🫶Advocating for Responsible AI

Introduction

At Trusst AI, we recognise the transformative potential of artificial intelligence (AI) and machine learning (ML) technologies to enhance customer experiences and operational efficiencies across various industries. Our flagship product, TrusstGPT, exemplifies our commitment to leveraging AI responsibly, ensuring our solutions augment human capabilities while adhering to ethical standards, privacy, fairness, and transparency.

Our approach to responsible AI encompasses the entire lifecycle of AI/ML development and deployment: from design and development through to deployment, and ongoing use. This document outlines our principles, processes, and practices that align with and extend beyond those advocated by industry leaders such as AWS.

Design and Development

Evaluating Use Cases and Bias Consideration

TrusstGPT is developed with a keen awareness of the societal impact of AI. We meticulously evaluate use cases for their potential benefits and risks, especially regarding human rights and safety. An integral part of our process includes bias consideration, where we employ explicit selection and filtering of training data, coupled with human evaluation and labelling to ensure the fairness and safety of model outputs.

Central to the development of TrusstGPT is our commitment to employing fine-tuned, task-specific proprietary models. This deliberate strategy ensures our AI solutions are precisely tailored to meet the unique requirements of our customers' specific use cases. This practice not only elevates the efficacy of our AI applications but also markedly shrinks the models' size. The result is a significant reduction in operational costs and risks for our clients, while simultaneously enhancing performance.

Data Sources and Quality Assurance

Our model training leverages a Trusst AI proprietary dataset, ensuring a rich and diverse data foundation. Quality assurance is rigorously maintained through human review and verification, quantitative evaluations, and red-teaming to identify vulnerabilities and emergent risks. This comprehensive approach ensures our models are robust, reliable, and aligned with our values of responsible AI.

Deployment and Operationalization

Transparency and Automation Bias

Transparency is a cornerstone of TrusstGPT's deployment. We aim to make the capabilities and limitations of our AI systems clear to all users, addressing the challenge of automation bias where overreliance on AI could lead to overlooking human judgment and expertise. By providing detailed documentation and explicit warnings, we ensure users understand the probabilistic nature of AI predictions and the importance of human oversight in critical decision-making processes.

Safeguards Against Model Hallucinations and Adversarial Attacks

To combat model hallucinations and ensure the integrity of our AI outputs, we employ strategies like low-temperature settings, prompt-engineering, and clustering for categorical extraction. Protecting against adversarial attacks is paramount; hence, we deploy models in isolated environments (VPCs) for each customer, ensuring no PII is used in model training and that customer data remains within their control, safeguarding privacy and security.

Ongoing Use and Continuous Improvement

Feedback Mechanisms and Model Testing

Continuous improvement is integral to TrusstGPT's lifecycle. Our built-in feedback mechanisms allow users to contribute to the model's evolution, helping us identify areas for enhancement. Periodic and customer-requested fine-tuning ensure our models adapt to new data and evolving needs, maintaining relevance and accuracy.

Legal Compliance and Ethical Governance

Engagement with legal advisors ensures our compliance with evolving AI and ML regulations globally. We are committed to ethical governance, involving diverse perspectives in our development teams and considering the broader societal impacts of our technologies.

Conclusion

Trusst AI's advocacy for responsible AI use is embedded in every aspect of TrusstGPT's lifecycle. Our policies and practices reflect a commitment to ethical AI, emphasising fairness, transparency, and security. By continually assessing and refining our approaches, we aim to set a benchmark for responsible AI in the industry, ensuring our technologies serve humanity's best interests while driving innovation forward.