Frontier AI Election Evaluation Framework

active

Developing a practical evaluation framework to identify governance failures in frontier AI systems during elections, strengthening democratic legitimacy and the institutional capacity needed to reduce catastrophic risks from AI.

Endorsed by

Gavin

Ryan+4

Endorsed by

Gavin

Ryan+4

Project Details

I am developing a practical framework that adapts traditional election observation methods to test how AI systems perform during elections. With more voters using LLMs and AI-mediated search as primary information sources, we need a reliable way to monitor the risks. This project gives civil society and policymakers the evidence they need to develop policy recommendations and advocate for safe AI and reduce the risk that AI failures erode democratic legitimacy.

Over the next four months, I will build and pilot an open-access toolkit for domestic election monitors and democracy practitioners. The framework will include prompt libraries covering common election-related information requests, testing protocols, documentation templates, implementation guidance, and recommendations for interpreting and communicating findings. Because elections occur across diverse political and linguistic environments, the project is designed to produce an evaluation framework that can be adapted across contexts.

To build this, I’m combining classic election-monitoring principles with modern AI evaluation practices. I'll be refining the methodology alongside election experts, relationships I've built through a decade of work in elections, including representing the US government as an international election observer in countries like Azerbaijan and Kyrgyzstan. I will also consult experts in AI evaluations and model testing to ensure the methodology is statistically sound and avoids overstating conclusions.

Theory of Impact

Managing existential risk requires high-functioning states that can cooperate, pass smart regulations, and enforce guardrails. If AI systems break the public's trust in elections, then state capacity disappears.

This project tackles x-risk in three ways:

If AI-driven misinformation breaks society's shared reality, political systems break down, and globally we will not be able to coordinate on existential threats. Keeping the information ecosystem stable keeps the door open for safety governance.
This framework brings safety testing into the wild, across diverse languages and political environments. It creates a massive, independent feedback loop that forces developers to fix structural vulnerabilities before their models become more powerful and autonomous.
By training domestic monitors, civil society groups, and ordinary citizens to rigorously test these systems now, we are building the global auditing infrastructure we will desperately need as AI capabilities scale toward critical thresholds.

People

Genevieve Shea

Team Member

Grants Received– no grants recorded

Add grant

Discussion

Private UserPrivate3mo?

Anton Makiievskyi 🔸3mo

Isn't this application AI generated, Genevieve?

Genevieve Shea3mo

Only where permitted by the application!

Matt Brooks3mo

Ah - sorry about that @Genevieve Shea! We might have had a bug in the AI evaluator. We have fixed it and re-ran all applications. Your application comes back as human-written.

MMark AikenEndorsement3mo(edited)

Endorsed, as I think this is a good idea. However, please note that UNDP has released two papers on the use of AI in elections over the past month. Good if you can also take note of these, to avoid reinventing the wheel to whatever extent your ideas overlap.

UNOPS is normally heavily involved in elections administration and elections monitoring, and at some point they will probably also come up with some thinking on this topic. I havent seen anything from them so far.

Genevieve Shea3mo

Hi Mark!

Thank you for your endorsement and thoughtful feedback. I’m familiar with the recent UNDP work. As you noted, that work focuses on election administration - i.e. how election management bodies are using or could use AI across the electoral cycle (e.g. voting, counting, tabulation), and the governance frameworks needed for responsible adoption.

My project focuses on the electoral information environment. Having worked on elections in authoritarian and backsliding contexts for most of my career, I have seen many times that a technically well-administered election does not necessarily mean that election was free, fair, or competitive. Observer assessments that focus on election-day procedures can actually legitimize unfair elections by validating their administration without accounting for manipulation, repression, and distortion that shaped the environment before e-day. AI will amplify these vulnerabilities considerably: scaling manipulation, reinforcing information asymmetries, and enabling dangerous concentrations of power while preserving the appearance of legitimate electoral processes.

My project will address this so far neglected area by equipping independent domestic election observers with methods to evaluate voter-facing AI systems on whether they provide accurate election information, exhibit political or demographic biases, tailor responses in ways that reinforce or manipulate users’ existing preferences, or can be exploited to distort political information at scale. The goal is to give observers credible evidence to escalate identified risks to developers, regulators, and policymakers.

I suspect much of this context may already be familiar to you, but I wanted to spell out the distinction for others who may be evaluating the proposal and want to understand how it is different than existing work in the field.

MMark Aiken3mo(edited)

Thanks for the reply. In relation to the electoral information environment, please note for example the UNDP / UNESCO paper from last year on freedom of expression in elections (https://unesdoc.unesco.org/ark:/48223/pf0000393473) and the case study on UNDP's work in digital monitoring of elections information environment in the Nepali Federal elections held in March 2026 (which found 388 violations of the electoral code of conduct on disinformation, hate speech etc), using a combination of AI social media monitoring and domestic election observers (https://www.undp.org/sites/g/files/zskgke326/files/2026-05/final-consolidated-report-2026.04.08bb.pdf).

I know you're familiar with UNDP's work, so this is just a mention :) And for other readers, I continue to endorse this application.

AAbeer SharmaEndorsement3mo(edited)

This stands out as an AI governance project because it is practical. I would like to see more work of such a nature.

Genevieve Shea3mo

Thank you Abeer!

Pim KangEndorsement3mo

I support this project! It sounds really promising and crucial.

Genevieve Shea3mo

Thank you Pim! Lots of thanks is owed to the team at Successif for being the first sounding board for this idea.

Karolina GruzelEndorsement3mo

When talking to Genevieve, you can clearly see that she is very knowledgeable and passionate about this subject. In the past, we saw a lot of claims about targeted political content and its impact on polarisation and elections being made ahead of sufficient empirical support, so a rigorous evaluation is much needed in this area. Also, WEF's report identified misinformation and disinformation as the second-most severe global risk over the next two years and the fourth-most severe over the next ten years.

Olin Thakur3mo

I think this is an important and impactful project. Best wishes.

Genevieve Shea3mo

Thank you Olin!