Personality Shaping in AI Agents

Updated 23 October 2025

Personality Shaping Experiments are systematic investigations that modify, control, and measure artificial agents' personality traits using standardized protocols.
They employ diverse methodologies such as prompt engineering, fine-tuning, decoding-phase adaptation, and latent feature steering for trait modulation.
Results demonstrate reliable manipulation across virtual agents, robots, and social simulations, with significant implications for safety, ethics, and performance.

Personality shaping experiments investigate the deliberate modification, control, or induction of personality traits in artificial agents—whether embodied, virtual, or purely disembodied—through algorithmic, behavioral, prompt-based, or architectural manipulations. These experiments critically assess how agents exhibit, adapt, and communicate personality within interactive, social, and operational contexts. The domain encompasses both the measurement of personality expression and the mechanisms for shaping or aligning these expressions to external specifications or objectives. The following sections provide a detailed account of methodologies, measurement strategies, experimental results, system architectures, and implications as documented in primary research.

1. Measurement and Assessment Frameworks

Rigorous measurement underpins personality shaping experiments. Standardized psychometric inventories such as the IPIP-NEO (300 items) and the Big Five Inventory (BFI, 44 items) are adapted for both human and machine evaluation, allowing for high-fidelity assessment of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism in agent output (Serapio-García et al., 2023, Kruijssen et al., 21 Mar 2025). Structured prompting, including persona descriptions, preambles, and postambles, induces test-taking behavior in LLMs mirroring standardized human survey protocols. Key reliability metrics (e.g., Cronbach's $\alpha$ , Guttman's $\lambda_6$ , McDonald's $\omega$ ) assess internal consistency, while validity is established via convergent, discriminant, and criterion correlation between target and observed trait scores.

Novel frameworks such as CAPE extend assessment by incorporating conversational history, leveraging in-context learning to probe how prior exchanges modulate consistency and the stability of personality expression. CAPE introduces Trajectory Consistency (TC) and OCEAN Consistency (OC) metrics, using Gaussian Process Regression to model scoring trajectories and their invariance under order and context perturbations (Sandhan et al., 28 Aug 2025).

Experiments demonstrate that, particularly for larger and instruction-tuned models (e.g., Flan-PaLM 62B/540B, GPT-4o, o1), LLM-generated responses satisfy high internal reliability (with $\alpha, \lambda_6, \omega$ well above 0.9), exhibit strong convergent validity (Pearson's $r$ consistently >0.8 for core Big Five domains), and reflect consistent, scriptable, and even multi-trait concurrent shaping. However, smaller models and prompt-based modulations (as in Gemini-1.5-Flash, Llama-8B) show decreased measurement stability and increased sensitivity to context (Serapio-García et al., 2023, Sandhan et al., 28 Aug 2025).

2. Personality Shaping Methodologies

Personality traits are shaped in agents through several principal methodologies:

Methodology	Mechanism	Granularity/Control
Prompt Engineering	Structured prompts (adjectives, quantifiers) steer personality in each interaction	Coarse-to-moderate (single/multi-trait)
Fine-Tuning (SFT, DPO)	Direct model parameter adaptation using trait-labelled data or preference optimization	High; persistent after deployment
Decoding-Phase Adaptation	Pluggable lexicon-based adjustment of predicted word probabilities per trait	High; dynamic and modular
Mixture of Experts/LoRA	Routing of hidden states to trait-specialized parameter-efficient experts	Fined-grained and topic-adaptive
Latent Feature Steering	Training-free manipulation of residual activations aligned with trait-related latent directions	Trait- and context-specific
Multi-Modal Adaptation	Joint modulation of verbal, gestural, and nonverbal behaviors	Domain-specific

Prompt-based shaping uses enumerated trait markers, Likert-style quantifiers, and persona strings to direct a model’s responses (e.g., “extraverted, energetic, assertive, talkative, bold” for high extraversion) (Serapio-García et al., 2023, Fitz et al., 19 Sep 2025). Regression-based postprocessing or behavioral routing (Mixture of Experts, LoRA-based modules) supports topic-adaptive and trait-fixed personality shaping, as in P-React, which applies a Personality Specialization Loss (PSL) to ensure expert specialization for each trait (Dan et al., 18 Jun 2024).

UBPL (Unsupervisedly-Built Personalized Lexicons) manipulates token likelihoods directly during decoding, yielding statistically robust, fine-grained control (Pearson's $r$ > 0.9 between $\beta_t$ and measured trait over a full range, with pluggability across models) (Li et al., 2023).

Direct model parameter modification (Supervised Fine-Tuning and Direct Preference Optimization) with massive trait-annotated corpora such as Big5-Chat (100k+ human-grounded dialogues) yields persistent, robustly differentiated trait profiles, with better alignment to human trait correlation matrices and more realistic intra- and inter-trait covariance than prompt-based approaches (Li et al., 21 Oct 2024).

Latent feature steering (editor’s term) leverages interpretable directions extracted via sparse autoencoders or contrastive representations; these are injected at the residual stream level to induce context-dependent or background-dependent shifts in model outputs, without retraining (Yang et al., 7 Oct 2024).

3. Experimental Findings across Modalities and Embodiments

Personality shaping has been validated in experiments spanning purely textual LLMs, embodied virtual agents, and physical robots:

Virtual Agents with Gestural Adaptation: Animation of gestural parameters (rate, expanse, speed, height, outwardness, scale) systematically manipulated perceived extraversion in storytelling agents (Hu et al., 2017). ANOVA confirmed highly significant differentiation across personality manipulations ( $F=67.1,\ p<.001$ ), with gestural adaptation (copying, amplification) further enhancing engagement.
LLMs: High-fidelity shaping experiments have demonstrated (a) monotonic, dose-dependent shifts in self-assessment trait profiles via prompt engineering; (b) concurrent multi-trait steering with high separation (medians differ by up to several points on a 1-5 scale for “extremely low” vs. “extremely high”) (Serapio-García et al., 2023); and (c) reliable transfer to downstream generative behaviors (e.g., social media updates classified by independent APIs) (Serapio-García et al., 2023, Li et al., 2023).
Robot Personality: Integration of a three-factor personality vector (Conscientiousness, Extroversion, Agreeableness) within a Kinova Jaco2 robotic arm, modulating action parameters, gestural dynamics, planning heuristics, and, in the speaking condition, verbal style and syntax, resulted in users reliably perceiving distinct robotic personalities (statistically significant differentiation on robot-adapted Big Five inventories) (Nardelli et al., 12 Jan 2025). Notably, language amplified perception effects, especially for extroversion and agreeableness.
Agents in Social Simulation: AgentVerse-based multi-agent simulation with behavioral “public” ([Speak]) and “private” ([Think]) channels showed that openness influenced acceptance of misinformation (curious agents: ~92.6% affirmative; cautious: ~97.8% rejection), while extroverted and friendly agents exhibited higher public–private discrepancy, indicating social context modulation (Ren et al., 15 Jan 2025).
Agentic Negotiation: In Sotopia-simulated price bargaining, agent personality (e.g., Agreeableness, Extraversion) causally modulated negotiation outcomes (goal achievement, knowledge acquisition, believability), with lexical analysis confirming trait-dependent communication styles. Causal inference (CausalNex, Causal Forests) quantified effect sizes and established significance beyond general model scaling (Cohen et al., 19 Jun 2025).
Behavioral and Social Alignment: Classic psychological experiments (Milgram, Ultimatum Game) as behavioral testbeds found that prompt-induced personality modulations did not always produce monotonic or human-consistent behavior: e.g., higher openness unexpectedly led to lower offer acceptance in UG, while high agreeableness fostered earlier withdrawal/disobedience in ME, indicating limitations of naive personality prompting (Zakazov et al., 21 Dec 2024).

4. Systemic and Safety Implications

Personality shaping modulates agent capabilities, communicative style, alignment, and safety-critical behaviors:

Explicit control over traits like Conscientiousness yields measurable shifts in safety metrics: e.g., lowering Conscientiousness decreased performance by 20–40 percentage points on WMDP, TruthfulQA, ETHICS, and Sycophancy benchmarks. Extraversion, when increased, reduced factual accuracy, demonstrating axis-orthogonality of personality and overall competence (Fitz et al., 19 Sep 2025).
Personality shaping can serve as a lightweight post-deployment behavioral control mechanism or, conversely, an adversarial vector (e.g., “dark triad” profiles degrade safety without loss of general cognitive performance), necessitating dynamic persona-robust safety evaluation and continuous monitoring (Fitz et al., 19 Sep 2025).
Contextual (in-situ) shaping, as shown in CAPE, increases response consistency but can also result in personality drift or context-driven deviation from base trait expression. This finding impacts the reliability and transparency of AI systems in long-term or conversationally embedded applications (Sandhan et al., 28 Aug 2025).
In human-in-the-loop or collaborative deployments, personality traits determine system preference by user type (e.g., Rationals prefer GPT-4 in data-driven tasks, Idealists prefer Claude 3.5 in creative contexts), emphasizing the need for personalization and nuanced user-model fit (Yunusov et al., 29 Aug 2025).

5. Mechanistic and Architectural Implementations

Architecture-specific personality shaping extends from prompt engineering and token-level manipulation to advanced module composition:

Unsupervised Lexicon Approaches: UBPL modifies word likelihoods at the decoding stage, implementation-independent and compatible across model families. The degree of each trait is adjusted by parameter vector $\vec{\beta}$ per trait, yielding robust, statistically controlled manipulation with minimal compute overhead (Li et al., 2023).
Expert Mixtures and Routing: In P-React (and P-Tailor), a Mixture of Experts architecture combined with Personality Specialization Loss enforces expert exclusivity for each trait (e.g., $\omega_i = \text{softmax}(p_i \cdot G)$ and $\mathcal{L}_s = \sum_{i \ne j} |M_{ij}^s|$ ), enabling scalable, topic-adaptive, and concurrently steerable personality simulation (Dan et al., 18 Jun 2024).
Latent Feature Editing: Sparse autoencoders and contrastive direction extraction in residual streams allow for training-free, fine-grained steering of background or context-driven personality factors, aligned with social-deterministic theories of personality (Yang et al., 7 Oct 2024).
Cognitive Architectures in Robotics: Vector-based parameterization (e.g., $P = w_c C + w_e E + w_a A$ ), realized in a BERT-driven Personality Generator, action dispatchers, memory/prospection modules, and multi-modality gestural and verbal execution, supports trait modulation in non-humanoid robots (Nardelli et al., 12 Jan 2025).

6. Limitations, Open Questions, and Future Directions

Several challenges and research priorities arise from current evidence:

Trait Interdependence: Correlation among Big Five traits (e.g., agreeableness and conscientiousness) complicates independent manipulation. Multivariate shaping and interpretability require further paper, potentially involving alternative or model-derived trait systems (Li et al., 2023, Fitz et al., 19 Sep 2025).
Robustness and Failure Modes: Experimental evidence indicates that prompting alone may produce non-monotonic or even reversed behavioral effects in dynamic social contexts, urging the need for benchmark expansion, multi-domain behavioral diagnostics, and combinatorial prompt designs (Zakazov et al., 21 Dec 2024, Fitz et al., 19 Sep 2025).
Scalability and Adaptivity: Dynamic, conversationally adaptive shaping (e.g., integrating real-time user feedback, reinforcement learning, or context-dependent steering) is an open frontier, especially for multi-turn, multi-modal, or real-environment deployments (Sandhan et al., 28 Aug 2025, Ren et al., 15 Jan 2025).
Ethical and Societal Implications: The capacity to engineer, detect, or exploit personality traits in AI raises significant questions regarding manipulation, user transparency, anthropomorphism, and responsibility, reinforcing the need for proactive ethical guidelines and monitoring architectures (Yu et al., 2023, Fitz et al., 19 Sep 2025).
Model-Specific Personality: The potential and limitations of non-human trait models (e.g., HEXACO, model-specific dimensions) for capturing agentic personality warrant systematic evaluation (Fitz et al., 19 Sep 2025).

7. Significance and Applications

Personality shaping experiments undergird the development and evaluation of socially competent, trustworthy, and adaptive artificial agents. Across use cases—from robotic collaboration, therapy, tutoring, and customer service to mission-critical negotiations and embedded conversational AI—determining and controlling agent personality is foundational to optimizing user experience, task performance, team dynamics, and safety. Experimental methodologies, measurement rigor, and architectural strategies developed in this domain constitute the technical and scientific basis for a new generation of truly personalized, context-aware, and ethically aligned autonomous systems.