8 Core Principles for Responsible and Ethical AI
Commitment to transparency and accountability
We ensure that our AI models and algorithms are transparent and understandable by thoroughly documenting the complex models we've created, trained and fine-tuned. This transparency guarantees that our customers understand how data is found, sorted, and extracted, giving them insight into how their fine-tuned models will continue to improve and learn over time.
Explainability is a top priority
We create AI that provides clear and understandable insights to support decisions and actions. Our interpretable algorithms and tools include traceability and field level confidence scoring, which use customizable thresholds to tell users when document processing can be completed automatically or when manual review is needed by your subject matter experts.
.png?width=300&name=artificial-intelligence%20(1).png){kind=icon}
Embedding AI reliability and safety
Roots Automation was developed by experienced insurance operators, and our technology helps humans make better underwriting or claims decisions. We extensively constantly evaluate our InsurGPT model to ensure that it does not hallucinate and responds based solely on the documents presented. Once in production, we monitor models in real time to detect and fix anomalies, errors or model drift as they occur.
Safeguarding privacy and data security
Our philosophy is simple: Your data belongs to you, not us, and we will keep it that way. We respect and protect the privacy of individuals and companies by ensuring maximum data security and adhering the principle of data minimization, gathering and using only the data required to enable model performance. Unlike public AI solutions (e.g., OpenAI, Meta, LLaMA and others), client data is never exposed to the open internet, and we never use client data to train models built for use by another customer.
Removing bias and prioritizing fairness and inclusivity
The insurance industry deserves, and regulators require, AI systems that are fair, equitable, and inclusive, with no biases that could lead to discrimination. At Roots we regularly audit, test, and fine-tune our models to reduce any biases, particularly those linked to race, gender, age, and socioeconomic status. We actively involve different teams and stakeholders in the AI development process to discover and eliminate potential biases early.
Model testing, reproducibility, robustness and validation
We use rigorous processes for testing, reproducing, and validating AI models to ensure that they are robust, reliable and perform consistently in real-world circumstances. Our testing protocols evaluate AI models under diverse scenarios, including edge cases and stress tests, to ensure reliable performance in real-world conditions. To ensure accuracy and robustness, we routinely verify models with new data and real-world scenarios and update them on a regular basis to reflect changes in data patterns or underlying assumptions.
.png?width=300&name=shield%20(1).png){kind=icon}
Establishing data governance and compliance
Effective data governance is critical for ensuring the integrity, security, and privacy of data in AI systems. Compliance with legal and ethical standards, such as GDPR or CCPA, ensures responsible data practices that respect individuals' rights. This principle supports the trustworthiness of AI systems by ensuring that data is managed securely, ethically, and compliantly. Roots conducts yearly third-party review for SOC 2 Type 2 attestation and complies with ISO 27001, NIST, HIPAA, CCPA, GDPR, and 23 NYCRR Part 500 standards.
.png?width=300&name=user%20(1).png){kind=icon}
Fostering human-AI collaboration
AI works best when it enhances human skills. Our solutions encourage collaboration between human subject matter experts and AI, and we want our AI to supplement and improve human decision-making rather than replace it. Roots’ patented Human-in-the-Loop (HITL) capabilities allow underwriting and claims professionals control over low-confidence extraction scenarios. Users can define confidence thresholds and create default alerts that prompt human intervention for the specific response area. Our AI then collects the users’ feedback to continuously improve the system.
