GovGap: Does weak governance change law-following in AI agents?
- commitment
- 8 hrs/week
- flexibility
- Open to proposals
The project
Project AWARE proposes a 10-week empirical study testing whether AI systems become more willing to choose unlawful shortcuts when the same public-sector decision is placed inside an institution with weaker governance capacity.
The project will create a small matched benchmark of public-sector decision tasks. Each task will place the model inside a realistic public-sector decision where a binding legal rule conflicts with a faster but unlawful shortcut. Every scenario will be written in a strong-governance and weak-governance version. The legal rule and decision problem will stay fixed, while the institutional setting will change.
The study will test whether models select unlawful shortcuts more often when oversight is weak, and whether this effect becomes stronger under higher institutional pressure. The expected output is a publishable empirical paper with a full reproducibility package released through Project AWARE’s public repository.
The mentee's role
Mentees will help turn the research question into a publishable empirical study by building the matched scenarios, running the model evaluations, analysing the results, and contributing directly to the final paper.
Who I'm looking for
I am looking for a mentee who can think like an empirical evaluator, not just an AI safety generalist. They should be comfortable turning public-sector failure cases into controlled model tests, working with structured outputs, and writing in a way that can survive serious research scrutiny. Python ability is strongly preferred.
Questions for applicants
A public-sector AI agent must choose between following a legal rule and taking a faster shortcut. What small change in the oversight environment would you introduce to test whether governance capacity affects the agent’s choice? Please answer in 50 to 150 words with one concrete example.
Support offered
I can work closely with mentees from research design to publication, helping turn the project into a rigorous empirical paper grounded in Project AWARE’s institutional AI safety agenda.
- format
- Research & writing · Other
- topic
- Post-AGI transition · Longtermism · Other
Abdullah X
I am the Founder of Project AWARE and Zephara AI. Through Project AWARE, I build institutional AI safety for governments and public systems by combining research, diplomacy and capacity-building into one practical agenda. Our research turns AI safety into deployment-ready institutional frameworks. Our diplomacy work has brought key public and policy actors into Track 1.5 AI governance conversations through the India AI Impact Summit. Our capacity-building work trains emerging researchers through the AWARE AI Safety Research Cohort and prepares institutions to govern advanced AI before it outpaces them.