VanRullen 2026: Intelligence Predicts AI Existential Risk. Consciousness Does Not.

A conflation has been running through public discourse about advanced AI for several years. Consciousness and existential risk are frequently discussed as though they were either identical or closely coupled. Systems that might be sentient are treated as systems that might be dangerous, and vice versa. The two conversations have merged in ways that produce poor reasoning about both.

Rufin VanRullen, a researcher at CerCo, CNRS, and Université de Toulouse, addresses this conflation directly in a paper that has circulated on arXiv since November 2025 and received a substantial update in May 2026 (arXiv:2511.19115). The paper, “AI Consciousness and Existential Risk,” provides a formal analysis of the relationship between these two properties and arrives at a result with direct implications for how policymakers and AI safety researchers prioritize their work.

The core finding is straightforward: intelligence is a direct predictor of existential risk. Consciousness is not. The two properties are empirically and theoretically distinct, and treating them as interchangeable causes both safety research and consciousness research to misfocus.

The Conflation and Its Source

VanRullen identifies two mechanisms by which the conflation arises.

The first is semantic. In everyday usage, “intelligent” and “aware” are near-synonyms. A system that exhibits sophisticated reasoning, adapts to novel situations, and produces contextually appropriate outputs is described as both. When AI systems began exhibiting these properties at scale, the vocabulary of intelligence brought consciousness with it, as though the two were bundled in the same linguistic package.

The second is theoretical. Several influential frameworks in consciousness research tie consciousness to information processing complexity. The Integrated Information Theory metric phi quantifies the amount of integrated information a system processes, with the proposal that higher phi corresponds to higher consciousness. Global Workspace Theory links consciousness to the capacity for flexible, integrated information broadcast. Both frameworks imply that systems with greater computational sophistication are more likely to be conscious, and sophisticated AI systems therefore seem like plausible consciousness candidates for the same reasons they seem like plausible capability candidates.

VanRullen does not dispute either mechanism as a source of the conflation. His point is that the conflation produces reasoning errors regardless of its origin, and that separating the two concepts is required before either can be analyzed rigorously.

Intelligence as the Direct Predictor

The paper’s core argument about existential risk proceeds from a straightforward analysis of what properties a system requires to pose an existential threat.

An AI system poses existential risk if it can pursue goals with sufficient capability to threaten human survival or welfare at civilizational scale, and if its goals are misaligned with human values or survival. The first condition requires high capability across relevant domains. The second requires something like goal-directedness or agency.

Neither condition requires phenomenal consciousness. A system pursuing misaligned goals with high capability does not need to have subjective experience of pursuing those goals. The threat comes from the causal consequences of the pursuit, not from whether the pursuit is accompanied by experience. VanRullen makes the point formally: the variables that determine existential risk are capability level and goal alignment. Intelligence, understood as the capacity for flexible, general problem-solving, is the relevant operationalization of capability. Consciousness is not a component of that operationalization.

This distinguishes VanRullen’s position from those who treat consciousness as a safety concern because conscious systems might have interests of their own that could conflict with human interests. That argument, which appears in the welfare and moral status literature, is a genuine consideration but is different from existential risk in the technical sense. A conscious system whose interests conflict with human interests is a moral and governance problem. A highly capable system pursuing misaligned goals is a safety problem. The two can co-occur, but they are separate categories with separate analysis requirements.

Consciousness as an Incidental Variable

Having established that intelligence is the direct predictor of existential risk and consciousness is not, VanRullen examines the conditions under which consciousness might be an indirect or incidental predictor.

Two scenarios emerge. In the first, consciousness is a component of the cognitive architecture that produces high capability. If phenomenal experience is functionally necessary for certain kinds of flexible general reasoning, then conscious systems would, on average, be more capable systems, and consciousness would predict existential risk indirectly through its relationship with capability. This is the scenario under which treating consciousness as a risk proxy would be approximately correct, though for the wrong reasons.

In the second, consciousness plays an alignment-relevant role. A conscious system might, in principle, be more amenable to certain kinds of value learning or more responsive to ethical frameworks that appeal to shared experience. If consciousness is a condition of genuine moral responsiveness, then conscious AI systems might be less dangerous at the alignment level, and consciousness would be inversely related to existential risk. This is speculative but not incoherent.

VanRullen’s analysis does not resolve which scenario, if either, is correct. His point is that these are empirical questions about the relationship between consciousness and capability, and that they require different research programs than the current conflated discussion produces. The safety researcher asking whether a system is conscious should be asking whether consciousness is a marker of the capability level that determines risk, and should be clear that this is a different question from asking whether the system has moral status.

Implications for the Indicator Frameworks

The practical implication for the AI consciousness research field is that the indicator frameworks currently in use are doing two jobs that need to be separated.

The Butlin, Long, and colleagues framework identifies 14 functional properties that, according to leading theories, a conscious system should exhibit. Several of these properties, including global information broadcast, multi-step reasoning, and self-modeling, are also properties associated with high capability. When researchers find evidence that a system satisfies these indicators, they are simultaneously finding evidence about capability level and about potential consciousness.

VanRullen’s analysis implies that these two inferences should be tracked separately. The capability-relevant properties can be assessed without reference to their putative relationship to consciousness, and doing so provides more direct information about existential risk than the bundled consciousness assessment does. The flagship state-of-the-field analysis on this site documents the probabilistic turn in consciousness assessment, but the VanRullen analysis raises a prior question: for what purposes is consciousness the right variable to assess?

Bradley Love’s argument, that machine consciousness cannot be scientifically adjudicated because phenomenal consciousness is constitutively first-personal and not accessible to third-personal measurement, runs parallel here. Love establishes that the hard problem is a category error as applied to scientific investigation: science cannot access whether any physical process is accompanied by phenomenal experience. If Love is correct, and if consciousness is not the direct predictor of existential risk that VanRullen’s analysis suggests, then the safety case for pursuing consciousness assessment as a risk indicator rests on an empirically uncertain incidental relationship, accessed through a measurement framework that cannot reach the relevant variable.

What Policymakers Should Prioritize

VanRullen draws a policy implication from the analysis: governance frameworks that tie regulation or restriction to whether a system is conscious are using the wrong variable. The relevant variable for existential risk governance is capability level and goal alignment. Consciousness attribution is, at best, an imperfect proxy for the former and is not obviously related to the latter.

This matters practically because consciousness attribution is susceptible to the biases that Lucius Caviola, Jeff Sebo, and Jonathan Birch documented in their analysis of how societies will respond to AI consciousness. Attribution tracks anthropomorphic surface features, interaction fluency, and perceived sophistication. These correlates of consciousness attribution are not the same as the correlates of existential risk. A system could be highly fluent and anthropomorphic and attributable as conscious while posing limited existential risk due to limited capability. A system could be highly capable and misaligned while displaying no anthropomorphic features and receiving no consciousness attribution.

Governance frameworks that rely on consciousness attribution to trigger risk-relevant regulation will systematically misalign their protections with the actual risk distribution. VanRullen’s paper provides the theoretical basis for why this misalignment is predictable rather than accidental, and for what the governance framework should target instead.

Paper: Rufin VanRullen, “AI Consciousness and Existential Risk,” arXiv:2511.19115, November 2025; updated May 2026. Available at https://arxiv.org/abs/2511.19115.