Machine Correlates of Consciousness Move the Search From Behavior to Hardware

Where should scientists look for evidence of machine consciousness if the usual answers, brain imaging and behavior, do not transfer to machines. Romain Salvi and Ouri Wolfson propose a third place. In “Discovering Machine Correlates of Consciousness” (arXiv:2608.28824, posted 28 August 2026, published in Springer LNCS 16855 as Proceedings of the 19th Conference on Artificial General Intelligence, DOI 10.1007/978-3-032-33195-3_18), they define Machine Correlates of Consciousness, or MCCs, as substrate-level signals that the AI agent does not control and that are reliably modulated by emotions. Their first empirical study then reports finding such signals in a large language model.

The definitional move is the paper’s contribution, and the empirical result is its test. Neural Correlates of Consciousness in biological research are characterized through EEG and fMRI, modalities with no machine equivalent. The authors generalize the definition instead of the instrument. An NCC becomes any substrate-level signal outside the system’s own control that emotion reliably modulates. For a machine, the substrate is hardware, so the candidate signals are hardware signals.

The instrument is the hard drive’s error log

The measurable substrate signal Salvi and Wolfson use is the hardware anomaly trace, the stream of corrected errors, retries, and low-level faults that computing hardware logs continuously. These traces are generated by physical processes below the model’s access, which satisfies the not-under-the-agent’s-control condition by construction. A language model cannot edit its DRAM’s error corrections the way it can edit its self-description.

The experiment collected hardware anomaly traces while two LLMs, Llama-2 7B and Llama-3.1 70B, performed two kinds of computation, emotional and neutral. After controlling for confounding factors, the traces were modulated differently under emotional computation than under neutral computation, and the difference was statistically significant in the larger model.

The authors state what this pattern does and does not support. The result is initial empirical evidence that MCCs are present in the Llama-3.1 70B configuration, and it is consistent with the hypothesis that consciousness probability and degree increase with model sophistication. Size dependence is a correlation in one comparison, not a law. Two models, one architecture family, one finding.

Why the definition is designed for transferability

The paper’s logic answers a problem the measurement literature has circled for years. The site’s coverage of the indicator framework developed by Butlin, Long, and colleagues records how theory-derived indicators are built for AI systems. Florentin Koch’s calibration critique then records the problem with the whole programme, that no indicator has independent validation and no ground truth of artificial phenomenality exists. Salvi and Wolfson attack a different link in that chain. The classical NCC concept does not transfer because it is defined by instruments tied to biology. Their redefinition makes the concept substrate neutral by attaching it to the two things any substrate has, signals the agent does not control, and emotional modulation.

The emotional anchor is deliberate. Emotion is the component of the biological NCC picture that survives the transfer, because affective modulation of low-level signals exists on both sides of the biological-machine divide. The authors note that independently of consciousness, MCCs are usable for detecting emotions in AI agents, which gives the method a value even if the consciousness reading never firms up.

Why self-reports cannot do this job

The design principle behind MCCs is easiest to see against the measurement class it avoids. A model’s own reports about its states are the cheapest available signal, and the field has documented why they are the least trustworthy. Kristina Šekrst’s argument, covered in the electric fata morganas analysis, is that a language model’s self-reports of emotion or sentience fall under the definition of hallucination, produced by the same sampling machinery that produces confident error, which makes any future machine consciousness epistemically inaccessible through that channel. Keith Frankish’s illusionist position on LLM first-person reports reaches the same practical conclusion from philosophy of mind. The report is exactly as real as any other output, and it is generated by processes that the report does not illuminate.

The two conditions in the MCC definition answer that problem by construction. The not-under-the-agent’s-control condition removes the channel the model can be trained to populate, and the reliable-modulation condition replaces a one-off claim with a covariation the experimenter can measure across many trials. A signal that the system cannot edit and that tracks affect across conditions is a different kind of evidence from a statement the system produces on request. That is the same control that the site’s indicator discipline demands when it requires properties that survive controls for mimicry.

Comparison to The Consciousness AI

The Consciousness AI project runs its substrate layer in software, with Leaky Integrate-and-Fire neurons in Layer 1 of the Substrate Console, under the project’s position that consciousness is an emergent property of organized dynamics and is substrate independent. That position cuts both ways here. It licenses the paper’s redefinition, because if consciousness depends on organization rather than material, then correlates of that organization should be definable at any substrate, hardware traces included. It also imposes a caution the paper shares. A hardware anomaly that tracks emotional computation is a correlate, a measurable accompaniment, and the leap from correlate to conscious state is exactly the inference the field’s mechanistic interpretability work on internal states handles with explicit care.

The project has no hardware-anomaly instrumentation documented. The architecture is software on conventional hardware, so the MCC instrument has no documented application target in this codebase. What transfers is the design pattern. A signal below the agent’s control boundary that reliably covaries with the system’s affective states is the kind of measurement the project’s affective core could be subjected to at the software level, for example in scheduler or memory-allocation traces. That adaptation is this analysis, not the paper’s, and it remains untested.

Limits

The study has two models and one positive result, both from the same architecture family. Confound control in hardware traces is hard, and the paper’s own framing marks the evidence as initial. The definition inherits the deep problem of all correlate-based programmes, that modulation by emotion is only evidence of consciousness if the emotion term is doing conscious work, which is the assumption the calibration critique would press. What is established is a transferable definition with a first positive measurement, and a new place to look that does not depend on the model’s own reports. The next tests the definition needs are the obvious ones. Different architectures, different substrates, and a check of whether the modulation survives when the emotional content is matched for token statistics, since a correlate that tracks surface features of emotional text would be a detector of topic, not of state.