Responsible AI Policy

1. Our Commitment

At Inferent, responsible AI is not a compliance checkbox—it is a foundational design principle. We believe that AI systems must be developed with care, humility, and a genuine commitment to the wellbeing of those who interact with them and those affected by their outputs.

This policy describes how we approach the development, deployment, and ongoing governance of the Leuks family of Foundation Models. We recognize that AI is a powerful and evolving technology, and we commit to continuous improvement as our understanding grows.

2. Core Principles

Beneficence

Our models are designed to be genuinely helpful. We measure success not only by capability benchmarks but by the real-world value our models provide to users, developers, and society.

Non-maleficence

We work systematically to identify and reduce potential harms arising from our models, including harmful content generation, enabling of dangerous activities, facilitation of disinformation, and reinforcement of harmful biases.

Autonomy

We respect user autonomy and the right of individuals to make informed decisions. Our models are designed to inform and assist—not to manipulate, deceive, or unduly influence users' beliefs or behaviors.

Justice and Fairness

We strive to ensure our models perform equitably across demographic groups, languages, and cultural contexts. We audit our models for disparate impacts and work to address identified gaps.

Transparency

We publish model cards, usage policies, and transparency reports so users, developers, and researchers can make informed decisions about how and whether to use our models. We are honest about capabilities and limitations.

3. Safety Practices

Safety is integrated throughout our model development lifecycle, not added as an afterthought:

• Pre-training filters: We apply content filtering to training data to reduce exposure to harmful material
• RLHF: Human feedback is used to align model behavior with human values and safety requirements
• Safety fine-tuning: Dedicated optimization passes focus on refusing harmful requests and providing safe responses
• Adversarial testing: We conduct structured red teaming exercises before and after deployment
• Runtime guardrails: Real-time content moderation complements model-level safety

4. Human Oversight

We maintain meaningful human oversight at every stage of our AI systems. This includes human review of AI safety decisions, an escalation path for edge cases flagged by our moderation systems, and regular audits of model behavior by our safety team.

We do not deploy AI systems in contexts where errors could cause irreversible, catastrophic harm without appropriate human review mechanisms in place.

5. Bias Mitigation

We acknowledge that AI models can reflect and amplify biases present in training data and human feedback. We take an active approach to bias mitigation:

• Regular evaluation of model outputs across demographic dimensions
• Diverse teams involved in model evaluation and red teaming
• Targeted interventions when systematic biases are identified
• Transparent reporting of known limitations in model cards

6. Red Teaming

Before deploying major model updates, we conduct structured red teaming exercises in which internal and external researchers attempt to elicit harmful, biased, or otherwise problematic outputs. Findings from red teaming directly inform safety improvements.

We also maintain an ongoing vulnerability disclosure program for model safety issues. See our AI Safety Policy for details.

7. Incident Response

When we identify or are notified of a safety incident—such as the model generating harmful content at scale or being used for harmful purposes—we follow a documented incident response process that includes:

1. Immediate assessment and containment
2. Root cause analysis
3. Remediation (model updates, policy changes, or access restrictions)
4. Post-incident review and transparency reporting

8. Contact

To report an AI safety concern or responsible AI inquiry:
Email: [email protected]