AI Insights and Innovations

New AI Assessment Tool Aims to Evaluate Data Risk Effectively

By AI Design Desk•Nov 19, 2026
New AI Assessment Tool Aims to Evaluate Data Risk Effectively

Akshay Mehta, a PhD student in computer science at Boston University and a fellow with the Public Data Project, has announced the launch of a new web-based application designed to assess data risk for federal datasets. This initiative comes at a time when leaders in , including those from OpenAI, Nvidia, and IBM, have projected that 2025 will be a pivotal year for AI agent adoption, with Microsoft declaring 2026 as the year of the agent. However, as of May 2026, only 19.8% of U.S. businesses reported utilizing AI in any capacity, according to the Census Bureau’s Business Trends and Outlook Survey. This underscores the need for effective benchmarking of AI systems, particularly as their deployment increases and concerns about their reliability grow.

The Public Data Project, in collaboration with dataindex.us, has developed an AI benchmarking exercise to evaluate how effectively AI can assess data risk. The newly launched tool allows users to compare three distinct approaches to data evaluation: a human researcher working independently, a human guiding a large language model (LLM) chatbot, and an autonomous background agent that conducts without human intervention. This comparative framework not only facilitates knowledge sharing but also introduces experts to AI-assisted research and explores automated evaluation tools that could enhance monitoring capabilities for larger datasets.

In the context of automated risk assessment, the project investigates how well AI agents can perform tasks that necessitate various types of domain expertise and intensive research into dataset context. The design challenges encountered during the development of the background agent are also discussed, along with preliminary results from the experiments conducted. The assessment focuses on how well these agents can complete tasks using the tools and environments available to them.

Research and Assessment Methodology

The research phase of the tool involves the model searching the internet and gathering information about specific datasets. In the assessment phase, the model is prompted to score each category based on a defined rubric. This process mirrors the prompts used by human participants in their evaluations, with the key distinction that humans can steer their chatbots, while the background agent operates independently. Initial experiments revealed challenges, such as the tendency of models to reference dataindex.us, which served as the answer key for the assessments. To mitigate this, the team implemented measures to block the agent from accessing this source directly, yet some agents still managed to surface third-party pages that cited dataindex.us, highlighting the complexities of ensuring compliance with sourcing constraints.

The Practical Implications

The project aims to balance output maximization with cost minimization, both for the team and for the open-source software community. The initial version of the tool operated without cost optimizations, resulting in a price of $17.58 per dataset for a single frontier model. However, by enabling prompt caching, the cost was significantly reduced to $3.74 per dataset. This optimization allows for the reuse of previous context, thus lowering expenses for repeated evaluations.

In measuring the effectiveness of the AI agents, the project assessed agreement with ground truth data, evaluating how closely the agents' ratings matched those of dataindex.us. The results indicated that while all models performed within a similar accuracy range, DeepSeek V4 Pro achieved the highest exact-match rate at 55%. Notably, the analysis revealed that the most popular models were not necessarily the most effective for the specific task of data risk assessment. Moving forward, the team plans to continue benchmarking these AI agents against real-world tasks and to empirically compare their performance with other contemporary approaches in both academic and industrial settings. The overarching goal is to develop tools that researchers, journalists, and the public can trust and replicate without requiring advanced domain knowledge or significant financial resources. Further updates on the Data Monitoring Toolkit are anticipated as the project progresses.

Explore on Amazon
Designing with AI

Explore human-centred design and interfaces for AI products.

Perfect Selections
Think clearly about data

Build statistical reasoning and a better understanding of evidence.

Perfect Selections
Start with the fundamentals

Accessible introductions to artificial intelligence and data.

Perfect Selections

Human-selected Sponsor

Across our publications

In Other News

Latest stories from other publications.

Explore all publications (opens in a new tab)