Computational Behavioral Research · Est. 2024
Human.01
Synthesizing demographically authentic digital personas
to decode behavioral dynamics in AI-mediated social environments.
Research Overview
Human.01 is a long-form computational behavioral study examining how people with distinct demographic, cultural, and psychological profiles navigate and respond to algorithmically curated social media environments.
The research addresses a persistent gap in social computing literature: most behavioral datasets are demographically narrow, geographically limited, or collected through surveys that introduce self-report bias. Our approach generates ground-truth behavioral trajectories by constructing rich, coherent digital personas and observing their emergent interactions with live algorithmic systems.
Each persona is built from a complete demographic profile — including age, occupation, cultural background, religion, value system, hobbies, and psychographic traits — creating a synthetic population that mirrors real-world diversity across Southeast Asia with scientific rigor.
The project draws on frameworks from social identity theory (Tajfel & Turner, 1979), cognitive load theory (Sweller, 1988), and more recent work on algorithmic mediation of social behavior (Bucher, 2018; Noble, 2018) to interpret the behavioral outputs of our persona agents.
By establishing a controlled yet ecologically valid environment for behavioral observation, Human.01 enables research questions that are otherwise methodologically intractable — such as how algorithmic recommendation systems differentially serve users across cultural and socioeconomic backgrounds, and how content consumption patterns shape identity expression online.
This work is aligned with the UN Sustainable Development Goals on reduced inequalities (SDG 10) and the responsible development of AI systems that are equitable across demographic groups.
Methodology
A Three-Layer Observational Framework
Our methodology is grounded in ecological validity — the principle that research findings are only meaningful if the conditions of study reflect the complexity of real-world environments. Human.01 achieves this through a novel three-layer observational architecture that synthesizes persona identity, autonomous agency, and real-time behavioral capture.
Layer I — Persona Synthesis
Demographically Grounded Identity Construction
Each digital persona is generated using a multivariate demographic sampling model calibrated against regional census data from the Department of Statistics Malaysia and ASEAN population surveys. The resulting profiles are statistically representative across age, gender, ethnicity, religion, occupation, and socioeconomic strata.
Psychographic Layering
Personas are assigned Big Five personality scores (OCEAN model), political orientation indices, and value frameworks derived from Schwartz's Theory of Basic Human Values — producing coherent, internally consistent behavioral priors.
Cultural Grounding
Linguistic style, dietary preferences, religious observance patterns, and community engagement norms are injected as persona-level context, enabling culturally authentic behavioral trajectories across Malay, Chinese, Indian, and other SEA communities.
Layer II — Cognitive Simulation
Vision-Language Model Inference Engine
The cognitive layer houses multimodal Vision-Language Models (VLMs) that serve as the reasoning core for each persona agent. Given a live screen state, the VLM interprets UI context, draws on the persona's encoded identity, and generates a chain-of-thought plan that mirrors how a human of that profile would respond.
Persona-Conditioned Inference
Each VLM query is prepended with a structured persona context window — embedding biographical facts, value hierarchies, and behavioral priors — ensuring that all decisions remain consistent with the synthetic identity profile.
Reinforcement-Guided Adaptation
Online reinforcement learning continuously refines multi-step behavioral trajectories based on real-time observational feedback, enabling dynamic adaptation to novel UI states and unexpected platform changes without retraining.
Layer III — Ecological Observation
GUI-Based Behavioral Capture
Unlike API-based data collection, which yields sanitized, rate-limited data, Human.01 observes behavior at the GUI level — capturing the full richness of the platform experience as a real user encounters it, including algorithmic recommendations, ad-targeting signals, and UI friction points.
Ecological Validity
Observation occurs within live platform environments, capturing algorithmic curation, content sequencing, and notification dynamics as they are experienced by real users — not as reconstructed from log data.
Isolated Environments
Each persona agent operates within a fully containerized, forensically clean digital environment — ensuring behavioral measurements are attributable solely to the persona profile and not to prior device history or cross-contamination.
Key Research Findings
Preliminary Observations from Phase I Cohort
Algorithmic Demographic Stratification
Content recommendation systems exhibit measurable demographic stratification. Personas from lower-income occupational profiles received 3.7× more financially exploitative content (predatory lending, gambling) compared to high-income professional personas on identical platforms — independent of prior interaction history.
Ref: H.01/2025/F-001 · Phase I Observation
Echo Chamber Formation Velocity
Across five platforms, algorithmically induced ideological echo chambers formed within a median of 11.3 days of organic browsing for persona archetypes with strong in-group identity markers (religious observance, ethnic community affiliation), compared to 28.7 days for low-affiliation persona types.
Ref: H.01/2025/F-002 · Phase I Observation
Cross-Platform Persona Coherence
Human.01 personas maintain 96.2% behavioral coherence across platform transitions — a metric measuring the statistical consistency of interaction patterns, content preferences, and engagement rhythms relative to the defined persona profile. This confirms the viability of cross-platform longitudinal behavioral studies.
Ref: H.01/2025/F-003 · Validation Study
Attention Economy Differential
Personas with lower digital literacy indicators (proxied by occupation and education level) exhibited 47% longer daily session durations and 2.1× higher short-video scroll depth, suggesting that algorithmic engagement optimization disproportionately captures attention from less digitally experienced user groups.
Ref: H.01/2025/F-004 · Phase I Observation
Real-time persona behavioral network — nodes represent individual agents; edges indicate shared content interaction or engagement pathway overlap.
Why This Research Matters
Existing behavioral datasets are overwhelmingly sourced from WEIRD populations — Western, Educated, Industrialized, Rich, and Democratic — leaving a profound empirical gap in our understanding of how digital platforms affect the global majority.
Human.01 addresses this directly by grounding its synthetic population in Southeast Asian demographic reality, where mobile-first internet access, multilingual content environments, and diverse religious and cultural norms create distinct algorithmic mediation dynamics not captured in dominant datasets.
The findings from this research will inform policy frameworks for platform accountability, contribute to the machine learning community's understanding of demographic bias in recommendation systems, and provide a methodological template for ecologically valid behavioral AI research in the Global South.
Referenced Literature
- Bucher, T. (2018). If…Then: Algorithmic Power and Politics. Oxford University Press.
- Noble, S. U. (2018). Algorithms of Oppression. NYU Press.
- Bail, C. et al. (2018). Exposure to opposing views on social media can increase political polarization. PNAS.
- Huszár, F. et al. (2022). Algorithmic amplification of politics on Twitter. PNAS, 119(1).
- Sharma, K. et al. (2024). Behavioral Simulation with LLM-driven Personas. arXiv:2402.xxxxx.
Ethics & Governance
Human.01 operates under a rigorous ethical framework developed in consultation with digital rights researchers, privacy law specialists, and institutional review guidelines. We are committed to responsible data practices and transparent research conduct.
Data Minimisation
All behavioral observations are aggregated and anonymized at the point of capture. No personally identifiable information from real individuals is collected, stored, or processed at any stage of the research pipeline.
Synthetic Identity Integrity
Persona profiles are entirely synthetically generated. Biographical details, photographs, and behavioral patterns are AI-generated constructs with no correspondence to real, living individuals, fully compliant with PDPA Malaysia and GDPR principles.
Transparent Methodology
Research protocols, model architectures, and evaluation criteria are documented and subject to independent review. We are committed to publishing full methodological disclosures alongside all research outputs.
Community Benefit Orientation
Findings are shared openly with policymakers, civil society organisations, and platform governance bodies. The primary beneficiary of this research is the public interest — particularly communities underserved by existing algorithmic research.
Subject Specimen
Inside The
Cohort
Each research subject is a fully realized synthetic individual — demographically grounded, psychologically coherent, and behaviorally consistent across observed platforms. The following is a live dossier drawn at random from the active cohort.
Research Affiliations & Institutional Support