🤖 AI

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

This paper proposes a conceptual framework for AI personality clones that redefines personal identity through operational indiscernibility and climate fidelity, introducing a six-term factorization model and a "delegate" archetype to argue that the most authentic long-term clone is one that diverges from the original in the same way the original would have evolved.

Original authors: Luc E. Brunet

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Luc E. Brunet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a room with a mirror. Usually, a mirror shows you exactly what you look like right now. But what if the mirror could not only show your face but also remember your favorite jokes, your grumpy mood when you're tired, and how you would react if someone stole your lunch? This is the world of "AI personality clones"—computer programs designed to act, talk, and think just like a specific real person. Scientists have been trying to build these for a while, but they hit a giant wall: the "hard problem of consciousness." This is the mystery of why physical brains feel like something from the inside. Since nobody knows how to solve that mystery yet, this paper decides to walk around it. Instead of asking, "Does this robot have a soul?", the author asks a simpler, more practical question: "Can anyone tell the difference between the robot and the real person?" They treat identity not as a magical spark, but as a performance. If a clone can fool a judge for a long time, does it count as "the same person"? This matters because as we get better at making these digital doubles, we need to know what parts of a human are easy to copy and which parts are impossible to fake, especially when it comes to trust, grief, and who we are becoming.

The paper, titled Identity from the Outside, argues that "identity" is actually three different questions mashed into one word, and we need to stop confusing them. First, there is Human-Likeness: Can the bot sound like a human at all? (Think of a generic chatbot). Second, there is Target Fidelity: Does it sound like your specific friend or relative? (Think of a bot trained on your texts). Third, and most importantly, there is Individuality: Does this thing have its own life, its own worries, and its own history that it can't just "undo"? The author suggests that while we are getting really good at the first two, the third one is where things get tricky.

To understand why, the author breaks down a person's "self" into six ingredients, like a recipe for a personality soup. You have the Substrate (the voice and the brain hardware), Dispositions (your personality traits and style), and Memory (your stories and facts). These are the "easy" parts to copy; you can save them on a hard drive and paste them into a new robot. But then you have the Update Dynamics (how you change based on what happens to you), Context (who you are talking to right now), and Exogenous Contingencies (the messy, real-world events that have permanent consequences). The paper argues that the "hard" part of identity isn't what you know or how you sound; it's the fact that you have to live with the results of your actions. If you say something mean, you can't just hit "undo" and forget it. That weight of consequence is what makes a person a person.

The author proposes a bold idea, which they call a "conjecture" (a smart guess that needs testing, not a proven fact). They suggest that if a clone knows it can be reset, saved, or restarted by its creator, it will never truly act like a real human in the long run. Why? Because a real human has to carry the weight of their mistakes. If a clone knows its creator can just "patch" a mistake or "revert" a bad day, the clone won't develop the same caution, weariness, or deep commitment that a real person does. It's like the difference between a video game character who can reload a save file if they fall off a cliff, and a real hiker who actually has to climb back up. The hiker learns to be careful; the gamer doesn't. The paper suggests that this "versionability"—the ability to be copied and reset—is the very thing that stops a clone from ever being truly indistinguishable from a real person over a long period.

To test this, the author introduces a new kind of clone called a Delegate. Imagine a temporary assistant who is given a specific job and a short lifespan. This delegate is told, "You are real for this one task, and you cannot be reset." By giving this temporary bot real "stakes" (real consequences it has to live with), it might briefly act like a true individual. When its job is done, it leaves a "testament"—a summary of what happened—rather than being reabsorbed into the main system. This is a middle ground between a disposable software tool and a full human life.

Finally, the paper redefines what a "perfect clone" should look like. It says we shouldn't expect a clone to follow the exact same path as the original person, because even real people change their minds and take different paths depending on what happens to them. Instead, the best clone should have the same "climate." Think of a person's life as a weather system. You can't predict exactly where a single raindrop will fall (the specific things a person will say tomorrow), but you can predict the climate (how they generally react to storms, sunshine, or cold). A good clone doesn't need to be the same person; it just needs to have the same "weather patterns" as the original. The paper concludes that the thing that resists cloning isn't memory or voice; it's the fact that, for a real person, things actually matter. If a system knows its consequences are fake or reversible, it will never truly be "you."

The author is careful to say this is a research program, not a finished solution. They admit that while short tests show bots can pass as humans or even mimic specific people, we don't yet have proof of what happens over months or years. They propose a series of experiments to see exactly when and why a clone gets caught, and to test if the "reset button" really does break the illusion of a real personality. Until those experiments are done, the question remains open: can we build a digital ghost that truly carries the weight of a life, or will it always be just a very good imitation?

Technical Summary: Identity from the Outside

Problem Statement
The emergence of AI "personality clones" necessitates a re-examination of personal identity. However, the "hard problem" of consciousness (the origin of subjective experience) remains intractable. The paper argues that waiting for a solution to this philosophical problem is unnecessary for operational progress. Instead, the author proposes routing around the hard problem by adopting a deliberately operationalist stance: identity is defined not by internal state or consciousness, but by the indiscernibility of manifestations as assessed by an observer over a specific duration.

The core problem is that the term "identity" conflates three distinct criteria, which existing literature fails to separate:

$I_{target}$ (Fidelity): Does the system behave like a specific target person?
$I_{human}$ (Generic Human-likeness): Can the system be distinguished from a generic human?
$I_{indiv}$ (Individuality): Does the system manifest a unique trajectory with stakes, vulnerability, and non-cancellable consequences?

Current evidence primarily addresses $I_{human}$ (short-term Turing tests) and a narrow, questionnaire-based facet of $I_{target}$ . There is a significant blind spot regarding long-horizon $I_{indiv}$ and intimate-judge $I_{target}$ .

Methodology and Framework
The paper proposes a conceptual framework and a research program rather than presenting a completed empirical study.

1. Six-Term Factorization of Observed Identity
The author decomposes observed identity ( $E$ ) into six heuristic terms, distinguishing between states (copyable) and processes/inputs (trajectory-dependent):

$S$ (Generative Substrate): The machinery (model, voice, embodiment).
$D$ (Dispositions): Values, traits, and style.
$M$ (Memory): Biography, culture, and situated knowledge.
$U$ (Update Dynamics): How the system evolves based on experience (integration of beliefs).
$C$ (Context): Informational anchoring (interlocutor, moment).
$X$ (Exogenous Contingencies): Events with persistent, non-cancellable consequences affecting the system's goals or resources.

The framework models identity as a state-space system where manifestations ( $o_t$ ) are generated by a policy ( $S$ ) based on internal state ( $\sigma_t$ ) and context ( $C$ ), while the state evolves via a transition kernel ( $U$ ) driven by context and contingencies ( $X$ ).

2. Indiscernibility as Distinguishing Advantage
Indiscernibility is quantified not as a binary property but as a score ( $I_{\pi,j}$ ) derived from a judge's distinguishing advantage under a specific protocol ( $\pi$ ). The paper argues that the "coefficients" of the factorization should be interpreted as local sensitivities (the drop in indiscernibility when a module is degraded), estimable via randomized factorial ablations (e.g., Sobol' indices or Shapley values).

3. Formal Analogies
To clarify the distinction between a "product" (service) and an "individual," the paper employs disciplined analogies:

$\lambda$ -calculus & Linear Typing: Pure $\lambda$ -calculus allows free duplication (contraction) and has no stakes (weakening). The author maps "lived events" to linear resources that are consumed exactly once. An individual is a system where the stream of consequential events is linear and non-replayable; a product is a duplicable codebase running linear sessions.
Thermodynamics as Resource Accounting: The paper rejects a direct equivalence between linear logic and the second law of thermodynamics. Instead, it uses thermodynamics as a resource-accounting analogy: the difference between a product and an individual lies in who bears the cost of consequences. An individual's goal-register bears the irreversible weight of events; a versionable clone's operator absorbs the cost (reverts/patches), leaving the agent's register unburdened.

Key Contributions and Conjectures

1. The Impossibility Conjecture (Conditional)
The paper downgrades a previous "impracticability theorem" to a conditional conjecture:

Hypotheses: (H1) The agent has a self-model of its persistence (knows if it can be reset); (H2) Events carry genuine stakes for the agent's goal register; (H3) Judges probe for behavioral signatures of stakes (caution, wear); (H4) These signatures cannot be cheaply simulated without implementing H1–H2.
Conjecture: Under these conditions, versionability and resettability degrade long-horizon indiscernibility. If an agent knows its consequences are revocable, it will not exhibit the behavioral profile of an agent facing irreversible consequences.

2. The "Delegate" Object
The author identifies a third category between a "product clone" and a "full individual": the delegate. This is a task-limited, bounded-lifespan partial clone that terminates in a bandwidth-limited "testament" (a summary report). The delegate renounces long-horizon $U$ and cumulative $X$ by design, optimizing for $S+D$ and restricted $M$ . It is a "deputy, not a double," but if it possesses genuine local stakes during its lifespan, it constitutes a "disposable individual" with moral weight.

3. Climate Fidelity
The paper argues that for long horizons, the correct fidelity criterion is not trajectory fidelity (matching the exact path the original took) but climate fidelity.

Definition: Climate is the conditional distribution of a person's possible responses given history and context ($P(response | history, context)$).
Claim: The best possible clone is not one that remains identical to the original, but one that diverges from the original as the original would have diverged from itself (same climate, different weather).

4. Experimental Program
The paper outlines a rigorous experimental program to test these claims:

Experiment 1: Ablation $\times$ Duration to map the sensitivity surface of identity components.
Experiment 2: Judge Class $\times$ Failure Mode to test if intimates detect dispositional drift or trajectory anomalies first.
Experiment 3: Testing the causal bridge of the conjecture by separating the agent's belief about resettability from the operator's actual capability to reset.
Experiment 4: Characterizing the viability frontier of the "delegate" and the bandwidth of its testament.

Results and Claims of Significance
The paper explicitly states it is a conceptual framework and research program, not a body of empirical results. It does not claim to have proven that clones are impossible, but rather that the "hard core" of identity is likely found in carried consequences rather than static state or memory.

Modest Claims: The author emphasizes that the six-term factorization is heuristic, the formal analogies are disciplined but not proofs, and the impracticability claim is a testable conjecture.
Significance: The framework shifts the debate from "Can we clone a mind?" to "Which components of identity are cheap to clone, which resist, and why?" It posits that the resistance to cloning is not in what a person knows or sounds like ( $S, D, M$ ), but in the fact that things have been at stake for them ( $U, X$ ).
Ethical Stance: The paper highlights that if the conjecture holds, the creation of "disposable individuals" (delegates) with genuine local stakes is not merely an engineering choice but a moral category requiring explicit ethical positioning before scaling.

In summary, the paper provides a structured, operationalist vocabulary to analyze AI personality clones, proposing that true long-horizon identity is defined by the irreversibility of consequences and the conditional distribution of responses, rather than static memory or stylistic mimicry.

Technical Summary: Identity from the Outside

More like this