Ethics, Transparency
& Data Governance

Our Ethical Foundation

DivyCHI was founded on a single conviction: that artificial intelligence, to be genuinely useful, must also be genuinely trustworthy. We define trustworthy AI not merely as AI that is accurate, but AI that is honest about its limits, fair in its treatment of people, transparent about how it works, and governed by humans who remain accountable for its outcomes.

Our approach to ethics is not a compliance exercise. It is embedded in how we recruit, how we design our systems, how we train our models, and how we respond when things go wrong. Every team at DivyCHI — engineering, research, product, and operations — operates under a shared ethical charter that is reviewed and updated at least annually.

We acknowledge that intelligence systems of this scale carry significant societal responsibility. We take that responsibility seriously and hold ourselves to a higher standard than the law requires.

Core Principles

Six principles govern every product, feature, and policy decision at DivyCHI:

Model Development Ethics

Our models are developed under a structured ethical review process that spans the entire lifecycle — from data sourcing through deployment and ongoing monitoring.

Pre-Training

• Training data is subject to a provenance review to assess copyright status, consent, and representational balance before inclusion.
• Data sourced from the public web is filtered to remove known harmful content categories, including child sexual abuse material, non-consensual imagery, and incitement to violence.
• We maintain a training data registry — a structured record of all datasets used in each model version — accessible internally and summarised in our transparency reports.

Fine-Tuning & Alignment

• All models undergo alignment training designed to reinforce helpfulness, honesty, and harmlessness before any user-facing deployment.
• Red-teaming sessions — conducted by both internal teams and external researchers — are mandatory before each major model release.
• Human feedback used in alignment training is collected from contractors operating under fair-pay agreements and clear task guidelines. We do not use deceptive or coercive feedback collection methods.

Deployment Gates

• Every model must pass a multi-dimensional safety evaluation covering factual accuracy, refusal behaviour, bias metrics, and adversarial robustness before release.
• Models that fail any safety gate are returned to the research team. Release timelines are never allowed to override safety outcomes.
• Post-deployment monitoring begins immediately on release. Anomalous behaviour triggers an automatic escalation to the Safety Review Board.

Transparency Commitments

We commit to being transparent with our users, regulators, and the public about how DivyCHI's systems work, what they can and cannot do, and where we fall short.

Data Governance Framework

DivyCHI's data governance framework defines how data is classified, accessed, stored, retained, and disposed of across the organisation. It applies to all data handled by DivyCHI — user data, training data, operational data, and internal data.

Data Minimisation

We collect only the data necessary for the stated purpose. We do not collect data speculatively or build behavioural profiles for advertising purposes. Data fields that are no longer needed are deleted on a rolling basis as part of our quarterly data hygiene process.

Cross-Border Data Controls

Data transfers across jurisdictions are governed by our International Transfer Policy. We do not transfer personal data to countries without adequate data protection unless Standard Contractual Clauses or equivalent safeguards are in place. See our Privacy Policy for full details.

Training Data & Consent

The data used to train DivyCHI models is subject to strict governance requirements. We believe that how a model is trained is as ethically significant as what the model does.

User Conversation Data

• Consumer plans (Standard, Pro, Max, X-Max): conversations may be used to improve model quality on an anonymised, aggregated basis. Users may opt out at any time via Settings → Privacy → Training Preferences.
• Enterprise plans: conversation data is excluded from model training by default. Opting in requires an explicit, written instruction from the account administrator.
• API users: data submitted via the API is not used for training by default. Opt-in is available for organisations that wish to contribute to model improvement under a formal data contribution agreement.

Third-Party & Web Data

• We respect robots.txt directives and do not scrape data from sources that explicitly prohibit it.
• Licensed datasets are used under written agreements that include representations about the data's provenance and consent status.
• We maintain a public summary of the categories of data sources used in each major model version, updated with each release.

Bias & Fairness

We recognise that AI systems can perpetuate, amplify, or introduce bias — and that the populations most harmed by biased systems are often those with the least power to challenge them. Addressing bias is therefore not optional; it is a core design requirement.

What We Measure

• Demographic parity: do model outputs differ in quality or tone across gender, race, ethnicity, religion, nationality, or disability status?
• Language equity: do non-English-language users receive comparable quality outputs to English-language users for equivalent tasks?
• Refusal symmetry: do refusal rates for sensitive queries differ across demographic groups in ways that cannot be justified by safety considerations?
• Representation in outputs: do default-generated content and examples reflect a diverse range of human backgrounds and experiences?

How We Act

• Bias evaluation results are reviewed by a cross-functional team including external advisors before each major model release.
• Models that fail fairness thresholds on any measured dimension are not released until the bias is remediated or mitigated to acceptable levels.
• We publish aggregated bias evaluation scores in our model cards and transparency reports.
• Users who identify biased behaviour in our products are encouraged to report it via ethics@divychi.com. All reports are reviewed within 10 business days.

Human Oversight

DivyCHI operates on the principle that consequential decisions should not be made by AI systems alone. We design our products to support, not supplant, human judgement.

Internal Oversight Structures

• Safety Review Board: a standing committee of senior engineers, ethicists, and external advisors that reviews all new model capabilities, high-risk deployments, and material incidents. Meets monthly and on-demand for urgent matters.
• Ethics Council: an independent advisory body comprising external academics, civil society representatives, and domain experts. Reviews DivyCHI's ethical framework annually and publishes its own assessment.
• AI Product Owner programme: every deployed AI system has a named internal product owner who is personally accountable for its behaviour and must respond to escalations within 24 hours.

Design Constraints

• DivyCHI products operating in high-stakes domains (medical, legal, financial, safety-critical infrastructure) include mandatory human review checkpoints and must not be configured to act autonomously.
• Agentic features — where DivyCHI takes actions on behalf of a user — require explicit user confirmation for each consequential action and can be interrupted or reversed at any time.
• We do not build features designed to circumvent human oversight, even at a user's request.

Prohibited Use Enforcement

Our Terms & Conditions define categories of prohibited use. This section explains how we detect and respond to violations.

Detection Methods

• Automated classifiers: real-time content classifiers flag queries and outputs that match known patterns of prohibited use (e.g., requests related to CBRN weapons, CSAM, or targeted violence).
• Human review queues: flagged interactions are reviewed by trained Trust & Safety analysts within 24–48 hours. Analysts follow a structured decision framework and are subject to inter-rater reliability testing.
• User reports: users can flag any interaction via the in-product report button. Reports are reviewed within 48 hours and contribute to classifier improvements.

Enforcement Actions

Third-Party Audits

We believe that self-assessment alone is insufficient for systems of this consequence. DivyCHI commissions independent third-party evaluations of our models, infrastructure, and governance practices on a regular schedule.

• Safety audits: conducted annually by an independent AI safety organisation. Results are shared with our Safety Review Board and summarised in our transparency report.
• Security penetration testing: conducted bi-annually by a certified external firm. Critical findings are remediated within 30 days of report delivery.
• Bias & fairness review: conducted with each major model release by an external team with expertise in algorithmic fairness and demographic representation.
• Data governance audit: conducted annually to verify compliance with our data classification, retention, and access control policies.
• SOC 2 Type II: infrastructure and security controls audit conducted by a licensed CPA firm. Available to Enterprise customers under NDA upon request.

Incident Response

When something goes wrong — whether a security breach, a model safety failure, or a data governance violation — we follow a structured incident response process designed to contain harm, notify affected parties, and prevent recurrence.

Material incidents — those affecting a significant number of users or involving sensitive data — are summarised in our annual transparency report. We do not suppress incident disclosures to protect our reputation.

Research & Responsible Disclosure

DivyCHI is an AI research company, and we believe that open scientific exchange makes AI safer for everyone. We participate in the broader research community and support responsible disclosure practices.

Our Research Commitments

• We publish safety and alignment research findings that do not create undue risk if disclosed — including research that reveals limitations or failures of our own systems.
• We support academic researchers who wish to study our models through a structured researcher access programme. Applications are reviewed by our research team at research@divychi.com.
• We do not suppress publication of research that reflects negatively on DivyCHI systems, provided the research meets basic standards of scientific rigour.

Security Vulnerability Disclosure

If you discover a security vulnerability in DivyCHI's systems or models, please report it to security@divychi.com using our PGP key (available at divychi.com/security). We operate a responsible disclosure programme:

• We acknowledge all vulnerability reports within 48 hours.
• We commit to remediation timelines based on severity: critical within 7 days, high within 30 days, medium within 90 days.
• We will not pursue legal action against good-faith security researchers who follow our disclosure guidelines.
• Researchers who discover and responsibly disclose qualifying vulnerabilities are acknowledged in our security hall of fame, published annually.

Your Rights Under This Policy

In addition to the rights granted under our Privacy Policy, this ethics and governance framework gives you the following specific rights:

• Right to know if AI is involved: you can ask DivyCHI at any time whether a product, feature, or communication you are interacting with involves AI-generated content, and we will answer honestly.
• Right to opt out of training use: you may opt out of having your conversation data used for model training at any time, with immediate effect, via account settings.
• Right to report bias or harm: you may report any AI output you believe is biased, harmful, or ethically concerning to ethics@divychi.com and receive a substantive response.
• Right to appeal enforcement actions: if your account is restricted or terminated for an alleged ethics or policy violation, you may appeal the decision within 14 days of notification.
• Right to be notified of material policy changes: changes to this document that materially affect your rights will be communicated to you at least 30 days before they take effect.

Contact & Reporting

We welcome questions, concerns, reports, and criticism relating to our ethics, transparency, and data governance practices. The right contact depends on your concern: