Using ‘Trust Layers’ to make Responsible AI

In today’s blog post, we get a bit technical. We introduce trust layers and provide something of a shopping list.

A trust layer is a mechanism that aims to ensure AI systems behave reliably, safely and ethically.

In the last couple of years, tech giants have been pioneering such mechanisms for their products. For example, Salesforce has created the Einstein Trust Layer for its generative AI applications utilising Large Language Models (LLMs).

According to Salesforce:

·        it uses a standard protocol to encrypt a prompt when it is sent to a LLM.

·        in some apps, you can mask personally identifiable information (PII), replacing PII in a prompt with placeholder data

·        zero retention agreements are in place with LLM providers, so LLMs forget both the prompt and the response as soon as the response is sent back to Salesforce.

But there are also trust layer tools for AI developers. A key feature of these packages is machine learning interpretability (MLI). This just means these tools are supposed to help us understand why an AI did a certain thing or came up with the answers that it came up with.

Examples of these products include:

IBM Watson OpenScale - Helps track AI model performance, explain outcomes, mitigate risks and bias, and govern models. Provides transparency and interpretability for AI.

Microsoft’s Azure AI - A toolkit and governance package for identifying issues with machine learning systems like fairness, interpretability, and privacy. Helps organizations build responsible AI practices.

Google Explainable AI - A set of techniques and tools to help understand and audit machine learning models, with explanations for individual predictions. Helps improve transparency with human-interpretable explanations of models.

H2O Driverless AI - Includes automatic feature engineering, model validation, explainability, and automatic documentation to increase transparency and trust in models.

What are the benefits of trust layer tools?

- They can encode ethics directly into the AI, guide it towards fair, explainable and unbiased decisions that serve all stakeholders. This helps uphold principles around safety, security and privacy.

-They can help demonstrate transparency and accountability. Trust layers are akin to documentation, testing and audit trails. Having it in place can show regulators and customers how your AI system works and help assure everyone it will behave properly.

- They can verify system performance over time. Ongoing monitoring via trust layers can detect if the AI begins acting erratically or produces unfair outputs, allowing the issues to be addressed. This helps reliability and consistency.

- They can simplify coordination with partners. With trust layers that document inputs/outputs and boundaries of the AI, it can more easily integrate safely into third-party products and ecosystems.

Off-the-shelf trust, governance and responsible AI solutions

There are some more targeted off-the-shelf solutions emerging that can help build AI with decent trust and governance in place.

Explainability Libraries: Libraries like InterpretML, Alibi Explain, and DALEX provide common model-agnostic explainability algorithms out-of-the-box.

Bias Detection Services: IBM AI Fairness 360 is an open source toolkit that can help you examine, report, and mitigate discrimination and bias in AI models. Fairlearn allows you to assess and mitigate fairness issues using their Python toolkit. Pymetrics provides methods to audit and test models for bias in hiring and employment.

Model Monitoring: Tools like Weights & Biases, Comet and Valohai enable quick setup of ML experiment tracking, model management, and production monitoring.

Trust Assessment Frameworks: Initiatives like the EU's Ethics Guidelines for Trustworthy AI provide a comprehensive assessment list focused on areas like transparency, fairness, and accountability.

The Bottom Line

Leveraging these off-the-shelf solutions can save AI teams significant development overhead. Though some level of customization may still be required for the specific application and use case. Regulatory standards around algorithmic transparency and accountability are also continuing to evolve quickly, so constant vigilance is still required even when leveraging pre-built trust packages and tools.

When Waltzer Consulting is engaged to help an organization with responsible AI, we will often work with stakeholders to determine how some of these valuable tools can be deployed.