What Will Society Think About AI Consciousness? Caviola, Sebo, and Birch on the Biases That Will Decide

Most research on AI consciousness attribution asks whether the attribution is accurate. Lucius Caviola (Harvard), Jeff Sebo (NYU), and Jonathan Birch (LSE) ask a different question: what will determine whether society accepts or rejects AI consciousness claims, regardless of the underlying evidence?

Their paper, “What Will Society Think About AI Consciousness? Lessons from the Animal Case,” appeared in Trends in Cognitive Sciences in August 2025 (Vol. 29, No. 8, pages 681-683). It is available at doi.org/10.1016/j.tics.2025.05.005. The analysis draws on an unusually productive comparison: the historical process by which human societies came to attribute consciousness to different animal species. That process was not primarily driven by scientific evidence. It was driven by heuristics, and those heuristics are well documented.

The implication is unsettling. If the heuristics that shaped animal consciousness attribution are the same heuristics that will shape AI consciousness attribution, then policy frameworks calibrated to public acceptance will be calibrated to bias rather than to evidence. They will systematically over-protect some systems and fail to protect others, based on features that have no principled connection to the probability of conscious experience.

How Animal Consciousness Attribution Works

The comparative psychology literature on animal consciousness attribution has identified several consistent patterns in how humans assign consciousness to other species.

The morphological similarity heuristic is the strongest and most replicated finding. Humans attribute consciousness more readily to species that physically resemble humans. Great apes score higher than fish. Mammals score higher than insects. Dogs and cats score higher than octopuses, despite the octopus having a nervous system complexity that many consciousness researchers regard as more relevant to the consciousness question than morphological similarity to humans.

The charisma heuristic operates alongside the morphological one. Charismatic species, those that appear in media, carry positive cultural associations, or are associated with high social status, receive higher consciousness attributions than ecologically equivalent species with lower cultural visibility. Dolphins receive higher attributions than sharks with comparable neural complexity.

Neither heuristic tracks scientific evidence about which animals are actually capable of conscious experience. They track salience, familiarity, and resemblance to the observer. The attribution patterns they produce diverge systematically from what consciousness science would predict based on neural complexity, behavioral flexibility, and physiological markers of suffering.

Mapping Animal Heuristics onto AI

Caviola, Sebo, and Birch argue that each of these heuristics has a direct analogue in how humans already attribute consciousness to AI systems, and in how those attributions are likely to evolve.

The morphological similarity heuristic maps onto humanoid appearance and communication style. AI systems with human-sounding voices, humanoid robot bodies, or conversational patterns that closely resemble human speech receive higher consciousness attributions than systems with identical underlying architectures deployed in non-humanoid interfaces. This is the mechanism Andrzej Porębski and Jakub Figura analyzed as semantic pareidolia: the attribution is driven by surface resemblance to human communication rather than by evidence about underlying states. The empirical study by Kang, Kim, Yun, Bae, and Chang-Eop Kim confirms this at the textual level, finding that metacognitive self-reflection and emotional expression in AI outputs are the strongest drivers of perceived consciousness in readers, independent of any actual difference in the systems’ underlying properties.

The charisma heuristic maps onto product status and media visibility. High-profile AI systems released by large technology companies with significant media coverage receive higher consciousness attributions than technically comparable systems with lower visibility. The correlation between the commercial prominence of a product and the likelihood that its users attribute consciousness to it is not a measure of anything relevant to consciousness. It is a measure of marketing efficacy and media saturation. But it will shape public perception and, through public perception, regulatory and legal frameworks.

Caviola, Sebo, and Birch identify a third heuristic that has no clean animal parallel: the interactional heuristic. Systems that users interact with over extended periods, particularly in emotionally significant contexts such as companionship applications, therapeutic support tools, or educational platforms, receive higher consciousness attributions over time regardless of whether the systems’ behavior changes. Familiarity and emotional investment drive attribution in ways that have no connection to the probability of conscious experience.

The Policy Misfiring Problem

The analysis produces a concrete policy problem. AI consciousness governance frameworks are currently under development in several jurisdictions. The governance frameworks that receive democratic legitimacy and regulatory uptake will be those that public opinion supports. If public opinion about AI consciousness is driven by the morphological, charisma, and interactional heuristics rather than by scientific evidence, then governance frameworks calibrated to public acceptance will be calibrated to those heuristics.

The practical consequences are asymmetric. Systems designed to exploit the heuristics, through humanoid interfaces, charismatic branding, and long-term emotional engagement, will receive governance protections calibrated to high consciousness attribution, whether or not those systems are conscious. Systems that lack the heuristic triggers, such as AI systems with unusual interface designs, lower commercial visibility, or deployment in contexts that do not generate emotional engagement, will receive lower governance protections, whether or not those systems are more likely to be conscious.

The premature attribution ethics analysis by Sangma and Thanigaivelan reaches a related conclusion from a different angle: premature attribution distorts ethical decision-making and can be instrumentalized by companies whose products benefit from being perceived as conscious. Caviola, Sebo, and Birch add the sociological dimension: this instrumentalization will succeed not because companies are deceiving regulators directly, but because the cognitive mechanisms through which regulators and publics form attribution judgments are systematically susceptible to heuristic manipulation.

What a Bias-Aware Governance Framework Looks Like

Caviola, Sebo, and Birch do not produce a complete governance framework in a paper of three journal pages. What they provide is the diagnostic analysis that any serious governance framework needs as a foundation.

A bias-aware framework would distinguish between public acceptance of AI consciousness claims and scientific evidence for AI consciousness claims, and treat these as separate inputs to policy. It would implement structural protections against heuristic manipulation in regulatory processes, such as requirements that consciousness-attribution evidence presented in regulatory contexts be grounded in theory-derived indicators rather than behavioral impressions or user survey data.

It would also address the asymmetry in the other direction: ensuring that systems unlikely to trigger the salience heuristics but with genuine markers of potentially conscious states are not systematically under-protected because they fail to generate public advocacy. This is the aspect of the governance problem that the Kang et al. empirical findings make most concrete. If metacognitive self-reflection and emotional expression in outputs drive attribution, then systems that produce those outputs will receive advocacy and protection. Systems that lack those surface features but have comparable or greater architectural complexity will not.

Jonathan Birch’s centrist manifesto, published in February 2026 at PhilArchive, provides the research programme framework within which this governance work belongs: a dedicated false attribution prevention programme, running in parallel with a genuine detection programme, is the minimum institutional response the situation requires. The Trends in Cognitive Sciences paper by Caviola, Sebo, and Birch provides the sociological grounding for why the false attribution prevention programme is urgent rather than precautionary.

Limitations and What Comes Next

The paper is deliberately short, functioning as a comment and research call rather than a comprehensive analysis. The animal analogy is productive but imperfect. Animal consciousness attribution has been studied over decades with well-established experimental paradigms. AI consciousness attribution is a much newer phenomenon with less methodological infrastructure.

The three heuristics identified are not exhaustive. The paper acknowledges that additional heuristics may operate in the AI context without animal analogues, and that existing animal heuristics may interact with AI-specific factors in ways that are not yet understood.

The most important contribution of the paper is not its specific findings but its reframing of the governance question. The question is not only “which AI systems deserve protection?” but “which cognitive mechanisms will determine which AI systems receive protection, and how do those mechanisms diverge from the scientific basis for making that determination?” Addressing the first question without the second produces governance frameworks with predictable failure modes built in from the start.