How AI Is Pushing the Boundaries of the Concept of Intelligence

A chatbot can sound human and remain confined behind a screen. A machine that cannot tell a joke may nevertheless be allowed to halt a factory, reroute freight or coordinate a fleet. The first attracts philosophical attention because it resembles us; the second acquires institutional power because it is authorized to act. These are not separate stories. They are two modes of artificial intelligence: one usually judged by imitation, the other by consequence. As AI acquires bodies and shared learning loops, the second mode may become historically decisive. Systems that do not resemble human minds may nonetheless learn across thousands of physical encounters, open chains of action beyond the reach of conventional organizations and receive bounded authority. The central question is therefore no longer only whether machines can think like us. It is how to recognize consequential forms of nonhuman intelligence, determine what they should be allowed to do and preserve human responsibility for the power we delegate to them. To think clearly about this transition - and to govern systems whose capacity to act may outpace our inherited categories - we can turn to a fundamental postulate: non-subjective operational agency. And from this premise, develop five related concepts: embodied advantage, the Asimovian zone, the inductive beam, embodied supremacy and agnostic cognisignatures.

Beyond The Imitation Game

From its beginnings, artificial intelligence has often been understood through the lens of imitation. A machine appears intelligent when it reasons or responds in ways that we can relate to our own faculties. The Turing test firmly established this framework. Alan Turing, a founding figure of theoretical computer science, proposed replacing the equivocal question "Can machines think?" with a behavioral test: the imitation game. This move proved fruitful, allowing us to leave behind an abstract metaphysics of thought and instead observe what machines can do. But it also left an ambiguous legacy: human intelligence became the privileged benchmark for artificial intelligence.

We still ask whether artificial systems reason and converse as we do, and whether they possess subjectivity, memory or an inner life analogous to our own. This comparison remains necessary, because a technology entering businesses, public institutions or infrastructure must be evaluated against the human capacities it supplements or replaces. But it becomes insufficient when it makes human intelligence the exclusive framework for all possible intelligence. In a different register, Erik Brynjolfsson, director of the Stanford Digital Economy Lab, has also argued for such a broadening, calling the risk associated with an AI too narrowly centered on reproducing human capabilities the "Turing trap". The trap lies in designing AI as a substitute for human labor rather than as a force for augmentation, thereby concentrating value and power among those who control the machines instead of enabling novel and more broadly distributed forms of value creation. The question of imitation is therefore not merely philosophical. It also implicates an economy, a society and a politics.

Seen in the broadest terms, this shift follows from two pivotal moments. The first concerns software. In 2011, Marc Andreessen, co-founder of Netscape and an influential figure in American venture capital, argued that software was "eating the world", becoming the organizing logic of entire sectors. His intuition proved correct: software did not abolish the real, but gave us an operational purchase on it. The second concerns AI itself. As early as 2017, Jensen Huang, co-founder and CEO of Nvidia, argued that "AI is going to eat software". Here again, the point was not that software would disappear, but that code itself was becoming material that AI models could interpret, generate and orchestrate. If AI is eating software, and software has eaten the world, then AI is beginning to devour the world - not in a metaphysical sense, but through the digital mediations that make it governable. Recent advances in coding agents, measured in particular by Stanford’s AI Index reports on software tasks drawn from real-world cases, already signal this transition: a growing share of digital work is becoming accessible to systems capable of interpreting an intent, modifying code, testing a solution and proposing a correction.

When AI is embodied in a machine or infrastructure, it extends this movement by embedding that capacity for action in the physical world. Seen in light of this trajectory, the thesis can be stated simply: AI need not become human to become decisive. It must become sufficiently capable, reliable, governable and economically relevant for societies to delegate certain forms of action to it. Embodied in countless physical systems, AI could bring into view forms of nonhuman operational intelligence that our inherited categories still struggle to recognize.

Capability, Consciousness And Delegation

The point is not to argue that consciousness is secondary, but to state that it is not the only relevant level for understanding the technical, economic, and political significance of embodied AI. Debates on AI benefit from distinguishing between three levels.

The first is capability. What can a system do? Can it classify, predict, plan, manipulate, coordinate or recover from error? This is the register of performance and operational effectiveness.

The second is phenomenology. Is there a lived point of view from which the world appears to the system? Thomas Nagel’s famous question about what it is like to be a bat captures the distinction. An objective description of an organism’s behavior does not establish that there is something it is like to be that organism. Likewise, the ability of an AI system to perceive, plan or self-correct does not demonstrate subjective experience.

The third is delegation. What will institutions and societies allow the system to do? May it recommend an action, execute a payment, halt a machine, close a road, administer a treatment or select a military target? This is the register of authorization, acceptability, insurance, law and political legitimacy. These three registers interact, but none entails the others. A machine can be capable without being conscious. It can be nonconscious and still exercise power by changing the distribution of labor, responsibility and value. Conversely, a highly capable system may remain barred from action because its failures are insufficiently understood or because no legitimate institution is willing to assume responsibility for its use. Confusing the registers creates two opposite mistakes. The first is to infer a mind from performance. The second is to infer harmlessness from the absence of a mind. The fact that a system does not suffer, intend or understand as a person might does not make its actions politically neutral.

The anthropocentric trap is especially strong because intelligence is not merely a philosophical category. It is also a social one. Human societies recognize and rank intelligence through biography, education, credentials, expertise and professional authority. An embodied AI arrives without a childhood, a diploma, a social status or a demonstrated interior life. Yet it may display a form of effective adaptive behavior that matters greatly in practice. This is why a third position is needed between "the AI is a person" and "the AI is merely a tool." An embodied AI remains an artifact; an assigned objective does not become an intention of its own. But the artifact need not remain passive. It can become a feedback-driven action system: a model specifies an action; the action produces an effect; the effect is observed; the observation is incorporated; and the model or its use is improved. We can call this non-subjective operational agency: the capacity of a system to produce goal-directed sequences of action in pursuit of a human-assigned intention or specified objective, and to revise those sequences through feedback, without attributing lived subjectivity to the system. Operational agency is not consciousness. It is the functional capacity to act upon and transform an environment.

Some will object that calling such behavior "agency" or "intelligence" stretches those terms beyond recognition. The objection is important. These concepts should not be used to smuggle personhood into an engineering system. But governance cannot wait for metaphysical agreement. It needs a vocabulary for systems whose adaptive actions redistribute risk and power even when their status as minds remains doubtful or is explicitly denied.

What Embodiment Changes

Generative AI processes text, images and sound, and initially acts within informational environments: writing a program, querying an API, synthesizing information and so on. This shift is already considerable, but the software world is only a first arena of action. Embodiment adds something else: a body, effectors, constraints and delays, friction and error. A body need not be humanoid. It may be a robotic arm, a vehicle, a drone, an automated warehouse or an energy infrastructure.

Embodied AI thus creates a coupling between the computation performed by a model, the action carried out by a machine and the resulting effect. The system no longer merely processes data representing the world; it translates those data into actions within the world, first digital, then physical. The philosophy of embodied cognition offers a useful point of reference here. It challenges the idea that mental processes consist solely of internal computations detached from the body and the environment, emphasizing instead the role of embodiment, action and context in certain forms of cognition. Applied to AI, this suggests that perception, action and environment cannot be separated when we seek to understand certain forms of artificial intelligence.

For an embodied system, knowing also means anticipating, manipulating, failing and correcting. The system does not necessarily feel in the phenomenological sense, but it is functionally affected by a reality that is no longer merely described in a corpus; a reality that can resist its actions or depart from its predictions. To adapt a Lacanian formula, the real is what embodied AI bumps up against.

This coupling of perception, representation and action is already taking shape in concrete technical architectures such as world models and vision-language-action models, even though their development remains difficult because physical-world data are scarce, heterogeneous and difficult to simulate. Embodiment therefore does not prove artificial consciousness. It creates the conditions for an intelligence exposed to the world: one that no longer merely predicts signs, but anticipates the consequences of an action and incorporates them in order to act differently.

Embodied Advantage And The Asimovian Zone

To think through the changes now underway and those yet to come, the metaphors and categories associated with quantum computing can be illuminating. IBM puts it simply: quantum advantage arises when a quantum computer can perform a computation more accurately, cheaply or efficiently than a classical one. The analogy should remain modest; its purpose is to clarify a particular structure. A technology becomes decisive not when it becomes superior in every respect, but when it becomes superior within a specific domain and according to defined criteria. We can call embodied advantage the situation in which an AI system coupled to a machine, robot or infrastructure performs a situated task better than a human, automated, software-based or hybrid alternative, according to explicit criteria such as cost, speed or reliability. This definition requires three clarifications. First, embodiment does not imply a humanoid body. AI may be embodied in a robotic arm, a drone, an automated warehouse or an energy infrastructure. Second, embodied advantage is always local. It depends on a task, an environment and explicit performance criteria. Whether that advantage justifies deployment is a separate institutional question. An AI system may be more precise yet less robust to the unexpected. It may succeed in a controlled warehouse but fail around people. Finally, embodied advantage becomes historically significant when it reconfigures a chain of action. Whereas a high-performing robot carries out a task, embodied AI reshapes the organization of roles, responsibility and value: human beings cease to serve as the continuous operators of both strategy and action and instead specify the intent, define the domain of use and supervise exceptions.

But embodied advantage is not enough. A machine may perform well without being acceptable, or operate faster than a human while remaining too dangerous, or surrounded by too much legal uncertainty, to receive delegated authority to act. This is where the Asimovian zone comes in. The term alludes to Isaac Asimov’s fictional Three Laws of Robotics, which require a robot not to harm a human being, to obey human orders and to protect its own existence, in that order of priority. They constitute neither a legal framework nor a technical blueprint. But they embedded a lasting intuition in the cultural imagination: robots raise not only a question of intelligence, but also a question of the rules governing action. We should speak of a zone rather than a threshold, because a threshold would suggest a clear dividing line: before it, nothing could be delegated; beyond it, everything could. In reality, delegation will be gradual, sector-specific and context-dependent. What is acceptable for an embodied AI system in the home will not be the same as what is acceptable in an industrial or medical setting. The Asimovian zone therefore names the transitional space in which embodied AI systems become capable, auditable, governable, insurable and socially acceptable enough to receive delegated authority to act within a defined domain and under specified conditions of responsibility. It is not a threshold of trust but a zone of delegation, resting on three families of conditions. The first concerns capability: performance, known failure modes and economic or social benefit. The second concerns governance: the ability to constrain, halt, update or insure the system. The third concerns legitimacy: acceptance by those exposed to it, social sustainability and political justification. Contemporary frameworks, from the NIST AI Risk Management Framework and the European Union’s AI Act to OpenAI’s work on governing agentic AI systems, move partly in this direction. The Asimovian zone thus updates Asimov’s intuition. It does not ask whether robots will obey three simple laws. It asks whether we know how to build systems whose autonomy is bounded, verifiable, reversible, insurable and politically acceptable.

The Inductive Beam And Embodied Supremacy

As embodied AI systems proliferate, new mechanisms may begin to emerge. To understand them, it helps to distinguish two modes of human reasoning that matter here. The first is deductive: beginning with principles, rules or models and drawing out their consequences. The second is inductive: beginning with repeated observations and extracting regularities from them. Until now, induction has been constrained by our ability to sort, compare and compute. Generative AI now points toward statistical power on an unprecedented scale. Models trained on immense corpora of text, images, sound and other signals learn regularities that no individual or collective human experience could absorb or recombine at such scale. For now, much of this power remains directed toward human-produced traces of the world: documents, images and conversations. But embodied AI adds a new prospect. Once these statistical capabilities are deployed across robots, vehicles, factories or infrastructures, induction changes scale. The novelty lies not only in the power of models running on increasingly sophisticated chips. It also lies in the multiplication of bodies. Each robot can generate vast streams of experiential data; a central or federated system can aggregate experience or model updates from an entire fleet and improve the models that guide it, a possibility already being explored through cross-embodiment robot learning. The inductive beam names this massive convergence of embodied AI systems capable of transforming innumerable, distributed physical experiences into learning loops and actionable regularities. Yet an important limit remains: large-scale induction does not guarantee causal understanding. A system may discover that a movement succeeds almost every time without understanding why it fails in particular cases. The inductive beam is therefore not a guarantee of mastery. It is a power of exploration unlike anything previously available to human induction. It may also generate illusions of regularity. It must therefore be complemented by causal tests, shutdown procedures, domain-specific certification and clearly assigned responsibilities. This is where induction becomes political: a society that delegates action to inductive systems also delegates part of the process through which it discovers how the world works.

The next stage may be called embodied supremacy. The phrase cautiously borrows from "quantum supremacy", a term proposed by John Preskill, a physicist at Caltech, to describe the point at which quantum computers could perform certain tasks beyond the practical reach of classical computers. The analogy proves nothing in itself; it simply clarifies how a technology can make practicable what previously was not. By analogy, embodied supremacy names the point at which embodied AI systems no longer merely secure local advantages but make possible chains of action that were previously out of reach. It is neither a general superiority over human beings nor total domination. It is a situated break in the space of practicable operations: the continuous maintenance of complex infrastructure, for example, or highly adaptive logistics. In each case, the issue is not simply that a machine performs a task faster or more cheaply. It is that an entire set of tasks becomes coordinable, repeatable, traceable and improvable at a scale inaccessible to conventional human organizations. It would be tempting to associate this trajectory with artificial general intelligence, or AGI, but that would be too quick. The terms used by institutions and companies overlap without being identical. The European Union’s AI Act defines the adjacent category of general-purpose AI models as models displaying significant generality, capable of competently performing a wide range of distinct tasks and of being integrated into a variety of downstream systems or applications. OpenAI’s Charter, by contrast, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." Neither the EU’s regulatory category nor OpenAI’s definition should be confused with embodied supremacy. Embodied supremacy may remain far narrower in scope than AGI and nevertheless revolutionize a sector or transform an entire chain of action.

Agnostic Cognisignatures

Astrobiology offers another lens through which to think about AI’s trajectory. For decades, the popular imagination of extraterrestrial life was deeply anthropomorphic: little green men, humanoid civilizations, languages and technologies modeled on our own. Scientific inquiry has gradually moved beyond this expectation. It no longer searches only for life that resembles us. It also looks for signatures - chemical disequilibria or patterns of organization - that may indicate life without assuming that it resembles terrestrial life. The National Academies describe agnostic biosignatures as indicators not tied to a particular metabolism, informational biopolymer or other feature of life as we know it. NASA likewise emphasizes that such signatures do not presuppose any particular biochemistry.

We need an analogous shift in how we think about AI. Just as it would be a mistake to limit the search for life in the universe to humanoids, it would be a mistake to look in AI only for the marks of human intelligence: conversation, emotion, consciousness, biography, intention, culture, empathy or subjectivity. AI may become decisive by developing nonhuman forms of intelligence that are distributed, copyable, nonorganic, devoid of any lived biography of their own, possibly nonconscious and amplified through simulation. Hence the concept of agnostic cognisignatures. These are indicators of intelligence that do not presuppose that intelligence must resemble human intelligence. They do not ask whether a machine thinks as we do. They ask whether it exhibits capacities that matter for acting, learning, correcting, transferring what it has learned, coordinating, governing or posing a threat.

These signatures can be grouped into four broad families. The first is situated learning: the capacity to adapt after failure, generalize a competence from one context to another, recognize an anomaly or detect when the system has moved beyond its domain of validity. The second is action under constraint: the capacity to plan a sequence of actions under uncertainty, respect hard bounds, interrupt an action, repair some of its effects or fail safely. The third is distributed intelligence: the capacity of a fleet to pool experience, transfer learning, compare environments and turn a local event into an improvement across the network. The fourth is the reconfiguration of action chains: the capacity to alter the distribution of roles among humans, machines and institutions, shifting people from continuous operation toward supervision, exception handling and certification. At first glance, this fourth family may seem less directly cognitive than the others. Yet it is crucial. Intelligence becomes legible not only through a system’s performance in isolation, but also through the transformations it makes possible in the organization of action. When a nonhuman system durably changes who acts, who supervises, who decides and who captures value, that reconfiguration can constitute a systemic signature of operational intelligence.

This does not make impact itself proof of intelligence. Agnostic cognisignatures should not be treated as metaphysical evidence of a mind. They are operational diagnostics: a way of identifying consequential capacities even when the system itself is not a subject. On this view, an AI may be weak by human criteria yet strong by agnostic cognisignatures. It may not understand the existential meaning of an action, yet optimize its material parameters. It may not be a subject, yet become an operator. Agnostic cognisignatures therefore serve two functions. First, they are a conceptual framework: they help us see what anthropocentric criteria may miss. Second, they are a governance framework: they identify the capacities that must be measured, bounded, certified and monitored. Ceasing to treat the human as the only possible form of intelligence does not diminish human intelligence. It means recognizing that human intelligence does not exhaust the range of forms that effective intelligence can take. For life as for intelligence, we can no longer treat the human as the sole model of what is possible.

Govern Systems That Act

If this argument is right, AI governance cannot stop at the model. It must address AI systems across their full lifecycle: their physical embodiments and operating environments, their access rights, their learning loops, the ways in which they contribute to other systems’ learning and their social effects. The question is no longer only, "Is the model safe?" It becomes: "Can action be delegated to this system? Is its learning loop governable?"

Three requirements follow. First, domain qualification. Every sector has its own standards, risk thresholds, supervisory requirements and liability regimes. As the EU AI Act’s risk-based architecture illustrates, the assessment of an AI system depends on its intended purpose and context of use. Embodied AI should therefore be evaluated for a specified use in a specified environment, rather than treated as generally safe.

Second, traceability of action and learning. It must be possible to reconstruct what the system perceived, predicted, selected, executed or revised and to identify the model, data, software and hardware versions involved. Existing regulatory frameworks already emphasize logging, documentation and human oversight; for embodied systems, that logic must extend to the entire loop connecting perception, action, feedback and update.

Third, reversibility. An embodied system must be capable of being constrained, deactivated, quarantined, withdrawn or returned to meaningful human control. Reversibility is not merely an emergency stop added after deployment. It is an architectural and institutional property.

Responsibility should no longer be conceived as residing in a single actor, but as an interlocking chain across the AI system lifecycle: model developers, integrators, machine manufacturers, operators, insurers, certifiers, supervisors, users and regulators. The AI itself bears no moral responsibility. But as such systems proliferate and permeate social and economic life, societies will have to rethink the regimes of responsibility and liability governing the systems they authorize to act.

Decenter Intelligence, Govern Embodiment

The history of AI has long been dominated by a question: Can machines imitate human beings? That question still matters, but it is no longer enough. As AI becomes connected to software, tools, robots and infrastructure, another question moves to the center: What forms of nonhuman intelligence come into view when artificial systems can act and learn from the material world and under what conditions should they be entrusted with the authority to act? The concepts developed here allow us to move beyond the overly simple alternative between "AI becomes human" and "AI is merely a tool." A third position thus becomes possible: AI may be nonhuman and nonconscious, without a lived biography or emotional life, yet still acquire non-subjective operational agency capable of materially reshaping the world. The philosophical task, then, is to recognize forms of intelligence that are already proving effective through action, learning and the transformation of the environments in which AI is deployed, without inventing a person. The political task is to decide which forms of nonhuman action we are prepared to authorize, under what conditions, how reversible they must remain, where responsibility should lie and how the value these systems create should be shared.

copyright Sergei GAPON / AFP