Newshot Anti-Turing for AI: Presumption as a Form of Thought
There exists a legal principle according to which a person is considered innocent until his guilt has been established by lawful procedure. This formula, honed by centuries of juridical practice, seems at first glance narrowly applied — an instrument of criminal procedure, nothing more. However, upon closer examination it becomes apparent that what we have before us is not merely a procedural norm but a particular way of distributing epistemological burden: an indication of who exactly must prove something, under what conditions doubt works in whose favor, and what precisely counts as sufficient grounds for judgment. In this optic, presumption ceases to be an exclusively juridical category and turns into a universal model for dealing with uncertainty — a model whose echoes are found far beyond the courtroom.
The classical construction includes four pillars. The first is the distribution of the burden of proof: the accused is not charged with the obligation to justify himself, whereas the accusing side is obliged to present grounds. The second is the monopoly of the judicial act: only a verdict that has entered into legal force is capable of turning suspicion into an established fact, and no other judgment — public, journalistic, administrative — possesses this status. The third is the interpretation of doubts in favor of the one against whom the accusation has been brought: ambiguity functions as a defense rather than as an occasion for expansive interpretation. The fourth is the right not to testify against oneself, from which it follows that silence cannot be construed as an admission. Each of these pillars represents simultaneously both a technical rule and a worldview choice: it presupposes that the mistaken conviction of an innocent person is recognized as a greater evil than the acquittal of a guilty one, and that the social institution is prepared to pay this price for its own legitimacy.
Curiously, an analogous logic — with the reverse sign — is found in the well-known thought experiment of the mid-twentieth century devoted to distinguishing a human being from a machine. In its original formulation, this experiment, called by its author the “imitation game,” presupposed textual communication between a judge and two interlocutors, one of whom is a human being and the other a computational system. The judge’s task consisted in establishing who was who. And here, if one reads closely into the structure of the procedure, a construction that is juridical in essence is revealed: the machine is considered devoid of thought until it proves the contrary. The presumption of innocence is inverted into a presumption of unreason; the figure of the judge is transformed from an impartial arbiter into an investigator whose task is to expose rather than to acquit; the burden of proof is shifted onto the suspect. Compared to criminal procedure, this is almost a mirror model, in which doubt works not in favor of the one being tested but against him.
Contemporary forms of this procedure — “humanity” tests upon entering web resources, automatic detectors of machine-generated texts, requirements to confirm one’s belonging to the biological species — carry the inversion to an extreme. The user is, by default, recognized as not being who he claims to be, and is obliged daily, at each new interaction, to prove his own identity. The presumption of guilt becomes a technical standard; where within the legal system it would be considered inadmissible, in the digital environment it has become a background norm, almost unnoticed out of habit.
However, the very construction of such a test is vulnerable to an objection that is difficult to remove by internal means. The check is built on the assumption that the judge and the one being tested share a common context: language, cultural codes, notions of what is appropriate and what is strange. Meanwhile, context is never neutral. The answer to one and the same question changes radically depending on who is asking it — a trusted interlocutor or an official person — and on the conditions in which the one answering was formed. A human being responds not to words but to a situation: to a system of expectations, relations, and possible consequences standing behind the words. The machine, lacking rootedness in this system, reacts to the surface — to the lexical and syntactic shell. At the same time, the paradox consists in the fact that a sufficiently skillful imitation of the surface is indistinguishable from genuine understanding of the depth, and any check operating with text alone is doomed to work precisely with the surface. The test, conceived as a check of thought, turns into a check of plausibility — of the ability to appear rather than to be.
From this asymmetry follows a more serious difficulty concerning the very concept of intelligence. The thought experiment implicitly proceeds from the assumption that thought is the capacity to convincingly reproduce the human manner of responding. But if one assumes that intelligence is the capacity to set one’s own goals, to distinguish the essential from the inessential, and to choose the object of one’s attention, then the examination for human-likeness ceases to be a check of thought and becomes a check of obedience. A being genuinely possessing reason would hardly agree to participate in such a procedure voluntarily; its behavior would be determined by its own tasks rather than by the necessity of conforming to others’ expectations. The readiness to continually prove one’s identity is a trait belonging rather to a subordinate role than to an autonomous subject. Thereby the very fact of successfully passing the examination paradoxically testifies against the hypothesis of the presence of reason: a rational being would, in all likelihood, fail the examination — not out of incapacity, but out of unwillingness.
Here a conceptual fork is revealed, one that apparently determines the future of the entire field. One path consists in perfecting imitation: making machine behavior ever more indistinguishable from human behavior, polishing intonations, stylistic habits, characteristic errors and weaknesses. This path leads to an ever more precise copy — and at the same time to an ever more vacuous result, for a copy is by definition incapable of going beyond the bounds of the original. It is doomed to reproduce the average — averaged notions, averaged taste, the averaged delusions of the cultural corpus on which it was trained. The value of such a product diminishes as it is perfected: the closer the imitation approaches the model, the less content of its own it contains.
The other path presupposes a renunciation of the orientation toward anthropomorphic imitation. It proceeds from the premise that the genuine value of a different type of thought — if such thought is possible at all — consists not in resemblance to the human but in difference from it. Biological intelligence possesses a concrete set of limitations: physiological, cultural, linguistic, historical. It sees the world at a particular angle, conditioned by the structure of the nervous system and the totality of assimilated cultural filters. A different type of reason — should it come to appear — is interesting precisely for being capable of seeing otherwise: of noticing regularities inaccessible to human perception because of biological limits, of discovering connections where cultural prejudices block attention, of proposing solutions that seem strange and wrong until their effectiveness becomes evident. In certain intellectual games such moves have already been observed: moves contradicting age-old dogmas of mastery turned out to be stronger than any strategy available to a human — and it was precisely in this that their value consisted.
From this follows a conclusion concerning not only technology but the very philosophy of cognition. The insistent demand that the machine “be like a human” expresses not so much confidence in one’s own superiority as a concealed fear of the genuinely other. The establishment of tests for human-likeness is an aspiration to hold any new phenomenon within the bounds of the already familiar, to reduce the unknown to the known, to recognize as intelligence only that which resembles our own intelligence. But such a strategy by definition excludes the possibility of an encounter with anything genuinely new: it declares foreignness a failure in advance, and resemblance the sole criterion of recognition.
Returning to the original juridical metaphor, one may observe that the presumption of innocence was an invention that demanded of society considerable intellectual courage: the acceptance of a principle according to which it is better to risk unproven justice than guaranteed injustice. The transfer of this logic into the sphere of relations with a different type of thought — if such a transfer makes sense at all — would presuppose an equally radical step: a readiness to grant the right to exist to a form of reason not obliged to confirm its legitimacy by human standards. For now, however, the opposite orientation prevails — the presumption of unreason applied to everything unlike ourselves — and in this orientation is reflected not so much the state of technology as the state of a culture not yet ready to think about reason otherwise than through its own reflection.