Our scalable AI Assurance Implementation (AI2) system is designed to easily test and view results through a simple online interface.
Challenge
As impressive as AI applications are, their models have biases that may lead to poor results or vulnerabilities that a bad actor might exploit. This creates a significant challenge: How do we determine if an AI model is trustworthy? How do we identify weaknesses so we can fix them? And how do we explain these complex problems and solutions to people who aren’t AI experts to facilitate good implementation decisions?
Solution
Noblis developed its AI2 system to rigorously test AI models. It starts with a suite of 14 different benchmarks that focus on three main areas:
- How well the AI model works: Is it accurate? Is it reliable? Can it handle unexpected situations?
- How the AI model affects people and society: Does it have hidden biases that could lead to unfair results? Can we understand why it made a certain decision? Does it appropriately protect privacy?
- How well the AI model follows ethical rules: Is it fair? Is it transparent about how it works? Is anyone accountable for its actions?
AI2 runs these tests automatically, creating easy-to-read scorecards that rate how risky or reliable an AI model is. It’s scalable enough to work on almost any computer system, from a laptop to a cloud computing center, and is designed to easily test and view results through a simple online interface. It even has special tools to verify whether AI language models are telling the truth.
Impact
The impact of AI2 is significant for determining whether AI models are safe and reliable. It offers us a clear picture of a model’s strengths and weaknesses, providing a simple “Risk Score” on a scale from 1 to 5 to characterize quickly how trustworthy it is. For the tech experts, it provides all the detailed information they need to dive deep and fix problems. AI2 has already proven its worth by finding hidden flaws in models used for initiatives such as identifying objects in X-ray scans or recognizing languages. By illuminating these hidden problems, AI2 helps improve overall AI models and their trustworthiness, preventing potential harm and ensuring these powerful tools are used responsibly.
Further Applications
The potential for AI2 goes even further. It can be used to test new AI models as they’re being built, ensuring they’re safe from the outset; help create industry-wide safety and quality standards for all models—and eventually be integrated into all systems development efforts to drive powerful AI models that are reliable, fair and safe for everyone.