How Much of You Can an AI Reproduce?

Part III taught a system to adapt to a person without reproducing them. Now the boundary gets pushed deliberately: given enough evidence — interviews, policies, artifacts, interaction history — how much observable behaviour can an AI reproduce, and how is that claim measured? Prediction, not identity. This chapter measures the first and refuses the second so completely that Chapter 31 inherits a clean, hungry question.

The first discipline is vocabulary. “Reproduce” must not float between style, preferences, values, decisions, and identity. The chapter fixes a reproduction surface of observable targets — language/expression, declared preferences, factual self-knowledge, value judgements, task choices, scenario decisions, multi-step behaviour — governed by one rule:

Success on one surface licenses no claim about another.

Prose imitation teaches nothing about decisions. Survey prediction teaches nothing about sequential behaviour. Yesterday’s preferences teach nothing about changed situations. And no surface teaches identity. The book’s internal term is predictive replica or behavioural approximation — “digital twin” may appear as the literature’s word, but never as an earned conclusion, because it asserts equivalence no current experiment justifies. The chapter law: a model predicting some outputs of a person is a predictive model of those outputs, not a copy of the person. Never: wanted, understood, became, possessed identity. “The AI is 78% of you” is banned outright.

The strong positive: Park, measured against self-replication

Park et al.’s Generative Agent Simulations of 1,000 People (Stanford HAI brief May 2025, verified in depth: 1,052 US-representative participants, two-hour qualitative interviews with adaptive follow-ups, full transcripts injected with an LLM, evaluated on GSS core module, 44-item Big Five, five economic games, five experiments) is the chapter’s strong bounded result — and its exact wording is the methodology:

Agents replicated participants’ GSS responses 85% as accurately as participants replicated their own answers two weeks later — comparably on personality and experiments, less biased than prior simulation tools.

Not “85% accurate at reproducing people” but 85% of human self-replication. Hence the chapter’s first-principles move: the baseline for reproducing a person is not perfect consistency, because the person is not perfectly self-consistent. Three comparisons structure every claim — MODEL↔PERSON-NOW, PERSON-NOW↔PERSON-LATER (the stability denominator), MODEL↔PERSON-LATER — with both numbers kept visible, never collapsed into one clone score. Model limitation (retest high, prediction low) and target instability (both low — no stable mapping exists to reproduce) are different findings with different remedies, and the predictability map across target families (preferences, values, decisions, scenarios, style) exposes which is which.

The brakes: values, chains, caricature

Three counterweights prevent the positive from inflating. BehaviorChain (Li et al., Findings ACL 2025, pp. 15738–15763, verified: 15,846 behaviours across 1,001 personas, iterative inference in dynamic scenarios, SOTA models struggling) establishes that survey/bounded replication ≠ continuous behavioural reproduction — sequential divergence compounds, one wrong step-2 prediction rewriting the chain by step 6. IndieValueCatalog (Jiang et al., ACL 2025 long, pp. 6757–6794, verified: frontier models at 55–65% on novel individual value judgements from value-expressing samples, demographics insufficient) bounds value prediction specifically. And the personality-emulation study (Scientific Reports, s41598-024-84109-5, verified: GPT-4 convergent validities 0.90–0.94 but internal consistencies 0.97–0.99 against human 0.79–0.89, factor-pure loadings) contributes the subtlest failure — caricature fidelity: the replica cleaner than the person, regularities exaggerated (INTROVERTED applied to every answer), consistency mistaken for faithfulness. High coherence can signal over-regularisation, and the chapter connects it to Chapter 22: the label again devouring the mess it claims to model.

The battery: scoped frontier, honest baselines, held-out evaluation

EXP-30, the book’s headline measurement experiment, builds a Personal Reproduction Battery: evidence packages (interview, declared policy, artifacts, bounded interaction evidence) strictly separated from held-out targets across seven families (factual self-report, preferences, value judgements, technical choices, scenario decisions, explanation/writing, multi-step chains — difficulty measured, not assumed). Baselines run population model, minimal facts, Part-III model only, interview, history, combined — plus the critical HUMAN RETEST denominator, asking for each behaviour class whether more evidence closes the gap toward self-repeatability. Leakage discipline is severe: no evaluation responses, paraphrases, or post-decision information in evidence; retest sets (stability) separated from generalisation sets (prediction); evidence timestamped before target actions. Adjudication splits decision/reasons/qualifiers/confidence/style — stylistic similarity never standing in for conclusion agreement. Abstention is first-class: forced versus calibrated prediction compared on coverage, accuracy-when-predicting, overconfidence error, and appropriate silence — a lower-coverage replica that knows its boundary more faithful than a fluent guesser. Stress cases span near/far-domain transfer (confidence must not travel), reversals, novel value conflicts, context dependence, and sequential divergence. The deliverable is a reproduction frontier — target × evidence × horizon × confidence → fidelity — scoped, dated, and fenced. Identity language anywhere in its vicinity fails the chapter on sight.

Suppose the replica eventually becomes extremely good — predicting speech, choices, style, even its own uncertainty. What exactly has been reproduced? Prediction, however exact, never crosses into permission: evidence sufficient to evaluate predictive reproduction is not thereby authorised for unrelated proxy action, personalisation, or disclosure — data purpose travels with the evidence, never assumed from its availability.

References

• Park et al., Generative Agent Simulations of 1,000 People (Stanford HAI brief May 2025; arXiv:2411.10109). Verified in depth: 1,052 participants, 2h interviews, GSS 85%-of-self-retest, Big Five/games/experiments, bias reduction. Used as bounded positive with retest-relative framing.
• Li et al. (2025). BehaviorChain. Findings ACL 2025, pp. 15738–15763. DOI 10.18653/v1/2025.findings-acl.813. Verified: 15,846/1,001, SOTA struggles with continuity.
• Jiang et al. (2025). IndieValueCatalog. Proc. ACL 2025 (long), pp. 6757–6794. DOI 10.18653/v1/2025.acl-long.336. Verified: 55–65% frontier value prediction; demographics insufficient.
• Personality emulation (2024). Sci. Rep., s41598-024-84109-5. Verified: 0.90–0.94 convergent, 0.97–0.99 vs 0.79–0.89 consistency. Used for caricature fidelity.

Proposed experiment EXP-30: Personal Reproduction Battery

Status: PROPOSED. Per the design above (seven target families; A–F evidence baselines + human retest; leakage discipline; split adjudication; abstention comparison; stress battery; scoped frontier output). Identity vocabulary banned from all materials.