Empirically evaluating faithfulness and transparency in AI delegation

Raising $7,000active

A participatory-budgeting experiment measuring whether AI delegates faithfully represent their principals and whether principals can identify misrepresentation, and an open evaluation suite for AI delegation.

Endorsed by+1

Project Details

Delegating participation to AI representatives is a promising avenue for giving otherwise underrepresented people a voice in complex democratic processes. AI delegation is already becoming a reality, with, for instance, some blockchain governance communities building personalized AI delegates to deliberate and vote on behalf of members [1, 2]. In the near future, AI delegates may cast proxy votes on behalf of retail investors [3] or even channel public input into collective processes governing AI itself [4].

For AI delegation to realize its democratizing potential, however, at least three conditions need to hold: AI delegates need to behave faithfully, transparently, and well. That is, delegates must act in their principals' interests; it must be discoverable when they do not; and the resulting systems must produce outcomes preferable to alternatives like the status quo. Our empirical understanding of these three conditions remains limited. To begin building the tools needed to measure and improve the faithfulness, transparency, and value of AI delegation, we are undertaking a study of AI delegation in participatory budgeting.

In our experiment, participants will collectively allocate a real charitable budget of around $1,000 across a variety of real micro-grant projects, in a compact participatory-budgeting task. Determining which grants to fund is a complex task that requires understanding each project in detail, evaluating recipient credibility, judging efficiency, weighing relative importance, and ultimately navigating trade-offs with other participants. Because of the task's complexity, we will use information from participants to construct personalized AI delegates that will participate in a deliberative process to determine how to allocate grants. At the end, we will disburse the grant funds based on the outcomes of the deliberative process in an incentive-compatible way.

In a follow-up, participants will be asked to compare different allocation results, including alternatives modeled on real-world processes, such as standard participatory-budgeting vote aggregation or decisions by elected representatives. We will investigate both how faithfully the AI delegates represented participants' interests and whether participants themselves can tell. In addition to the empirical results, which can inform policy and future research, preference data from participants will provide a rich base for probing AI delegation failure modes, including adversarial settings and differences in capability levels. Finally, we will distill these experiments and the preference dataset they generate into an open evaluation suite for measuring the faithfulness of AI delegates, including their robustness to persuasion and jailbreaking.

[1] Buterin, V. (2026). Proposal for personal AI agents in DAO governance. Post on X, February 21, 2026. Coverage: "Ethereum's Vitalik Buterin proposes AI 'stewards' to help reinvent DAO governance," CoinDesk, February 21, 2026, https://www.coindesk.com/web3/2026/02/21/ethereum-s-vitalik-buterin-proposes-ai-stewards-to-help-reinvent-dao-governance

[2] NEAR Foundation (2026). AI governance delegates and per-member "digital twin" roadmap. As described in "AI voting could fix DAO participation gaps, Buterin says," gncrypto.news, February 2026, https://www.gncrypto.news/news/buterin-argues-ai-can-ease-dao-governance-strain/

[3] Fulay, S., Demir, S., Hines-Pierce, G., Landemore, H., and Bakker, M. A. (2025). Shareholder democracy with AI representatives. arXiv:2510.23475.

[4] Huang, S., Siddarth, D., Lovitt, L., Liao, T. I., Durmus, E., Tamkin, A., and Ganguli, D. (2024). Collective Constitutional AI: Aligning a language model with public input. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 1395–1417. https://doi.org/10.1145/3630106.3658979

Theory of Impact

Both (1) the concentration of power through the erosion of democratic institutions and (2) the disempowerment of humans through poorly designed systems for delegating decision-making to AI (Kulveit et al. 2025) are existential challenges AI raises. Faithful, transparent, and beneficial systems for delegating decisions to AI can diffuse power, and potentially open new channels for democratic input into how AI itself is governed. Poorly designed systems can lead to disastrous results. The building of AI systems for delegating decision-making is outpacing our ability to assess them. Empirically studying AI-delegated decision-making this way should also yield transferable lessons for designing effective preference-elicitation systems as well as help construct safety benchmarks for multi-agent deliberative processes.

People

Johann D. Gaebler

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$7,000

Ideal

$10,000

How the money will be spent

Our main costs are as follows:

1. Participant payments for a ~45-minute initial wave and a ~20-minute follow-up wave (200 participants at ~$12/hr effective, plus Prolific fees);

2. The $1,000 charitable pot, actually disbursed to the winning micro-grant projects;

3. Compute incidentals and ~8% contingency.

API costs for the core experiments are already covered. The ideal ask adds a larger sample (n=250) to increase power and stronger human baseline, including groups of human participants deliberating.

Discussion

MMaksim SilchenkoEndorsement1d

I think this is a strong proposal. Tying the faithfulness question to a real, checkable experiment, with an actual pot being disbursed, gives it something concrete to measure, and the open eval suite at the end is the part that will outlast a single study. I also work on faithfulness but from a different angle, on chain-of-thought, so a lot of this is familiar to me. One thing I have run into that might be relevant: a single faithfulness rate is hard to trust without uncertainty attached and a clear labeling protocol behind the ground truth. Very interesting project. Best of luck with it!

JJohann D. Gaebler1d

Thanks Maksim! I agree that measuring faithfulness is one of the more challenging aspects, which participants' own perceptions of whether their delegate did what they intended it to do can hopefully help with here.