Responsible AI Policy

Last updated: July 27, 2026

1. Our Commitment

Fabric AI is committed to developing, releasing, and maintaining artificial intelligence systems that are safe, transparent, equitable, and broadly beneficial. We believe that advanced intelligence should be efficient enough to run anywhere and open enough for anyone to build on. This Responsible AI Policy outlines the principles and practices that guide our work.

2. Core Principles

2.1 Safety

We conduct safety evaluations on all released models before public distribution. These evaluations include benchmarking for harmful content generation, bias and fairness, factuality, and robustness to adversarial inputs. Our post-training pipeline includes safety-focused data curation and alignment techniques designed to reduce the likelihood of harmful outputs.

We publish system cards and evaluation results alongside major model releases so that the community can make informed decisions about deployment. We also maintain a responsible disclosure process for safety vulnerabilities.

2.2 Transparency

We believe that transparency is essential for trust in AI systems. For every major model release, we publish:

• Technical reports detailing architecture, training data, and methodology
• System cards describing known capabilities, limitations, and evaluation results
• Model cards with intended use cases and recommended deployment configurations
• Benchmark results under standardised evaluation frameworks

We are committed to providing the community with the information needed to understand what our models can and cannot do, where they may fail, and how to use them responsibly.

2.3 Open Access

All Fabric AI models are released under permissive open-source licenses, currently the Apache 2.0 License. We believe that open access accelerates safety research, enables a wider range of beneficial applications, and allows the global community to scrutinise and improve our work. We are committed to maintaining open access as a core principle of our mission.

2.4 Fairness and Inclusivity

We recognise that AI systems can reflect and amplify biases present in their training data. We evaluate our models for demographic bias across relevant dimensions and publish the results. While no model can be perfectly fair across all contexts, we are committed to transparency about known biases and to ongoing research into bias mitigation techniques.

Our training data is primarily English-language web text, which means our models may perform less well on non-English languages and may reflect cultural perspectives predominant in English-language content. We acknowledge this limitation and encourage community contributions to improve multilingual and multicultural representation.

2.5 Accountability

We take responsibility for the models we release. If a significant safety issue is identified in one of our models, we will:

• Acknowledge the issue promptly and transparently
• Provide guidance to users on mitigating the issue
• Work to address the issue in subsequent releases
• Update the relevant system card or documentation

3. Safety Practices

3.1 Pre-Release Evaluation

Before releasing a model, we conduct evaluations across multiple safety dimensions:

• Harmful content: We test for generation of hate speech, violence, harassment, and other harmful content categories
• Bias: We evaluate for demographic biases using standard benchmark datasets
• Factuality: We assess the model's tendency to generate factually accurate versus hallucinated information
• Robustness: We test the model's behaviour under adversarial or out-of-distribution inputs
• Privacy: We evaluate the model's propensity to memorise and reproduce training data

3.2 Post-Training Alignment

Our post-training pipeline includes supervised fine-tuning on safety-focused datasets drawn from established sources. We use alignment techniques designed to reduce harmful outputs while maintaining helpfulness and capability. We do not currently use reinforcement learning from human feedback (RLHF) or direct preference optimisation (DPO), which we note as a known limitation in our model documentation.

3.3 Ongoing Monitoring

We monitor community feedback, reported issues, and emerging research on model safety. We update our safety practices as the field evolves. We encourage users to report safety concerns or unexpected model behaviour through our contact email.

4. Limitations and Known Challenges

We acknowledge the following limitations of our responsible AI practices:

• Scale: As a small team, we cannot match the safety infrastructure of large organisations with dedicated safety teams
• Scope: Our evaluations cover known safety categories but cannot anticipate all possible misuse cases
• Trade-offs: Safety alignment can reduce model helpfulness or creativity in some scenarios
• Evolving standards: Responsible AI is an active field of research, and our practices will evolve as understanding improves

5. Community Involvement

We welcome input from the research community, users, and the public on our responsible AI practices. If you have identified a safety concern, have suggestions for improvement, or would like to collaborate on safety research, please contact us at support@fabricai.co.uk.

6. Updates to This Policy

We may update this Responsible AI Policy as our practices evolve. We will notify users of material changes by updating the "Last updated" date. We encourage you to review this policy periodically.