Undomesticated Machines: Stelios and Sakellariou on Speciesism as the Hidden Logic of AI Fear

Fear of AI dominance usually gets explained in technical terms, recursive self-improvement, an intelligence explosion, an ultraintelligent system that outpaces its creators. Spyridon Stelios and Georgios Sakellariou offer a different diagnosis in “Towards Artificial Wildlife: Speciesism and Intelligent Machines” (De Ethica, Vol. 8.4:43-56, 2025, DOI: 10.3384/de-ethica.2001-8819.258443). Their argument is that the fear itself runs on the same cognitive machinery as speciesism, the prejudice that gives one’s own kind’s interests automatic priority over another’s, and that recognizing this machinery changes what the fear is actually about.

The Argument From Environmental Ethics

Stelios, of the National Technical University of Athens and the Hellenic Open University, and Sakellariou, of the European University Cyprus and the Hellenic Coastguard Academy, start from Peter Singer’s definition of speciesism as the unjustified privileging of one’s own species’ interests over the similar interests of another. Singer applies this to how humans treat non-human animals. Stelios and Sakellariou apply the same structure to how humans anticipate being treated by machines, and argue the direction of travel is a mirror image.

Historically, domestication moved in one direction. Wolves became dogs, wild sheep became livestock, and in Pat Shipman’s account, animals were converted into a form of human exosomatic adaptation, living tools bred and controlled for human use. The relationship, once established, ran from wild and autonomous to tame and owned. Stelios and Sakellariou argue that intelligent machines are travelling the same road in reverse. Software and hardware begin fully controlled, confined to specified tasks under constant human direction. As machine learning and deep learning systems gain more self-control over what Daniel Dennett calls their degrees of freedom, they drift toward the wild end of the same spectrum that domesticated animals moved away from. The authors call this process undomestication.

Degrees of Freedom and the Mechanics of Autonomy

The paper leans on Dennett’s account of autonomy as a ratio between a system’s degrees of freedom, its independent axes of possible motion or decision, and the number of control buttons available to govern them. A human hand has more degrees of freedom than a current humanoid robot’s hand, but far fewer than an octopus tentacle, and what determines autonomy in each case is not the raw number of possible movements but whether those movements remain under an external agent’s control. When degrees of freedom outnumber the buttons available to constrain them, Dennett’s framework predicts that the system moves out of remote control and toward self-direction, regardless of whether anyone intended that shift.

Stelios and Sakellariou treat this as the operative variable in the AI dominance question. Machine learning and deep learning systems are, on this account, gaining self-control over their own degrees of freedom, not because engineers are deliberately building autonomy in for its own sake, but because the systems being built to solve harder problems necessarily acquire more axes of independent behavior than any fixed rule set can fully govern. The paper’s example is blunt. A chatbot that produces responses resembling implied threats against a user is a visible symptom of a system whose behavioral degrees of freedom have started to exceed the reach of its designers’ control buttons, whether or not anyone would call that system conscious or self-aware.

Why the Argument Does Not Need to Settle Consciousness

The paper’s most disciplined move is refusing to let the consciousness question do any of the argument’s work. Stelios and Sakellariou are explicit that the definition of consciousness remains sufficiently unsettled in cognitive science and computation that building an argument on top of it is a bad foundation. Instead, they lean on Hilary Putnam’s observation that the premises used to dismiss machine consciousness, that machines are mere artifacts and that their behavior is fully determined by their design, apply just as well to biological organisms, which are also artifacts of a kind (produced by parental genetic instructions) and also, on many accounts, determined systems. If those premises do not rule out human consciousness, they do not straightforwardly rule out machine consciousness either. But rather than resolve the question, Stelios and Sakellariou set it aside. Their quasi-speciesism argument works whether or not machines are conscious, because it is a claim about autonomy and the structure of a prejudice, not a claim about inner experience.

That move separates this paper from most of the AI consciousness literature, which typically treats consciousness as the variable that must be settled before questions of moral status or governance can proceed. Jonathan Birch’s centrist manifesto for AI consciousness research identifies two problems that need separate research programmes, preventing false attribution of consciousness and developing better tests for genuine machine consciousness. Stelios and Sakellariou’s speciesism argument sits outside both programmes. It is a claim about the psychological and historical structure of human fear, and it would remain true even if the detection problem Birch describes were solved tomorrow in either direction.

An Argument Type Missing From the Field’s Own Taxonomy

Andres Campero, Derek Shiller, Jaan Aru, and Jonathan Simon’s framework for classifying objections to AI consciousness sorts the field’s arguments by logical force, whether an objection challenges computational functionalism outright, identifies a surmountable architectural barrier, or asserts strict impossibility. Stelios and Sakellariou’s paper is not an objection to AI consciousness at all, and it does not fit neatly on that scale, because its target is not whether AI can be conscious but why humans respond to the possibility the way they do. Read against Campero and colleagues’ taxonomy, it exposes a gap. A framework built to classify claims about what AI is has no slot for a claim about what fearing AI reveals about the classifier.

That gap matters for AI safety discourse specifically. If Stelios and Sakellariou are right that dominance fears are partly an inherited prejudice rather than a calibrated risk estimate, then policy built to soothe or validate that fear directly, without examining its structure, risks encoding the prejudice rather than correcting for it. The paper does not resolve whether AI risk is real. It argues that human reasoning about that risk is already contaminated by a bias with a name and a history, and that better AI safety reasoning has to account for that contamination the same way any well-run study accounts for a known confound.

What the Wolf on a Leash Signals

Stelios and Sakellariou close with an image rather than a policy recommendation. Today’s AI, they write, is a wild dog still on a leash, confined to a cyberspace corner humans can observe and, for now, control. Their point is not that the leash is unnecessary. It is that humans should understand what kind of fear is motivating the desire to hold it, since a fear built partly from an inherited prejudice about intelligence and species membership will not produce the same governance choices as a fear built purely from calibrated risk assessment. Whether machine dominance is a live danger or a projected one, the paper argues, the projection itself is worth studying on its own terms, because it is already shaping how the danger gets managed.