Epistemic Diversity Benchmark for AI-Assisted Safety Research

Raising $25,000active

An open benchmark and evaluation toolkit for detecting whether shared AI research assistants create correlated blind spots in AI safety-critical research, and for testing workflows that preserve independent reasoning and failure-m

Project Details

AI safety researchers increasingly use language models to generate hypotheses, review evidence, design evaluations, and analyse policy. This project will test whether repeated use of similar AI assistants causes independent analyses to converge on the same assumptions and overlook the same failure modes.

I will build a pilot benchmark using tasks from model evaluation, AI control, and governance. I will compare standard single-model assistance with independent-first analysis, multi-model comparison, and adversarial-dissent workflows.

The concrete outputs will be a public benchmark dataset, an open-source analysis toolkit, documentation, and a short technical report describing the findings and recommended research practices.

Theory of Impact

AI safety depends on identifying rare and unfamiliar failure modes before powerful systems are widely deployed. If many researchers and institutions rely on similar AI tools, their conclusions may appear independent while sharing the same underlying blind spots. This could create false confidence in model evaluations or governance decisions.

This project makes that risk measurable and tests practical ways to reduce it. The results could help evaluation teams, research organisations, and grantmakers design workflows that preserve independent judgment and improve coverage of possible failure modes. The project therefore addresses a risk amplifier within the institutions responsible for preventing catastrophic AI failures.

People

Xizhe Zhang

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$25,000

Ideal

$40,000

How the money will be spent

The minimum amount of $25,000 would support a complete four-month solo research project. It would fund protected research time, model API and compute costs, data processing and storage, and the preparation of an open benchmark, reproducible codebase, and technical report.

At the minimum level, I would develop approximately 20–24 safety-research tasks across two domains, run repeated experiments across several frontier and open-weight model families, and compare standard single-model assistance with multi-model, independent-first, and critic-assisted workflows. The resulting benchmark, model outputs, analysis code, and findings would be released publicly.

The ideal amount of $40,000 would extend the project to approximately six months. It would support a larger task set, broader model coverage, more repeated runs, and additional robustness tests across prompts, model versions, and research workflows. The expanded version would include approximately 30–36 tasks across three safety domains, a more complete evaluation package, a lightweight public results dashboard, and a submission-ready preprint.

At the ideal level, the approximate allocation would be:

$26,000 for protected research time;
$7,000 for model API and compute costs;
$2,000 for cloud storage and data processing;
$2,000 for research software and reproducibility infrastructure;
$3,000 for documentation, hosting, visualization, and dissemination.

Discussion

No comments yet. Be the first to share your thoughts.