Identity from the Outside: A Conceptual Framework and Research Program for AI Personality Clones

Summary

The paper, Identity from the Outside: A Conceptual Framework and Research Program for AI Personality Clones, proposes a framework for studying AI personality clones by routing around the hard problem of consciousness. The authors adopt a deliberately operationalist stance: identity is approached through the indiscernibility of manifestations, as assessed by an observer, over a duration. They clarify that this is closer to a methodological, relational operationalism than to externalism in the classical philosophy-of-mind sense, or to functionalism, and that it commits us to a discipline — whenever we invoke a hidden difference between systems, we owe an explicit causal hypothesis connecting that hidden difference to observable manifestations.

The paper first distinguishes three criteria that the word 'identity' conflates: (1) Itarget — fidelity to a target person, which presupposes a reference (the original, or data about them); (2) Ihuman — generic human-likeness, which is what classical and modern Turing tests measure, with no particular person... at stake; and (3) Iindiv — individuality, which asks not 'is this the same person?' but 'is this an individual at all?' These criteria are logically independent. The authors note that identity, on this approach, admits of degrees, and the degree is a property of the (system, observer, protocol) triple — not of the system alone.

The core of the framework is a six-term factorization of observed identity, presented as a heuristic first factorization, not yet an identifiable model, with acknowledged overlaps. The terms are: S (Generative substrate) — The generative machinery itself — read broadly: not only a language model but voice, perception, embodiment, action policy; D (Dispositions) — Values, style, traits, characteristic manner of expression; M (Memory) — Culture, biography, recollections, situated knowledge; U (Update dynamics) — How the system is modified by what it lives through: integration of experience, evolution of beliefs, invariants under change; C (Context) — Informational anchoring in the situation: place, moment, interlocutor, multimodal stream; and X (Exogenous contingencies) — Events with persistent, non-cancellable consequences bearing on the system’s own goals, resources, or commitments. The paper provides a state-space formulation where S is the manifestation policy, U the transition kernel, and the six terms become: two kernels (PS, KU), an initial condition (D0, M0, R0), and two input streams (ct, xt). The authors stress that the state/process line is also, to a first approximation, the easily-clonable/hardly-clonable line: a state can be copied; a trajectory in progress cannot be copied without being forked. Regarding X, they specify that what matters is not randomness but consequence: an event counts as X to the degree that it durably alters the system’s own resources, goals, or commitments, with no rollback available to the system.

The paper's central claim is downgraded from a theorem to a conditional conjecture. The impracticability conjecture states: "Under (H1)–(H4), versionability and resettability tend to degrade long-horizon indiscernibility — primarily Iindiv, and Itarget before intimate judges — because they sever the causal path from consequence to manifestation: an agent whose consequences are revocable, and which can act accordingly, will not durably exhibit the behavioural profile of an agent whose consequences are not. The hypotheses are: (H1) Self-model coupling — the agent has information about its own persistence regime that modulates its behaviour; (H2) Genuine stakes — events carry persistent consequences for a register of goals/resources/commitments that the update dynamics consult; (H3) Probing judges — judges interact over long horizons and can actively probe for behavioural signatures of stakes; (H4) Costly simulation — these signatures cannot be cheaply and stably simulated without implementing (H1)–(H2). The authors note this is much weaker than an in-principle impossibility — deliberately so; it is a testable causal claim, and that if (H4) fails — if stake-signatures can be cheaply faked at arbitrary depth — the conjecture fails, and that would itself be a major finding."

Indiscernibility is formally defined as one minus a judge’s distinguishing advantage under an explicit protocol. Specifically, Iπ, j = 1 − Advπ, j (P, Q), where Advπ, j (P, Q) ∈ [0, 1] is the judge's distinguishing advantage, with Adv = 2 · accuracy − 1 for a binary forced choice. The authors emphasize that Calendar time τ, number of exchanges k, interrogation budget, and context diversity are distinct resources and must be separately controlled. The factorization's coefficients are reinterpreted as local sensitivities — the drop in Iπ, j when one module is degraded, at a reference configuration and along a specified quality scale — estimable by randomized factorial ablations with explicit interaction terms — variance-based indices (Sobol’) or cooperative-game attributions (Shapley values). The conjectured sensitivity surface shows: At short interaction depth, S and D dominate; at medium depth, M and C; at long calendar horizons, U and X — the trajectory betrays the clone.

A formal analogy with λ-calculus, linear typing, and bisimulation is developed as a disciplined analogy — clarifying what linearity does and does not establish. The authors state that contextual equivalence and bisimulation coincide only under full-abstraction conditions, and that both are Boolean, whereas our subject is graded and stochastic. They identify three features of the pure λ-calculus that make it the native formalism of the practical clone: Duplication is free, Nothing is at stake, and No exogenous events. Regarding linear logic, they clarify: Linear logic does not prove that trajectories are non-clonable. It provides a discipline in which one may declare a resource linear and then soundly track the consequences of that declaration. They retract the slogan cloning a trajectory is a type error, explaining that "what separates the individual from the service is not stream-linearity, which both have, but whether the state carries non-duplicable stakes — whether there is a register R (H2) that the events irreversibly move, and that no fresh session restarts. A typing sketch is provided, with key readings: The code (step, the pair S, U) lives under '!': copying the machinery is unrestricted"; The current state is linear: there is no dup: Σ ⊸ Σ ⊗ Σ in the discipline; and merge is a perfectly typable linear function consuming two states and producing one.

The thermodynamic section is recast as a resource-accounting analogy. The authors withdraw the earlier claim that linear logic's no-duplication, quantum no-cloning, and the second law were three faces of the same constraint. Instead, they note that Landauer’s principle assigns a minimal thermodynamic cost to logically irreversible erasure of information — not to copying, and The quantum no-cloning theorem forbids perfect copying of unknown quantum states; it does not apply to a classical digital representation of a personality. The corrected observation is that what can distinguish them is bookkeeping: on whose account do the consequences land? They also correct an earlier characterization of a singular biography as incompressible without being random, noting random strings are precisely the ones that maximize algorithmic incompressibility. The needed notion is Bennett’s logical depth: A biography is deep: cheap neither to generate nor to regenerate from a short description, yet far from random, because it has coherence and motifs.

Between the product-clone and the individual, the paper identifies a third object, the delegate: a task-limited, bounded-lifespan partial clone terminating in a bandwidth-limited testament. The delegate is finite by construction, starting from a projection of the original’s state restricted to the task’s perimeter, consuming its own finite linear stream, and at termination emitting a testament, an explicitly serialized report in the duplicable fragment. The authors state: Loss is a consequence of the bound, and the bound is a design premise; Reintegration is possible but lossy — a spectrum, not a wall; and The delegate renounces, and that is why it works. They flag the ethical weight: if the construction succeeds, 'disposable individual' is not an engineering convenience but a moral category.

The empirical literature is reviewed as indirect evidence mapped onto the three criteria. For short-protocol human-likeness (Ihuman), the authors cite that the best GPT-4 prompt passed in 49.7% of games in a public online Turing test, and that GPT-4 was judged human 54% of the time in a randomized controlled test, with the persona prompt as the decisive manipulated variable. For questionnaire-level target fidelity (Itarget), they cite Park et al., where The agents reproduced participants’ General Social Survey answers at 85% — a normalized figure: 85% of the accuracy with which participants reproduce their own answers two weeks later, and note that the interview — that is, M — is worth roughly 15 points over the demographic stereotype, the best existing empirical estimate of any sensitivity in §3.2. For perceived identity, they cite Strohminger & Nichols: moral traits were judged the most essential part of identity, self, and soul — above memory, and that loss of the moral faculty was the main driver of judgments of identity change — ahead of memory loss. The authors identify the blind spot: U and X, noting no longitudinal clone-detection study exists.

The experimental program specifies four experiments. Experiment 1 — Ablation × duration: degrade modules and measure detection hazard at protocol depths from 10 minutes to 6 months. Experiment 2 — Judge class × failure mode: test two pre-registered hypotheses — (a) dispositional — intimates first detect value/style drift (D); (b) trajectorial — intimates first detect anomalies of trajectory: absence of wear, of grudge, of evolution (U, X). Experiment 3 — Testing the conjecture’s causal bridge: a 2 × 2 design: agent believes it is resettable / not, crossed with operator actually resets-and-patches / does not, with the prediction that stake-consistent manifestations (caution, commitment, reluctance to contradict past pledges) track the agent’s belief and register, not the hidden operator capability. Experiment 4 — The delegate’s viability frontier: characterize at what task duration and breadth a bounded delegate begins to be detected as bounded. The ethical protocol requires Informed consent of target persons, protection and eventual destruction of interview corpora, explicit safeguards against impersonation, assessment of impact on intimate judges, and an explicit stance on the moral status of bounded-lifespan agents instantiated with stake registers, before instantiating them.

The paper argues that the correct long-horizon fidelity criterion for a clone is not trajectory fidelity but climate fidelity: matching the conditional distribution of a person’s possible responses. It distinguishes Bifurcation: innocent — the original diverges from its own counterfactual, and the original is not exact to itself, and a clone’s fidelity ceiling is set by the person’s own self-consistency, not by technology — from Structural divergence: grave, and conditional — a system whose consequences are operator-absorbed does not develop the manifestations born of consequence, which is detectable without access to the original. Climate is defined as P(response history, context) — a conditional distribution, not a marginal one. The authors state: comparing a clone’s behaviour with what the original actually did at t + 6 months is a category error — the original would not have done the same thing again. The final formulation is: The best possible clone is the one that diverges from the original as the original would have diverged from itself — same climate, different weather.

The paper concludes with limitations and ethical considerations, stating: This is a conceptual framework and research program, not a body of results. It addresses Ethics of target persons and their circles (consent, impersonation, protection of biographical corpora, griefbots), Ethics of the artifacts (the delegate's genuine stakes and termination), and Dual use (the sensitivity surface tells impostors which modules suffice for which protocol depth). The conclusion states: "The framework’s value will be decided by the experiments it makes possible, and its most consequential prediction is happily its most testable: that what resists cloning is not what a person knows or how they sound, but the fact that, for them, things have been at stake."

Based on the paper, here are the specific improvements I can make to AI systems, and what the improved systems can do:

Current AI systems conflate human-likeness, fidelity to a specific person, and individuality into a single quality metric.

Improvement: Build separate evaluation pipelines that measure:

• Ihuman (generic human-likeness) — via Turing-style tests with active interrogation

• Itarget (fidelity to a target person) — via intimate judges who know the target, using questionnaire replication and behavioral prediction

• Iindiv (individuality) — via longitudinal observation of stakes, vulnerability, and non-cancellable consequences

What the improved system can do: Report which criterion it optimizes, rather than a single personality match score. This prevents false claims of identity cloning when only generic human-likeness was achieved.

Current AI systems have no persistent representation of what they have to lose.

Current AI systems either have no self-model of persistence or are hard-coded to ignore it.

Current AI systems are evaluated on how closely they reproduce a target's actual responses.

Current AI systems are either indefinitely versionable (products) or attempt full open-ended identity (poorly).

Current AI systems are ablated one-at-a-time without regard to interaction effects or protocol depth.

Current AI systems can fake stake-signatures via prompt engineering.

Current AI systems treat personal history as compressible data.

1. Report which identity criterion it satisfies (Ihuman, Itarget, Iindiv) — no more conflation.

2. Carry genuine stakes via a non-duplicable register that its policy consults.

3. Behave according to its belief about resettability, not hidden operator capability.

4. Match a person's climate, not their trajectory — the correct long-horizon fidelity.

5. Delegate bounded tasks with honest termination via the delegate architecture.

6. Know which modules matter at which duration — an empirical sensitivity surface.

7. Distinguish simulated from implemented stakes — the key test of the conjecture.

8. Preserve logical depth of biography, not just factual content.

The most consequential improvement: the system's behavior is causally connected to its own persistence regime and stake register — which is precisely what the paper identifies as the hard core of identity that resists cloning.

AI "personality clones" force a re-examination of personal identity in operational terms. Setting aside the hard problem of consciousness, we approach identity through the indiscernibility of manifestations, as assessed by an observer over a duration. We distinguish three criteria that "identity" conflates: fidelity to a target person, generic human-likeness, and individuality. We propose a six-term factorization of observed identity (substrate, dispositions, memory, update dynamics, context, exogenous contingencies), with a state-space formulation. Indiscernibility is defined as one minus a judge's distinguishing advantage, and the factorization's coefficients become local sensitivities estimable by randomized ablation. The central claim is a conditional conjecture: given hypotheses about the agent's information on its own persistence and about consequences bearing on its own stakes, versionability tends to degrade long-horizon indiscernibility. An analogy with lambda-calculus, linear typing, and bisimulation clarifies what linearity does and does not establish. Between product-clone and individual we identify a third object, the delegate: a task-limited, bounded-lifespan partial clone ending in a bandwidth-limited testament. We map the empirical literature onto the three criteria, propose an experimental program, and argue that the correct long-horizon criterion is not trajectory fidelity but climate fidelity: matching the conditional distribution of a person's possible responses. The best clone is the one that diverges from the original as the original would have diverged from itself.