The concept of the Turing Test, as proposed by Alan Turing, was intended to distinguish between a machine’s intelligent behavior and that of a human. Here we extend and generalize this idea to a broader framework, the Generalized Turing Test (GTT), as a thought experiment designed to distinguish between what is natural and what is manmade. The fundamental premise of the GTT rests on the idea that ‘natural thing’ and ‘artifact’ are categorically disjunct concepts, though the line between them can become blurred in actual experience. By premise, natural things are not made, but simply found, in the literal sense that they are encountered or come upon in experience. They seem to exist independently of human creation or intervention. Artifacts, on the other hand, are made; they are products of human agency and definition, though they might also be found in the above sense. Following Vico’s makers-knowledge principle, an artifact should be exhaustively knowable by the agent that made it. In contrast, the properties and relationships of a natural thing are indefinite for any cognitive agent.

In principle, finding and making are distinct relationships of subject (or agent) to object. In practice, they are ambiguous in some situations. For instance, in quantum measurement, it can be unclear whether the observer is finding or making the experimental result, since the observer physically intervenes in ways that affect the result. For another example, because it is not known how “neural networks” produce their results, it is unclear whether the programmer is making or finding the results. You can know that something has been made, because you made it or witnessed it being made. But a thing you find cannot be assumed natural simply because you did not knowingly make it. The matter is complicated by the fact that your perception of the world (in contrast to the world itself) is also an artifact—produced by your nervous system! Is the world that appears to you found or made?

A model is an artifact that simulates a found thing by attempting to formally reproduce its properties and relationships. A model is a product of an agent’s definitions. It consists of a finite list of properties and relationships, which are themselves products of definition. Any model can be exhaustively modelled, because it is well-defined. In principle, an artifact and its model(s) are finitely complex; any artifact or simulation constructed from a model can be perfectly simulated. In contrast, a natural thing may be indefinitely complex; it cannot be perfectly simulated because no model can list all its properties and relationships. On the other hand, there is no logical limit to the complexity and completeness of models, or thus to the apparent realness of simulations. In principle, any given thing can be simulated so effectively that a given cognitive agent (with its limited resources) cannot distinguish between the model and its target. However, there are practical limits to modeling and simulation, which involve limited computational resources. Deterministic chaos, for example, can be modeled only for a limited period of time before diverging from expectation. The question is whether these resources are sufficient to pass the GTT in a given instance—which means to convince a cognitive agent that the thing in question is natural.

William Paley’s watchmaker argument for intelligent design invokes an obvious difference between a rock, found on the ground during a walk in the forest, and a pocket watch found lying beside it. However, modern technology blurs this distinction: an artificial rock could theoretically be assembled through nanotechnology—or more conventionally, as with man-made diamonds. Machines can now be so complex and sophisticated that they appear natural, even organic. We can no longer rely on ordinary cognition to conclusively judge the difference between nature and artifice, especially when there is an intention to obscure the difference—as with generative AI and chatbots. Moreover, the distinction is only meaningful because we already have a category of ‘made’ (or ‘fake’) to contrast with ‘found’ or (‘genuine’). Such categories depend on conscious human agency. Absent a GTT, fake only need be good enough to fool our natural cognition.

Suppose we happened to live in an entirely artificial world—for example, a virtual reality, as some people imagine is possible. If so, everything encountered during the stroll through the virtual forest would seem “natural.” (That would be the sole category of existence until something is “made” in the virtual world by someone in that world fashioning it from the “natural” ingredients available there.) We may add to this concern the idea that “reality” is not an ontologically fixed concept or category. The “realness” with which our normal experience of the external world is imbued serves an evolutionary function for biological cognitive agents; its epistemic utility is relative to changing context. Historically, it refers to what affects us (humans) physically and what we can affect. As we co-exist ever more with artifacts—even conceptual ones—these become “real” as we interact with them, and come to seem “natural” as our new environment.

Of course, conventional ways to test for naturalness already exist. An object or substance can be analyzed chemically and structurally. (For example, there are microscopic tests to distinguish man-made from natural diamonds.) However, such procedures would not necessarily reveal a thing’s origin, granted the possibility that any natural chemistry or structure can be simulated to a finer degree than the resolving capabilities of the test procedure. While certain patterns (e.g., tree rings and other growth patterns) do characterize natural things, these too can be imitated. Though idealization, perfect symmetry, and over-simplification do characterize man-made things, a simulation could intentionally avoid obvious idealization, perfect geometric forms, or perfect regularity well enough to fool even a vigilant observer. Pseudo-randomness can be deliberately introduced to imitate naturalness. (The challenge then becomes to distinguish real from pseudo randomness.)

At least on the macro scale, natural things have individual identity as variations from a type. Manufactured items are intended to be identical, but minor imperfections may distinguish them. Yet, even such telltale marks can be simulated. An object might be deemed natural because it is older than any plausible agent that could have made it; or found in some location where there could not have been any previous agents. This does not strictly disprove agency, however, since absence of evidence is not evidence of absence. Robots or bioengineered organisms might display preprogrammed or abnormal behavior that seems incompatible with evolutionary adaptation. But this is relative to earthbound human expectations, which might not apply in alien environments. It also begs the question of whether “evolutionary adaptation” must be natural and how well it could be simulated.

Apart from specific conventional tests and their limitations, an absolute GTT would ideally determine, in an unrestricted way, whether any given thing or experience is natural or artificial. But is that feasible? If (a) all the properties and relationships of a given item can be listed, then it should count as an artifact. Similarly, if (b) it can be shown that not all the properties and relationships of an item can be listed, or that the list is infinite, then the item is by definition natural. If (c) a new property or relationship is found outside the list given in a model, then the item does not correspond to that model, yet could still correspond to some more complete model, augmented by at least the new property. But, just as it cannot be proven that all crows are black, it cannot be proven that all properties have been listed. So (a) is no help. In regard to (b), while it can be shown that not all properties have been listed (such as by finding a new property), this does not prove that the list could never be complete—that no further properties can be found. Finding such a new property, as in (c), does not establish that the item in question is natural, nor does failure to find it establish that the item is artificial. Hence, an absolute GTT does not seem feasible. There is still the option of relative GTTs, whose power of discrimination need only be superior to that of humans and superior to the power of the simulation to deceive.

External things can be put in various real situations that would test whether their response is unnatural or seems limited by inadequate computational resources. On the other hand, if the agent is having an experience from within what is suspected to be a simulation, the agent can look for glitches in the experience presented, as telltale errors that stand out with respect to the norm of previously known reality. Within the confines of the VR experience, however, the agent must have reliable memory of such a reality. (This poses a recursive problem, since the memory could itself be part of the virtual reality: the dilemma facing our brains all the time.) Similarly, digitation has a bottom grain (pixelation), which can be noted with reference to a known finer-grained “reality.” As above, however, there must be a perceivable or remembered experience of a contrasting reality outside the VR experience to serve as norm. In the case of the brain’s natural and normal simulation (i.e., phenomenal experience), there is nothing outside it to serve as norm for comparison. Digitation and discontinuity within the nervous system are normally ignored or glossed over when functionally irrelevant, as manifested in the visual blind spot and other forms of perceptual adaptation and “filling in.” Thus, normal perception is transparent. It does not normally occur to us that we are living in the brain’s simulation.