AI Guardrails Need Smarter Tools

The AI safety conversation is shifting. Beyond mere capability checks, the focus is now on real-world performance, especially within specific contexts and languages. This evolution highlights a critical area: the guardrails themselves, the mechanisms intended to keep AI in check.

Guardrails Under the Microscope

At the ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT), researchers argued that these guardrails deserve the same rigorous evaluation as the AI models they govern. Historically, guardrails operated as proprietary black boxes. However, the rise of open-source models and policy-prompting now allows for more independent assessment.

The central thesis presented was clear: guardrails must move beyond static rules. They need to be informed by context- and language-specific evaluations to address nuanced failures effectively.

Bridging Evaluation and Implementation

Connecting AI model evaluation to practical guardrail implementation is a complex challenge. A project undertaken involved community and language expertise to evaluate 120 refugee and asylum-focused scenarios across English, Farsi, Arabic, Kurdish-Sorani, and Pashto. Native speakers scored these scenarios based on six rights-based criteria.

The published results revealed recurring issues like unsafe referrals and stereotyped assumptions. These findings were then translated into concrete guardrail policies.

The Gap: Static Policies Fall Short

Testing these policies exposed a significant limitation: criteria like factuality and actionability cannot be reliably judged from text alone. Guardrails without external verification tools often approved responses with fabricated terms or missed factual errors.

This led to a hypothesis: AI guardrails need tools such as web search and fact-checking capabilities to perform more reliably.

Introducing Agentic Guardrails

The team demonstrated how agentic guardrails, leveraging tools like web search, could enhance evaluation accuracy. A hands-on session at FAccT allowed participants to compare agentic and non-agentic judges on AI responses.

While most verdicts agreed, the use of tools by the AI judges significantly impacted the supporting evidence. When tools were invoked, they could either verify factual claims, boosting a score, or identify overlooked errors, downgrading a verdict.

The effectiveness of these tools heavily depended on the underlying language model. Some models actively used web search, while others rarely did.

Tooling for Practicality

Making such experimentation feasible requires robust infrastructure. Tools like Mozilla AI’s open-source any-guardrail offer a unified interface for configuring guardrails. Otari, a new open-source LLM gateway from Mozilla, allows seamless switching of different LLMs for judging responses.

The ongoing research aims to further test if tool access makes AI-enabled guardrails more trustworthy across various use cases, including humanitarian, financial, and social-engineering scenarios.