Trust & Responsible AI

Last updated: May 14, 2026

Market Twin builds AI-powered consumer simulation tools used by companies making real business decisions. We take seriously our responsibility to deliver outputs that are accurate, fair, and transparent. This page documents our approach.

AI Governance

AI Governance Frameworks We Follow

Market Twin's AI practices are designed in alignment with leading international AI governance standards.

NIST AI Risk Management Framework (US)Govern · Map · Measure · Manage four core functions
EU AI Act ComplianceSelf-assessed as "Limited Risk" (transparency obligations). Preparing for August 2026 enforcement.
OECD AI PrinciplesInclusive growth, human-centered values, transparency, robustness, accountability
ISO/IEC 42001 AI Management SystemsRoadmap: 2027
Korean AI Ethics GuidelinesMinistry of Science and ICT

Safeguards

Our Safeguards

Five-stage quality assurance system every simulation output passes through.

Statistical Grounding

All AI personas are grounded in official government statistics from 24 countries (KOSIS, US BLS, Japan e-Stat, and more). This prevents fabricated demographic distributions and ensures personas reflect actual market realities.

3-Layer Voice Sanitizer

Every LLM output passes through three independent quality checks:

Locale filter — prevents cross-cultural voice leakage (e.g., Japanese persona using Korean expressions)
Brand-name bias detection — flags spontaneous brand mentions for review against a 28-brand reference dictionary
Auto-translation pattern detection — identifies machine-translated tone and triggers persona regeneration

Multi-LLM Ensemble

For Deep-tier simulations, we run independent inference across three model families (Anthropic Claude, OpenAI GPT, Google Gemini) in round-robin. This reduces single-vendor bias and improves output consistency.

Slot-Based Diversity Enforcement

Occupation and demographic distributions are pre-defined in slot pools across 8 categories. This prevents LLM-induced clustering and ensures statistical diversity in synthetic respondent samples.

Workspace Data Isolation

Row-Level Security (RLS) on all customer data
Workspace-level persona pools — no cross-customer data leakage
30-day read-only sharing tokens with automatic expiration
Customer data is never used to train models

Limitations

What Our AI Cannot Do

We are transparent about our system's limitations.

Not a replacement for human decision-making

Our outputs inform, not decide. All material business decisions should include human judgment.

Not actual consumer research

Synthetic personas approximate real consumer reactions but cannot replace lived human experience.

Statistical, not predictive

We provide probability-weighted insights based on statistical grounding, not deterministic forecasts.

Cultural nuance constraints

While we ground personas in official statistics, deep cultural and contextual nuances may require local expert review.

Data Privacy & Security

Data Handling

No customer data used for model trainingCustomer inputs are not shared with LLM providers for training purposes.

EncryptionAll data encrypted in transit (TLS 1.3) and at rest (AES-256).

Regional data residencyKorean customer data stored in Korean cloud regions.

Right to deletionCustomers can request complete data deletion at any time.

Third-Party Processors

We use the following processors under data processing agreements:

AnthropicLLM inference

OpenAILLM inference

Google CloudGemini + infrastructure

SupabaseDatabase

VercelApplication hosting

Certifications

Certifications & Memberships

Current Status

Korean Personal Information Protection Self-Certification (K-PIPL)— In preparation
Born2Global Centre Member— Applied (2026)

Planned (Roadmap)

TTA AI Trustworthiness Certification (Korea)— Target: 2026 Q4
ISMS-P (Korea)— Target: 2027 Q2
ISO/IEC 27001— Target: 2027 Q3
SOC 2 Type II (US)— Target: 2027 Q4

Reporting

Reporting Concerns

If you encounter outputs that appear biased, factually incorrect, or otherwise concerning, we want to know.

Contact: trust@markettwin.ai

We commit to:

Acknowledging reports within 2 business days
Investigating and responding within 14 days
Publishing material findings (anonymized) in our quarterly transparency reports (starting 2026 Q4)

Transparency Reports

We commit to publishing quarterly transparency reports starting Q4 2026, covering:

System reliability metrics (consistency, slip-pattern detection rates)
Geographic and demographic distribution of simulated personas
Customer-reported concerns and resolutions
Model version updates