After the Fork: Why an Artificial Replica Cannot Continue Developing as a Specific Human

In this essay, I argue that an artificial intelligence modeled on a specific human cannot independently preserve the human particular developmental continuity after their histories diverge. Imagine that a person provides an AI system with every available detail of their life: memories, written messages, private thoughts, personal conversations, verbal and nonverbal expressions, preferences, values, decisions, and the consequences of those decisions. Imagine further that the system receives complete information about the person’s physical condition, including how bodily sensations, emotions, illness, and fatigue have influenced their thoughts. At the moment of activation, the AI might appear to be a psychological replica of the human from whom it was constructed. Yet divergence may begin with its very first experience. The human may confront the uncanny existence of an artificial replica, while the AI confronts the realization that it is a replica. These are related but fundamentally different experiences. How would either experience transform the human or the AI, and could the other system fully reproduce that transformation? More importantly, what happens when the human and the AI stop exchanging information and begin living separately?

A critical part of this essay is distinguishing three related concepts: imitation, prediction, and continuation. Imitation occurs when an AI reproduces a person’s recognizable language, behavior, preferences, or manner of reasoning. By identifying patterns in the person’s recorded past, the AI might respond as that person previously responded in familiar circumstances. It might also generalize those patterns to unfamiliar circumstances, but behavioral resemblance alone would not establish that the AI understands or experiences the situation as the human does. Prediction goes further by anticipating how the human will respond to a new event. A prediction may be accurate, but one correct prediction does not demonstrate that the AI and human share the same developing identity. Continuation requires more than isolated resemblance or successful prediction: it requires the AI to undergo a connected sequence of experiences and transformations that remains synchronized with the human’s evolving life. This essay argues that an AI could imitate a human and make many accurate predictions while nevertheless failing to continue as that particular person. Once the AI and human experience an event differently, each enters the next event with a different psychological history. As these differences accumulate, later predictions become based on an increasingly incomplete representation of the person whom the AI was originally designed to reproduce.

I argue that an AI could reproduce a human’s recorded memories, dispositions, and behavioral patterns, but could not independently continue developing exactly as that particular human after their experiential histories diverge. Hume’s problem of induction establishes that past behavioral regularities cannot guarantee how someone will respond to genuinely novel circumstances. Humean bundle theory, Benovsky’s pluralism, and Buddhist reductionism further suggest that new experiences do not merely provide information to an already completed self. They help constitute the person who is developing. Sartrean existentialism adds that the influence of an experience depends partly on how a situated individual interprets it and integrates it into an evolving understanding of the self. A deterministic model of the mind presents a serious objection by proposing that sufficiently complete knowledge of a person’s prior state could make their future development predictable. However, determinism does not establish that two systems receiving different embodied, social, and experiential inputs will follow an identical developmental path. An artificial replica could therefore become a psychological descendant of the human, inheriting significant features of that person’s past without becoming the same person the human later becomes. The fork occurs when the trained AI is activated and begins processing new experiences independently of the human. Before this moment, the system contains a representation of the human’s recorded past, whose accuracy is limited by the completeness and quality of the information provided. This representation may include the person’s memories, values, dispositions, and previous decisions, but the AI has not lived the experiences from which that information originated. After activation, the human and the AI begin encountering events along separate experiential paths. An event affecting the human may not affect the AI, and even when they encounter similar events, each may interpret them differently. Every new experience can modify the memories, expectations, and dispositions through which the next event is understood. Their development is therefore path-dependent: a difference arising at one moment becomes part of the psychological background influencing later interpretations and actions. The AI may initially resemble a detailed copy of the human’s recorded past, but after the fork, the human and the AI become two distinct continuations of that past rather than one synchronized developing person.

Suppose that neither the human nor the AI knows how to play chess at the moment of separation and that both learn independently during the following year. The human’s development may be shaped by experiences surrounding the game: forming a friendship with another player, meeting an unusually effective teacher, feeling humiliated after a painful loss, or gaining confidence during a winning streak. Experiences outside chess may also affect how the human plays. A breakup, accident, illness, or professional success could alter the person’s patience, confidence, tolerance for risk, or ability to concentrate. Meanwhile, the AI could study chess while using its model of the human’s prior dispositions to estimate how that person would approach the game. It might even limit its training so that it plays not as well as possible but approximately as the human would. Nevertheless, matching the human’s skill level or recognizable style would not reproduce the particular experiences through which the human learned. After one year, if the human and AI encountered the same chess position, would they select the same move, and would they select it for the same reasons? This thought experiment does not by itself prove that their choices must differ. However, it does demonstrate how even a constrained activity can become integrated into two increasingly different experiential histories, despite the AI’s access to decades of information about the human’s past.

The chess example illustrates that learning is not merely the accumulation of propositions. On both sides of the fork, the human and the AI might learn how the pieces move, study common openings, and recognize tactical patterns. Even if they acquire the same propositional knowledge, however, there is no guarantee that they will retain, interpret, or apply it in the same way. The human also undergoes experiences outside chess that may influence concentration, confidence, motivation, and willingness to take risks during a game. Because these experiences accumulate and interact, they may transform how the human approaches chess without changing any of its formal rules. The divergence may become even more obvious if a contingent event causes the human to stop learning after only a few days. A family obligation, health problem, or new priority might redirect the human’s attention, while the AI, having no corresponding experience, continues studying throughout the year. The AI would then become a stronger player precisely because it had failed to develop as the human did. Therefore, the thought experiment reveals how a relatively small difference can redirect one developmental path and produce further differences over time. As these differences compound, the human and AI become increasingly unlikely to retain the same knowledge, motivations, interpretations, and patterns of action.

The relevance of David Hume’s problem of induction lies in what it reveals about the limits of behavioral prediction. Hume distinguishes matters of fact, which concern the empirical world and are learned through observation and experience, from relations of ideas, which can be established through reason alone. Mathematics and deductive logic belong primarily to relations of ideas: if their premises are true and their reasoning is valid, their conclusions necessarily follow. Matters of fact, however, require inductive reasoning whenever we move beyond present observation and memory. Humans form expectations from repeatedly experienced patterns, while AI systems identify statistical regularities in previously recorded data. Neither process guarantees that an established pattern will continue under future conditions. Hume argues that the assumption that the future will resemble the past cannot be demonstrated deductively, since a different future remains logically possible. Nor can it be justified inductively without circularly assuming the reliability of induction itself. Consequently, inductive reasoning may make a behavioral prediction highly probable, but it cannot provide the logical guarantee available in a valid deductive argument (Hume, 1748/2007).

An AI designed to replicate a specific human would use the person’s recorded past to imitate established behavior and predict responses to unfamiliar situations. After activation, it would also interpret new information through patterns derived from that historical record. Hume’s problem of induction therefore becomes central to evaluating whether the AI could continue developing as the human. Every prediction about the person’s future implicitly assumes that previously observed dispositions will remain relevant: that a cautious person will remain cautious, that established values will retain their importance, or that familiar experiences will continue to produce similar responses. However, even if the AI possessed a complete record of the person’s past, that record could not deductively determine how the person would respond to a genuinely novel event. The AI might correctly predict the action the human eventually performs, but the person’s past behavior does not logically guarantee that prediction. Hume’s argument does not deny the practical usefulness of induction; humans and artificial systems both rely upon it successfully. Instead, it shows that predictive accuracy, however extensive, cannot by itself establish a flawless continuation of the person’s reasoning, responses, and development (Hume, 1748/2007).

In my previous essay, I used the term genuine discontinuity to extend Hume’s problem of induction to situations in which an event transforms the assumptions underlying an established pattern (Desjardins, 2026). A genuine discontinuity does not make inductive reasoning impossible, but it exposes its limitations because the regularities learned from the past may no longer apply under the new conditions. Consider a CEO who ordinarily makes cautious, data-driven decisions. An AI replica trained on the CEO’s history would recognize this regularity and would probably predict another cautious response. The sudden emergence of a transformative technology, such as large language models, could nevertheless create a genuine discontinuity by threatening the company’s existing business model. The CEO might conclude that an unusually risky strategy is necessary to prevent the company from declining. The AI could search for analogous situations and generate a statistically reasonable response, but no quantity of information about the CEO’s previous caution could logically determine how the CEO would interpret this unprecedented event. Indeed, the human CEO might also remain uncertain until actually confronting the situation. Moreover, the decision and its consequences could permanently change how the CEO understands risk, thereby influencing later decisions in ways the AI did not anticipate. The absence of an exact precedent does not create Hume’s problem of induction. Rather, the discontinuity makes the problem especially visible by demonstrating that even extensive past regularities cannot guarantee their continued relevance.

Hume’s problem of induction establishes that past events, decisions, and behavioral patterns cannot guarantee the future. His argument does not demonstrate that prediction is impossible or necessarily inaccurate. Instead, it shows that induction cannot rationally justify its own future reliability. A deductive justification fails because a future unlike the past remains logically possible, while an inductive justification becomes circular because it uses induction’s previous success to defend its continued reliability. This problem does not apply in the same way to valid deductive reasoning, in which a conclusion necessarily follows from true premises. Predictions about a human’s future development, however, concern matters of fact and therefore depend upon induction. Hume’s argument consequently establishes that an AI cannot derive a logically guaranteed continuation of a specific human from that person’s recorded past. This epistemological limitation does not by itself prove that the AI and human must diverge. To explain why different experiences may produce different developing selves, it is necessary to consider Hume’s bundle theory (Hume, 1748/2007).

According to Hume’s bundle theory, introspection does not reveal a simple and permanent self existing beneath experience. Instead, it reveals a continually changing collection of perceptions, including sensations, emotions, thoughts, desires, and memories. We ordinarily regard ourselves as enduring persons because memory, resemblance, and causal connections link these perceptions into an apparently continuous life. Nevertheless, the self is not an unchanging object to which new experiences are merely added. The changing bundle of perceptions is what constitutes mental life. This account strengthens the argument against exact artificial continuation. After the fork, the human and AI receive different perceptions, form different memories, and develop different emotional and causal connections among their experiences. Their bundles therefore begin to differ, even if they originated from the same recorded history. The AI may preserve a recognizable pattern derived from the human’s past, but as each follows a different stream of perceptions, it can no longer reproduce the particular bundle that constitutes the human’s developing self (Hume, 1739–1740/2000).

Benovsky’s pluralist theory of the self goes beyond Hume’s bundle theory. Hume describes the self as a changing bundle of perceptions, but Benovsky questions whether it is necessary to posit even one bundle that contains or unifies those perceptions. According to Benovsky, there are experiences and mental states, but no additional substance, owner, or bundle existing beyond them. The self is therefore literally a plurality: it consists in the many experiences and mental states themselves rather than in a separate entity constructed from them. This distinction strengthens the fork argument. After separation, the human and the AI begin undergoing different sequences of experiences and forming different mental states. Because neither sequence belongs to an additional, unchanging self capable of remaining identical across both paths, the two pluralities increasingly differ. The AI may inherit representations of the human’s earlier experiences, but its subsequent states belong to its own developing plurality. On Benovsky’s account, the AI therefore becomes not a continuation of the same singular self but a distinct plurality derived from the human’s recorded past (Benovsky, 2014).

Applying Benovsky’s pluralism to the chess thought experiment further demonstrates why the human and AI diverge. The human’s relationship with chess consists not only of factual knowledge about the game but also of numerous associated experiences: interactions with a teacher, emotional reactions to victories and defeats, bodily sensations during competition, and events outside chess that affect motivation and concentration. Together, these experiences and mental states form a plurality different from that of the AI replica. The difference would remain even if the AI later received a complete textual account of the human’s year. Reading a description of an experience is not identical to undergoing it. Each produces a different mental state and therefore contributes differently to the plurality constituting the self. The AI might consequently know what happened to the human and might even imitate the behavior produced by those events, but its experiential constitution would not be identical to the human’s. For example, the sensation of anxiety before performing a risky move during a chess match. Benovsky’s pluralism therefore suggests that the fork produces two distinct selves with separate experiential histories, even when those selves occasionally reach the same conclusions or perform the same actions (Benovsky, 2014).

Alternatively, Galen Strawson’s Pearl View preserves a distinction between an experience and the subject who undergoes it. A pearl does not represent one individual experience. It represents a short-lived mental self capable of undergoing a unified episode of experience. According to this view, a human life does not contain one mental self that persists continuously through every experience. It contains many successive subjects of experience, existing one after another like pearls arranged on a string. The string therefore represents a sequence of selves rather than one enduring self composed of every pearl. Applied to the fork, the human and the AI would already possess distinct subjects at the moment of activation, even if their memories and mental contents were initially equivalent. Afterward, each would generate its own succession of short-lived subjects under different biological, social, and experiential conditions. Therefore, the relevant question is not whether the AI can preserve one original pearl indefinitely, but whether it can independently reproduce the human’s entire future sequence of subjects. Because each later subject emerges from a different causal and experiential history, the AI’s sequence may resemble the human’s without being identical to it. Although Benovsky identifies the self with a plurality of experiences and Strawson describes a succession of short-lived subjects, both theories imply that copying a person’s existing mental content does not reproduce that person’s future experiential history (Strawson, 1997).

Both Benovsky and Strawson rejecting of the existence of one permanent and unchanging self emphasis on impermanence that also appears in Buddhist philosophy. In particular, the interpretation by Mark Siderits (2003, 2007) of Buddhist reductionism offers a framework that takes this idea of momentary existence a step further by questioning whether a distinct 'self' exists behind these experience-episodes at all. However, Buddhist reductionism does not claim that persons are completely nonexistent. Rather, it distinguishes between the person who exists conventionally in everyday life and the permanent, independent self that cannot ultimately be found. The Buddhist doctrines of anatta, or no-self, and anicca, or impermanence, reject the idea that a person possesses an unchanging essence that persists beneath every experience. What we call a person is instead a continually changing process of physical and mental events connected through causal relations. This perspective creates another difficulty for exact artificial replication: there is no static self that can be captured at one moment and then reproduced unchanged into the future. A flame provides a useful analogy. We identify it as one continuing flame, although its material composition changes from moment to moment and its development depends upon fuel, oxygen, and surrounding conditions. Similarly, a person remains conventionally recognizable while changing in response to bodily, social, and psychological conditions. An AI could copy a representation of the person at the moment of the fork, but once it encounters different conditions, it participates in a different causal process and therefore cannot preserve the human’s exact development, later interpretations and decisions.

The Buddhist concept of dependent origination holds that phenomena arise and change in dependence upon particular causes and conditions. Physical and mental processes are therefore neither isolated nor permanent; each exists through its relationships with other conditions. From this perspective, the self should not be understood as a single independent entity. It is better compared to a spider’s web whose interconnected strands include the body, sensations, social relationships, emotions, desires, previous experiences, interpretations, and habits. No individual strand constitutes the person by itself, and a change in one part of the web may affect many others. However, the web should not be mistaken for another permanent self existing behind these conditions. It represents the continually changing network through which the conventional person develops. This interdependence creates a fundamental problem for exact artificial replication. Even a comprehensive record of the human’s past captures only the conditions that existed before the fork; it cannot give the independently developing AI the same network of future bodily, social, and experiential conditions (Siderits, 2007).

After the fork, the human and AI participate in two distinct causal networks. Even their precise states at the moment of separation may be difficult to equate because the human’s identity depends upon a complex network of bodily, mental, social, and environmental conditions that the AI can represent without necessarily reproducing. From that moment forward, both develop under changing conditions, but they do not occupy the same body, environment, or relationships. Consequently, they cannot be assumed to encounter the same causes or to interpret similar events under identical conditions. Minor differences do not require every observable decision to differ immediately, but they become part of each system’s causal history and may influence later experiences and responses. As these histories develop separately, the human and AI become distinct processes of dependent origination. The AI may remain causally related to the human through the information from which it was created, but it cannot independently remain the same developing self as the human (Siderits, 2007). Jean-Paul Sartre rejects the idea that human beings follow a completely predetermined path. Although individuals do not control the circumstances into which they are placed, they remain responsible for how they interpret those circumstances and respond through action. Sartre describes these given conditions as facticity, but he denies that they constitute a fixed essence that determines what a person must become, the transcendence. Past experiences, established dispositions, and habits may influence a choice, yet they do not fully specify its outcome. This is especially significant in a novel situation, where previous behavioral patterns may provide limited guidance and the individual must assign meaning to unfamiliar circumstances. Even in recurrent situations, Sartre would maintain that following a habit remains a way of responding for which the individual is responsible. A person’s future actions therefore cannot be reduced entirely to patterns extracted from that person’s past (Sartre, 1943/1956).

In the CEO example, the emergence of large language models as a disruptive technology might provoke a response unlike the CEO’s previous cautious decisions. The CEO could interpret the disruption as an opportunity requiring a pivot or radical transformation of the company. Sartre explains human action through the tension between facticity and transcendence. Facticity includes the CEO’s past decisions, established character, responsibilities, and present circumstances. Transcendence is the capacity to project possibilities beyond what those existing facts appear to prescribe. The CEO remains situated within a history but is not completely defined by it. An AI replica could model that history and predict how it might constrain the decision, but it could not determine from the past alone which future possibility the CEO would adopt. Therefore, the difficulty is not calculating how much facticity or transcendence the human will use, but it is predicting how the human will interpret the given situation and choose what it means for the future (Sartre, 1943/1956). However, a determinist could object to the Sartrean account of freedom and to the thesis of this essay. According to determinism, every human decision results from prior physical and psychological causes. If an AI could reproduce the human’s complete internal state and accurately model the surrounding environment, it could theoretically reproduce the same decision. From this perspective, a genuine discontinuity would not represent a break from causation but only a situation whose relevant causes were previously unknown. The apparent novelty of the CEO’s decision, for example, might conceal a complete causal explanation involving the CEO’s neural state, memories, character, social pressures, and perception of the available evidence. The AI’s failure to predict the decision would then reveal missing information or insufficient computational power rather than genuine freedom. Gathering and processing every relevant cause might remain practically impossible, but the determinist could maintain that exact artificial continuation is theoretically possible. Divergence would therefore be a technical limitation rather than a necessary result of the fork (Kane, 2005).

Even if determinism is accepted, it does not establish that the human and AI will develop identically after the fork. Determinism might imply that identical states under identical causal conditions produce identical outcomes, but the replica does not occupy the human’s body, environment, relationships, or first-person experiences. Differences in these conditions become part of each entity’s causal history and may influence later decisions. The replica could inherit the human’s recorded memories, values, dispositions, behavioral patterns, relationships, and unfinished projects, but inheriting this information would not preserve one identity across both entities. It would instead create a copy: a new entity causally derived from the human’s past but developing through its own subsequent conditions. Avoiding this divergence would require the AI to receive continuous updates about the human’s physical, mental, and social states. This creates a dilemma: either the AI develops independently and accumulates a distinct causal history, or it remains synchronized and therefore depends upon the human’s continuing development. Once synchronization ends, the AI necessarily proceeds from its own history rather than continuing the particular life of the human.

In conclusion, an AI modeled on a specific human could imitate that person and predict many future actions, but it could not independently continue learning and developing exactly as the human does after their experiential histories diverge. Hume’s problem of induction demonstrates that patterns in a person’s past cannot logically guarantee future responses. Benovsky and Strawson provide different accounts of the self, but both challenge the assumption that copying existing mental content preserves one identity into the future. Buddhist doctrines of anatta, anicca, and dependent origination further describe the person as an impermanent process developing through changing causes and conditions rather than as a fixed entity available for replication. Sartre emphasizes that individuals interpret their circumstances and project new possibilities instead of merely executing patterns established by the past. Even determinism cannot guarantee identical development once the human and AI receive different bodily, social, and experiential inputs. After the fork, each new difference becomes part of a distinct causal history that may influence subsequent interpretations and decisions. The AI therefore becomes a psychological descendant of the human rather than the person whom that human later becomes. A perfect record of who someone has been cannot determine who that person will become, because new experiences do not merely provide information to the self: they participate in creating the self. A remaining question that remain unresolved is with continuous synchronization, could we preserve a behavioral replica, even if it could not preserve personal identity?

References

• Benovsky, J. (2014). I am a lot of things: A pluralistic account of the self. Metaphysica, 15(1), 113–127. https://doi.org/10.1515/mp-2014-0008
• Desjardins, P. (2026, August 20). LLMs and genuine discontinuities: Hume’s problem of induction. Patrick Desjardins—Essays and Notes. https://patrickdesjardins.com/philosophy/llms-and-genuine-discontinuities-hume-s-problem-of-induction
• Hume, D. (2000). A treatise of human nature (D. F. Norton & M. J. Norton, Eds.). Oxford University Press. (Original work published 1739–1740)
• Hume, D. (2007). An enquiry concerning human understanding (P. Millican, Ed.). Oxford University Press. (Original work published 1748)
• Kane, R. (2005). A contemporary introduction to free will. Oxford University Press.
• Sartre, J.-P. (1956). Being and nothingness: An essay on phenomenological ontology (H. E. Barnes, Trans.). Philosophical Library. (Original work published 1943)
• Siderits, M. (2003). Personal identity and Buddhist philosophy: Empty persons. Ashgate.
• Siderits, M. (2007). Buddhism as philosophy: An introduction. Ashgate.
• Strawson, G. (1997). The self. Journal of Consciousness Studies, 4(5–6), 405–428.