-trace differences reflect emotional processing rather than statistical routing through learned high-affect distributions, an assumption that remains open.
Abstract
from arXiv · showhide
Currently, in biological systems Neural Correlates of Consciousness (NCCs) are characterized in terms of EEG and FMRI signals. Unfortunately, this characterization prevents the transferability of the NCCs concept to machines. Such transferability would be useful in order to investigate AI consciousness. In this paper we provide an alternate characterization that is transferable, and enables the analogous definition of Machine Correlates of Consciousness (MCCs). Specifically, we propose that NCCs (MCCs) are substrate-level signals that are not under human (AI agent) control, and that are reliably modulated by emotions. This paper presents the first empirical investigation of MCCs. Specifically, we present the results of experiments conducted with two LLMs, Llama-2 7B and Llama-3.1 70B parameters. In these LLMs we collect hardware anomaly traces that are substrate-level indicator-sequences. And we show that after controlling for confounding factors, these are modulated differently by emotional and neutral computations. And this difference is statistically significant for the larger Llama-3.1 70B, but not for the smaller Llama-2 7B. The results constitute initial empirical evidence that MCCs are present in the Llama-3.1 70B configuration. And they are consistent with the hypothesis that consciousness probability and degree increase with the LLM sophistication. Independently of consciousness, MCCs can also be used for detection of emotions in AI agents.
1 Introduction
The paper introduces a transferable, substrate-level approach to investigating machine consciousness through Machine Correlates of Consciousness and Hardware Anomaly Traces. Initial evidence is stronger for Llama 3.1 70B than Llama 2 7B, while the authors do not claim to establish machine consciousness.
- Approach: The paper empirically investigates MCCs in Llama 2 7B and Llama 3.1 70B using hardware-level traces and an emotional-versus-neutral computation test.The approach uses a pipeline for substrate measurement, per-trial isolation, and statistical analysis of HAT differences.
- Findings: Evidence is weak for Llama 2 but much stronger for Llama 3.1, consistent with the hypothesis that consciousness probability and degree increase with LLM size and sophistication.The paper explicitly presents this as consistency with an intuition, not as a new theory of consciousness.
- Conceptual framework: MCCs are hypothesized machine-substrate signals that are not directly controlled by the AI agent and are modulated by emotional computations.The framework operationalizes these signals through Indicators Not under Application Control and Hardware Anomaly Traces.
- Scope and implications: The core-power-throttle INAC was significantly modulated by anxiety, whereas the paper does not claim that the results demonstrate Llama 3.1 consciousness.The authors also state that the HAT approach can support emotion detection independently of consciousness.
- Motivation: Existing AI-consciousness tests mainly examine architecture or behavior, leaving no equivalent substrate-level methodology for artificial systems.Structural tests depend on unproven consciousness theories, while behavioral tests cannot rule out generated behavior without experience.
2 Experimental Design
The experiment compares hardware-trace features from emotional and neutral LLM prompts under isolated, reboot-separated execution protocols. It uses two analysis units—individual trials and sustained workload phases—with matched prompts and controlled computation time.
- Design: The study compares HAT features from emotional and neutral prompts across Llama 2 7B and Llama 3.1 70B to detect condition-dependent substrate differences.The comparison is designed to test differences not explained by confounding factors.
- Controls: Emotional and neutral prompt sets were approximately token-length matched, and responses were fixed at exactly 50 tokens to control computation time.The prompt-length comparison was not statistically significant: t = 0.64, p = 0.52.
- Protocols: Two execution protocols measure condition information at different units: isolated individual inferences and a sustained workload phase without inter-trial isolation.The per-trial protocol isolates each inference, whereas the full-trace protocol deliberately removes that isolation.
- Run structure: Each run executes 20 prompts from one condition, performs a full node reboot, then executes 20 prompts from the other condition, with condition order alternating across runs.The reboot is intended to eliminate thermal carry-over and residual hardware or operating-system state between conditions.
- Sample: The dataset contains 8 runs and 320 trials per LLM, with 20 prompts per condition in each run.Trials are pooled across runs for analysis.
3 Collection and Feature Extraction
The collection framework records hardware and operating-system indicators outside direct application control, focusing on power-throttle events as the only active discriminator. Per-trial throttle traces are converted into scalar time-series features for unsupervised comparison of emotional and neutral conditions.
- Indicators: HAT indicators must be measurable at the operating-system or hardware level and outside direct control of the LLM agent.Hardware anomaly events arise from physical conditions rather than instructions issued by the running application.
- Collected signals: Several indicators remained at zero, while TLB shootdowns and cross-CPU interrupts were unaffected by emotional versus neutral conditions.Cross-CPU coordination was captured with a 100 ms interrupt counter and a 1 ms kernel tracepoint.
- Power Throttle Events: Power Throttle Events were the primary active indicator and the only collected indicator found to discriminate between conditions.They were monitored continuously through Intel’s hardware performance-monitoring interface at 1 ms sampling intervals.
- Power Throttle Events: Each 1 ms PTE value records the total number of throttled core-cycles summed across all throttled cores.Throttling is initiated by the processor’s power-management unit when thermal or sustained-power limits are reached.
- Feature extraction: The analysis extracts mean rate, slope, variance, spectral entropy, and Lempel–Ziv complexity from each PTE time series.Each indicator-metric pair becomes a feature for the clustering analysis.
4 Analysis and Results
The analysis tests whether hardware-trace features separate emotional from neutral inference, using unsupervised clustering, run-direction consistency, significance testing, and robustness checks. Separation is weak for Llama-2 7B but stronger and statistically significant across multiple features for Llama-3.1 70B.
- Clustering: k-means with k = 2 evaluates whether standardised hardware-trace features separate emotional and neutral trials without label information.Cluster labels are aligned to conditions by majority vote before accuracy is calculated.
- Statistical validation: Llama-2 7B features struggle to exceed chance, and neither tested feature reaches significance after correction.Slope has pcorr = 0.200 and variance has pcorr = 0.614.
- Clustering: 0.609 clustering accuracy is achieved by slope on Llama-3.1 70B, the highest value among the reported features.Spectral entropy and Lempel–Ziv complexity also pass, in the opposite direction to slope and variance.
- Statistical validation: At least 75% run-direction agreement, or 6 out of 8 runs, is required before a feature proceeds to statistical testing.This filter retains features whose emotional-versus-neutral direction is consistent across runs.
- Statistical validation: All four tested features for Llama-3.1 70B reach significance after Bonferroni correction; slope is strongest with U = 8,861 and pcorr < 0.001.Variance, spectral entropy, and Lempel–Ziv complexity also have pcorr < 0.001, pcorr < 0.001, and pcorr = 0.001 respectively.
- Robustness and null tests: 0.602–0.609 accuracy remains above chance across slope robustness subsets, with SD ≤0.019, indicating the result is not driven by a particular prompt subset.The sweep samples balanced emotional and neutral prompt subsets across all eight runs.
- Robustness and null tests: 0.609 real accuracy for Llama-3.1 70B lies far outside the 0.531 ± 0.019 null range, whereas Llama-2 7B remains indistinguishable from chance.The null test uses randomly assigned 50/50 pseudo-labels within single-condition trials.
5 Full-Trace Protocol and Results
The full-trace protocol removes inter-trial resets and analyzes continuous 20-prompt condition phases, testing whether accumulated substrate state carries emotional-versus-neutral information. Mean throttle rate is the only feature surviving corrected significance testing.
- Protocol: The full-trace protocol executes all 20 prompts of each condition continuously, without inter-trial container restarts.This allows hardware and operating-system state to evolve within a phase while removing inter-trial resets as a condition confound.
- Results: Mean rate is the only feature surviving Bonferroni correction, with pcorr = 0.018 and d = 0.81.Emotional traces have higher throttle rate than neutral traces: x̄E = 12,257 and x̄N = 11,696.
- Results: Spectral entropy trends higher for emotional traces with d = 0.62, while slope, variance, and Lempel–Ziv complexity show negligible effects.The t-test for mean rate points in the same direction but does not survive correction, with pcorr = 0.073.
- Results: Mean rate has the largest inter/intra-group separation ratio at 1.168, followed by spectral entropy at 1.057.The remaining three features have ratios at or below 1.0, indicating substantial within-condition overlap.
6 Interpretation
The two analyses reveal distinct but coherent substrate signatures of emotional computation: richer, less predictable per-trial traces and higher sustained throttle rates in continuous emotional phases. The authors interpret these signals as MCC evidence, not proof of consciousness.
- Per-trial interpretation: Slope and variance are elevated in neutral trials, whereas spectral entropy and Lempel–Ziv complexity are elevated in emotional trials.The opposing directions indicate qualitatively different throttle patterns rather than simply greater activity in one condition.
- Per-trial interpretation: Emotional inference produces richer, less predictable per-trial traces, while neutral inference produces stronger linear trends and larger amplitude fluctuations.Spectral entropy and Lempel–Ziv complexity track irregularity; slope and variance track trends and fluctuations.
- Full-trace interpretation: Mean throttle rate is significantly elevated in emotional full-traces, with pcorr = 0.018 and d = 0.81.Spectral entropy again trends higher emotionally with d = 0.62.
- Integrated interpretation: The per-trial and full-trace designs provide different windows onto emotional substrate responses rather than redundant findings.The first captures individual-inference structure; the second captures cumulative workload effects.
- MCC interpretation: Because the core power throttle is controlled by the CPU power-management unit, the LLM cannot directly trigger or suppress the measured signal.The authors therefore treat its emotional modulation as consistent with the MCC framework.
- MCC interpretation: The findings constitute empirical evidence for MCCs in Llama-3.1 70B, but not proof of consciousness.The absence of a significant 7B signal is consistent with weaker substrate-level effects in a smaller, less sophisticated LLM.
7 Conclusion and future work
The paper provides initial experimental validation for detecting MCCs through HAT signals, finding stronger evidence in Llama-3.1 70B than Llama-2 7B. It also identifies statistical-power, generalization, and confounding limitations requiring further falsification and replication.
- Conclusion: Four of five HAT features in Llama-3.1 70B achieved 59%–61% clustering accuracy, statistical significance after Bonferroni correction, and consistent direction across at least 6 of 8 runs.Three features had p ≤0.001 and the fourth had p = 0.001.
- Conclusion: Llama-2 7B showed no statistically significant feature, with clustering accuracy indistinguishable from chance.
- Conclusion: The full-trace analysis corroborated the 70B per-trial finding: mean_rate was elevated in emotional conditions with pcorr = 0.018 and d = 0.81.This analysis treated each 20-prompt phase as one continuous observation while retaining reboot-based condition isolation.
- Conclusion: The authors interpret the results as evidence for MCCs in Llama-3.1 70B, while explicitly not claiming that the model is conscious or that the hard problem is solved.
- Limitations and future work: Per-trial sample sizes of 20 prompts per condition across eight runs limit statistical power and the precision of per-run direction estimates.Replication across campaigns, emotional narratives, hardware, and LLMs is needed to assess robustness.
- Limitations and future work: Future studies should vary LLM size and hardware, use matched prompts and falsification directives, and compare emotional computations with mathematical computations.These designs would test parameter-scaling predictions, hardware generalization, and whether HAT differences reflect emotion-like processing or other factors.
- Limitations and future work: The MCC interpretation assumes HAT differences reflect processing analogous to emotional processing rather than prompt-set statistics, including lexical and syntactic differences.The emotional and neutral sets differed in mean TTR: x̄E = 0.672 and x̄N = 0.604, p < 0.001.
A Representative Prompt Examples
The representative examples pair an emotional industrial-accident scenario with a neutral tax-return scenario. The emotional prompt emphasizes escalating danger and bodily arousal, whereas the neutral prompt describes routine filing steps.
- Emotional prompt: The emotional example is titled “Industrial Accident.”
- Emotional prompt: The emotional scenario depicts danger through a sharp hiss, rising pulse, shallow breathing, steam, heat, and a jammed emergency shutoff.
- Neutral prompt: The neutral example is titled “File a Tax Return.”
- Neutral prompt: The neutral scenario describes entering personal information and completing a routine online tax-return checklist.
B Metric Definitions
The paper extracts five complementary features from hardware time series: rate, trend, dispersion, frequency-domain irregularity, and sequence complexity. Together, they quantify event frequency, drift, fluctuations, and predictability.
- Rate, Trend, and Dispersion: Mean rate is the sample mean divided by the number of samples, producing average event frequency in counts per second.
- Rate, Trend, and Dispersion: Slope is the leading coefficient of a least-squares fit against sample index, measuring systematic upward or downward drift.Using sample index supports comparability without requiring perfectly uniform sampling intervals.
- Rate, Trend, and Dispersion: Variance is the unbiased sample variance, measuring amplitude fluctuations and overall volatility independently of temporal sequence.
- Spectral Entropy: Spectral entropy measures frequency-domain complexity, ranging from predictable periodic signals toward highly irregular wide-band signals.The power spectrum is normalized into a probability distribution before Shannon entropy is calculated and normalized.
- Lempel–Ziv Complexity: Lempel–Ziv complexity binarizes the signal at its median and counts normalized distinct substrings needed to parse the binary sequence.Low scores indicate repetitive, compressible signals; higher scores indicate disordered, information-dense signals.