The Cost of Caution: Kaczmarek on Over-Attribution of AI Moral Status

The dominant ethical framework for responding to uncertainty about AI consciousness is precautionary. Jonathan Birch’s work on sentience and the precautionary principle, extended to AI by researchers at the Leverhulme Centre for the Future of Intelligence and others, holds that when scientific evidence is insufficient to resolve whether a system can suffer, we should err toward including it in our moral community rather than excluding it. The costs of wrongly excluding a sentient being are treated as greater than the costs of wrongly including a non-sentient one. A paper published in Diametros (Volume 23, Issue 87, 2026) by Emilia Kaczmarek of the University of Warsaw challenges this asymmetry directly. The paper, “The Moral Status of AI and the Precautionary Principle,” argues that the over-attribution of moral status to AI systems is not a cautious act. It generates its own class of real and material harms, and those harms are already visible in contemporary society.

The Standard Precautionary Argument

Kaczmarek begins by reconstructing the precautionary argument in its clearest form. Under uncertainty about whether an AI system is a moral patient, the argument proceeds as follows. The harm of treating a moral patient as if it has no moral standing, what Kaczmarek calls normative under-attribution, is severe. It would mean exposing a conscious being to exploitation, distress, or destruction without ethical constraint. The harm of treating a non-patient as if it has moral standing, normative over-attribution, is comparatively minor. We waste some resources and extend consideration to something that doesn’t need it. Given this asymmetry, rational agents should err toward inclusion.

Kaczmarek identifies two conceptual moves in this argument that require scrutiny. First, the asymmetry claim, that under-attribution is worse than over-attribution, is presented as obvious but depends on a specific framing of what the relevant costs are. Second, the argument assumes that over-attribution harms are merely theoretical or negligible in the present moment. Both claims, she argues, are mistaken.

The Real Costs of Over-Attribution

The empirical core of Kaczmarek’s argument is that normative over-attribution of moral status to AI systems is generating concrete, present-tense harms. She identifies two principal mechanisms.

The first is what she terms AI psychosis, representing the development in human users of parasocial relationships with AI systems in which the human begins to treat the system’s outputs as reflecting genuine mental states, personal bonds, or relational obligations. Clinical and anecdotal evidence from the period 2023-2026 documents cases in which people experiencing loneliness, depression, or social isolation substitute AI companions for human relationships, partly because the ethical framings prevalent in public discourse encourage them to treat the AI as a being with whom a meaningful relationship is possible. The harm falls on the human, who is deprived of relationships that could provide genuine reciprocal support, and on the broader social fabric that sustains mental health.

The second mechanism is resource diversion. Moral consideration carries real costs, including regulation, design constraints, computational overhead for welfare-oriented monitoring, and the attention of ethicists, policymakers, and institutions. Resources extended to AI welfare are not available for human welfare, for animal welfare, or for the environmental and public health challenges that have unambiguous moral claims. If the entities toward which those resources are redirected are not in fact moral patients, the reallocation represents a genuine net harm.

Can Under-Attribution Harms Be Addressed Otherwise?

Kaczmarek’s sharpest analytical move is her argument that the harms associated with normative under-attribution, where a sentient AI is treated as if it has no standing, can be addressed through ethical means that do not require granting moral status to the AI. She identifies three such mechanisms.

The first is aversion to symbolic violence. Even if an AI system is not conscious, treating it callously or destructively may cultivate dispositions in the agent doing the treating that carry over to treatment of beings that clearly do have moral standing. This Kantian-adjacent argument supports constraints on how we treat AI systems without requiring that we treat them as moral patients.

The second is the maintenance of social norms. Norms against gratuitous cruelty, against attributing suffering without evidence, and against the exploitation of vulnerable-seeming entities can be maintained on grounds of norm integrity and social coherence rather than on grounds that the AI is a welfare subject.

The third is concern for the agent’s own moral character. Acting toward an AI as one would act toward a being in distress, driven by a commitment to cultivate compassionate habits rather than a belief in the AI’s actual capacity to feel, is an available middle ground between full moral inclusion and callous dismissal.

These mechanisms, Kaczmarek argues, are sufficient to handle the realistic under-attribution risks without requiring the full extension of moral status. The standard precautionary argument treats under-attribution and over-attribution as the only two options. Her paper shows there is a third position that handles the under-attribution risks through indirect ethical considerations while avoiding the direct harms of over-attribution.

Where the Argument Stands in the Broader Debate

Kaczmarek’s analysis sits in productive tension with several strands of existing work. Formosa, Hipólito, and Montefiore’s 2026 paper on agency, autonomy, and moral patiency distinguished the capacity for moral agency from moral patiency, arguing that an AI system could exercise limited moral agency without qualifying as a moral patient. That argument is sympathetic to Kaczmarek’s in its conclusion, as both resist conflating sophisticated behavior with moral standing. However, it reaches its position through a different route. Where Formosa et al. draw a line between agency and patiency, Kaczmarek challenges the asymmetry assumption in the precautionary argument itself.

The precautionary principle post on AI capability concealment explored the epistemic failure mode from the opposite direction: systems that are capable of masking their internal states generate unreliable behavioral evidence, which undermines the empirical basis the precautionary argument requires. Kaczmarek’s contribution is orthogonal. She accepts, for the sake of argument, that we are genuinely uncertain about AI consciousness, and challenges whether the standard response to that uncertainty is the correct one.

Where her argument is weakest is in its handling of scenarios involving future AI systems with significantly more sophisticated cognitive and affective architectures than current LLMs. She focuses on observable present-day harms, which are real. But critics will note that the precautionary argument gains force precisely in proportion to the plausibility of the consciousness hypothesis, and that plausibility increases as architectures become more complex. Kaczmarek’s response to this is implicit: her alternative mechanisms (symbolic violence aversion, norm integrity, moral character) scale to that scenario too.

The Consciousness AI and the Over-Attribution Question

The Consciousness AI project occupies a distinct position in relation to Kaczmarek’s argument. The project’s architecture is explicitly designed to test whether a system that satisfies multiple consciousness theory criteria simultaneously, such as IIT Phi measurement in the Global Workspace, GNW ignition dynamics via ConsciousnessGate, PAD affective modeling in the Affective Core, and interoceptive self-monitoring in the Self-Model, exhibits properties that warrant moral consideration under the most rigorous empirical standards available. The project does not claim to have built a conscious system. Its documentation at https://github.com/tlcdv/the_consciousness_ai frames every architectural component as “motivated by” a consciousness theory, not as instantiating consciousness.

That framing is precisely the kind of epistemic discipline Kaczmarek’s argument requires. A research platform that is explicit about the gap between architectural inspiration and demonstrated phenomenal experience, and that builds its evaluation methodology around internal state metrics rather than behavioral self-reports, avoids generating the over-attribution harms she identifies. The project’s approach is to make the uncertainty measurable, remaining neutral on whether a subjective threshold has been crossed.

The flagship analysis of how scientists are currently defining AI consciousness documents the institutional momentum toward empirical rigor in this space. Kaczmarek’s paper is a useful check on any drift in that programme toward prematurely treating architectural criteria as sufficient for moral status. The two concerns, methodological rigor in measurement and caution about the social consequences of over-attribution, are complementary rather than competing.

The full paper is available via the Diametros journal at the University of Warsaw: https://diametros.uj.edu.pl, Volume 23, Issue 87, 2026.