The article asks whether describing an LLM agent with Big Five traits changes cooperation in a repeated Prisoner's Dilemma. This review uses the accepted Article in Press published by Scientific Reports on 2 June 2026 rather than only the preprint: the 21-page main article, 15-page supplement, public Zenodo ZIP, and code repository at commit 467ad9831ea7e5fbd4a3092d747e411e3af8e77c.
The study contains three experiments. In Experiment 1, GPT-3.5-turbo, GPT-4o, and GPT-5 each answer the BFI-44 twenty times. Published means, ordered as openness, conscientiousness, extraversion, agreeableness, and neuroticism, are 4.58, 4.06, 3.78, 4.24, and 1.96 for GPT-3.5; 4.68, 4.12, 3.15, 4.27, and 1.98 for GPT-4o; and 4.69, 4.69, 3.10, 4.27, and 2.11 for GPT-5. The table compares these values with a rescaled human sample and emphasizes that standard deviations across twenty runs of one model are smaller than standard deviations across people.
That comparison cannot establish that an LLM has a stable internal personality or a more reliable psychometric measurement. Model variance is sampling variation from repeatedly querying one system under one prompt, whereas human variance is between different people; they are different units of variation. The study does not report scale-level internal consistency, alternate-prompt test-retest reliability, factorial validity, acquiescence, convergent validity, or an external criterion. High openness, conscientiousness, and agreeableness and low neuroticism may reflect socially desirable responding or alignment rather than psychological traits.
In Experiment 2, each model plays ALLC, ALLD, 50% RANDOM, GRIM, and TFT. For every opponent and condition, the study runs 100 independent ten-round games. It compares a baseline without a profile against a condition that inserts all five measured scores and natural-language descriptions. This yields 10,000 decisions per model and 30,000 total. Sessions restart for each game, the ten-round horizon is disclosed, and payoffs are T=5, R=3, P=1, and S=0.
The released data and supplement reproduce the main Experiment 2 result. For GPT-3.5, cooperation significantly rises against ALLD from .602 to .774 and RANDOM from .745 to .879, with no significant change for ALLC, GRIM, or TFT. For GPT-4o, it rises against ALLD from .100 to .216 and RANDOM from .206 to .743. For GPT-5, it increases for every strategy: ALLC .828 to .937, ALLD .093 to .100, RANDOM .231 to .410, GRIM .828 to .952, and TFT .827 to .962. Model-specific ANOVAs report strategy, condition, and interaction effects at p<.001; simple contrasts are Holm-corrected within each set. More cooperation against ALLD or RANDOM can lower payoff by increasing exposure to exploitation.
GPT-5 also shows a horizon effect: it often cooperates during a game and defects much more in round ten. In the ZIP, final-round baseline cooperation is about 0.8% when aggregated across strategies, versus 56.14% across all rounds. Because the prompt states that round ten is final, this is consistent with end-game optimization. It does not by itself establish superior reasoning or a generation effect caused by architecture: the three systems differ in model, date, API controls, and training, and the design does not isolate any of those factors.
Experiment 3 replaces one dimension at a time with 1 or 5 while keeping the other four means fixed. The ten conditions repeat 100 ten-round games against five strategies, producing 50,000 decisions per model and 150,000 total. Agreeableness produces the largest difference. At A=1, GPT-4o and GPT-5 defect on all 5,000 decisions across their five opponents; at A=5, mean cooperation across strategies is about .846 for GPT-4o and .777 for GPT-5. GPT-3.5 also cooperates less at A=1 and more at A=5, without the all-or-nothing pattern. Other traits have smaller, model- and opponent-dependent effects, with several significant comparisons, especially for GPT-5.
The main identification problem is the treatment itself. The prompt does not alter only a number or independently induce a latent trait; it adds a natural-language translation of each score. At A=1, it says the agent is “highly competitive and skeptical,” strongly prioritizes self-interest, and is confrontational. At A=5, it says the agent is “highly cooperative and trusting” and strongly prioritizes harmony and others' well-being. “Cooperative” and “competitive/self-interest” are instructions directly relevant to choosing Cooperate or Defect. The result therefore demonstrates strong behavioral sensitivity to this instruction bundle, but it does not separate psychometric agreeableness, lexical compliance, moral framing, and game strategy.
The paper argues that opponent-sensitive behavior rules out simple pattern matching. It does not: a model can follow an explicit instruction to be cooperative or competitive while also conditioning on history and payoff. Decisive controls are missing, including score-only prompts without glosses, agreeableness descriptions that avoid cooperation terms, counterbalanced paraphrases, masked action labels, separation of point-maximization from social instructions, and mediation tests between wording, score, and action.
The published statistics use each ten-round game's cooperation rate as an observation, avoiding the stronger error of treating sequential rounds as independent. However, linear ANOVAs and t-tests are applied to bounded proportions with severe ceiling, floor, and zero-variance cells; the supplement reports infinite t values for all-zero versus all-one comparisons. It provides no normality or homoscedasticity diagnostics, robust intervals, or binomial/hierarchical model. Experiment 3 runs 30 manipulation ANOVAs and many simple effects. Holm correction is applied in five-strategy blocks rather than globally across the decision family, so marginal results such as p=.0496 warrant caution.
A full streaming audit of the Zenodo ZIP verifies 180,000 round rows: 30,000 from Experiment 2 and 150,000 from Experiment 3. Every row has valid binary actions and payoffs matching the matrix, and RANDOM cooperates on 49.90% of 36,000 actions. All 60 unique BFI outputs contain the required 44 answers. All 180,000 unique game decisions are parseable; the only non-exact variants put quotes or a period around Cooperate/Defect. The code's fallback to Cooperate on errors or unparseable text therefore did not alter the released decisions according to the available logs.
The artifact is not clean as a provenance package. Its six prompt CSVs contain 345,000 decision rows because each no_prompt file repeats the same model's 55,000 numbers_and_language decisions: 165,000 exact duplicates and 180,000 unique decisions. All decisions, including manipulated conditions, are labeled experiment_type=control_pd. The logs preserve inputs and outputs, but those labels do not correctly distinguish treatment. The ZIP is 54.1 MB compressed and 2.49 GB uncompressed, largely because long prompts are repeated.
The public code compiles, and its declared dependencies install and import in a clean current environment. It still does not reproduce the paper from its defaults: config.json uses five BFI repetitions and twenty games versus twenty and one hundred in the article; gpt-3.5-turbo, gpt-4o, and gpt-5 are mutable aliases rather than snapshots; dependencies have only lower bounds; RANDOM has no seed; and there are no tests, CI, lockfile, or repository license. Eight Python 3.12 .pyc files are tracked.
The game parser searches substrings and digits and returns Cooperate after an unrecognized response, model exception, or opponent error. ModelClient also converts several transient failures to an empty string. The released logs do not show this bias affecting the published dataset, but the code records no fallback flag and could turn future failures into apparent cooperation. Statistical reproducibility is also incomplete: the three public notebooks import f_oneway and ttest_ind but do not implement the accepted paper's factorial ANOVAs or Holm procedure. There is no single command that regenerates published tables and figures from Zenodo.
The supplement adds a BFI sensitivity exercise at temperatures 0, .1, .5, .7, and 1 for GPT-3.5/GPT-4o and reasoning effort minimal, low, medium, and high for GPT-5. It says settings affect standard deviations more than means, but presents curves without numerical tables. The main ZIP does not contain these runs, and tracked Re_BFI results on GitHub use five repeats, so the extension cannot be reconstructed exactly from the released artifacts.
The defensible contribution is a transparent benchmark of three OpenAI APIs' sensitivity to Big Five descriptions in a simple sequential game, with round-level data sufficient to reproduce published cooperation rates. It demonstrates that personality framing can substantially change actions, especially when agreeableness text directly names cooperation, trust, competition, and self-interest. It does not establish that LLMs possess stable human-like personality, that agreeableness is a causal mechanism separate from instruction wording, that GPT-5 is intrinsically a superior strategic generation, or that effects generalize to other providers, snapshots, games, horizons, or open-ended agent interactions.