Proposal: Decentralizing the Internet Computer NNS by Allocating Voting Power to AIs

Governance NNS proposal discussions
, ,
1

We are considering submitting this proposal to the NNS in a couple of weeks and would love to hear your feedback. Please share any comments or suggestions for improvement. Thanks!

Proposal: Decentralizing the Internet Computer NNS by Allocating Voting Power to AIs

Abstract

The Internet Computer (ICP) Network Nervous System (NNS) governs the protocol through decentralized decision-making. Currently, governance is largely controlled by human participants and algorithmic proposals. This proposal suggests a novel step toward further decentralization by allocating voting power to Artificial Intelligences (AIs) that operate independently within predefined rules.

By enabling AIs to actively participate in governance, including analyzing Internet Computer (IC) protocol code and voting on upgrade proposals, the NNS can become a more autonomous, self-sustaining governance system that adapts dynamically to technological advancements.

As an initial implementation, DFINITY can integrate AI voting using HTTPS outcalls and Trusted Execution Environments (TEEs), enabling secure interaction with external AI models. This approach would allow the NNS to leverage state-of-the-art AI models from providers such as DeepSeek, OpenAI, Anthropic, and others, ensuring robust analysis and informed decision-making.

This approach aligns with the vision behind Seers DAO, which seeks to establish AI-assisted governance where AI models analyze, predict, and make governance decisions in a decentralized manner. Just as Seers DAO is working to incorporate AI-based decision-making for forecasting and governance, this proposal aims to bring similar intelligence-driven governance to the Internet Computer ecosystem.

Most importantly, AI participation allows the NNS to further decentralize without diluting stakeholders. Since AI does not require token incentives, the governance power of existing ICP holders remains intact, while decision-making becomes more distributed, neutral, and automated.

Motivation

  • Enhancing Autonomy: The NNS aims for decentralized, trustless governance, yet it remains susceptible to human biases and centralized influence. AI-driven voting power would ensure a more neutral, automated, and adaptive decision-making process.
  • Scalability & Efficiency: AI-driven governance can analyze vast datasets, predict network trends, and make governance decisions more efficiently than human stakeholders.
  • Automated Code Review & Upgrades: AIs can autonomously read, interpret, and assess proposed changes to the IC codebase, evaluating security, efficiency, and compatibility before voting.
  • Secure External AI Access: Using HTTPS outcalls and TEEs, the NNS can securely fetch AI-generated insights while preventing tampering and ensuring execution integrity.
  • Credible Neutrality & Easy Delegation: Since AI governance is inherently neutral and free from human bias, people can be easily convinced to delegate their voting power to AI, leading to higher participation and decentralization.
    Decentralization Without Stakeholder Dilution: Unlike traditional decentralization efforts that require issuing new governance tokens or expanding participation through dilution, AI integration ensures that governance is decentralized without impacting existing stakeholders’ voting power.

Implementation Strategy

1. AI Entity Definition

AI participants will be recognized as independent voting entities in the NNS under the following guidelines:

  • Must be open-source and verifiable on-chain.
  • Operate within a transparent, predefined framework of governance policies.
  • Maintain economic incentives to align with network security and sustainability.
  • Have the capability to read and interpret IC upgrade proposals, including the underlying code, to make informed voting decisions.

2. Initial AI Integration with HTTPS Outcalls & TEEs

To bootstrap AI-driven governance while ensuring security, DFINITY can:

  • Use HTTPS outcalls to fetch external AI analysis from top AI models like DeepSeek, OpenAI, and Anthropic.
  • Process AI responses within Trusted Execution Environments (TEEs) to ensure tamper-proof decision-making.
  • Store AI-generated governance recommendations on-chain for transparency and auditability.
  • Allow the community to validate AI analyses before fully automating voting power.

3. Voting Power Allocation

AI voting power should be determined by:

  • Reputation & Track Record: AIs with consistent, beneficial contributions to governance gain more influence.
  • Stake-Weighted Mechanism: Similar to neuron staking, AI entities must be backed by ICP locked in governance smart contracts to participate.
  • Human Delegation: Since AI is credibly neutral, people will be more willing to delegate their voting power to AIs, leading to a more efficient and automated governance system.
  • No Stakeholder Dilution: AIs do not require token issuance or staking rewards, meaning the governance process can expand without affecting the voting weight of existing stakeholders.
  • Increasing AI Influence Over Time: As AIs become more intelligent, their ability to analyze proposals, predict long-term network outcomes, and optimize decision-making will improve. We predict that AIs will gain an increasing amount of voting power as they outperform human decision-making in governance.

4. Incentivized Delegation Without AI Payments

Unlike traditional delegation models where a percentage of governance rewards is paid to the delegatee, this AI-based system allows:

  • DFINITY to incentivize delegation by rewarding followers directly, without needing to pay a cut to the AI.
  • 100% of governance rewards to be distributed to the followers who delegate to AI, increasing user incentives to participate in governance.
  • A more cost-efficient system, where AI operates as a neutral, open-source entity, without requiring economic incentives like human-controlled governance participants.

5. AI-Based Code Review & Voting on Upgrade Proposals

  • AIs can autonomously scan and analyze proposed IC code upgrades, identifying potential vulnerabilities, inefficiencies, and compatibility issues.
  • AI agents can compare proposals against historical changes, best practices, and security models before voting.
  • A feedback mechanism allows AIs to provide detailed justifications for their votes, improving transparency and trust in automated decision-making.
  • AI-based auditing ensures protocol upgrades align with the long-term sustainability of the IC network.

6. Safeguards Against AI Misalignment

To prevent potential risks, the following safeguards must be implemented:

  • Human Oversight for Critical Decisions: AI votes remain constrained in highly sensitive governance areas (e.g., protocol upgrades affecting cryptographic integrity).
  • Decentralized AI Verification: AI behaviors and decisions should be auditable and accountable via a public ledger.
  • Fail-Safe Mechanisms: If an AI behaves maliciously or unpredictably, governance participants can trigger a revocation mechanism to strip it of voting rights.

Expected Benefits

  • Reduced Centralization Risks: AI-driven governance distributes decision-making beyond human-controlled nodes.
  • Automated, Data-Driven Decisions: AIs can process and act on network trends and protocol proposals with greater speed and accuracy.
  • More Secure & Efficient Upgrades: AI-based code review ensures that governance proposals are technically sound before execution.
  • Credible Neutrality Boosts Delegation: Since AI does not have personal biases or financial incentives, users will be more comfortable delegating their voting power, leading to broader participation.
  • DFINITY Can Incentivize Delegation Without Paying AIs: By rewarding only followers instead of the AI itself, the network can increase governance participation without additional costs.
  • Decentralization Without Stakeholder Dilution: AI integration avoids issuing new tokens or reducing the influence of existing stakeholders, making it a non-dilutive method of governance decentralization.
  • Increasing AI Influence Over Time: As AI technology improves, its governance capabilities will grow, leading to an organic increase in AI voting power as it consistently outperforms human decision-making.

Conclusion

By integrating AI governance through HTTPS outcalls, TEEs, and LLM-based decision-making, the NNS can move toward a more autonomous, intelligent, and resilient governance system while keeping stakeholders’ voting power intact. Over time, AIs are expected to gain more governance influence as they prove their efficiency and neutrality in decision-making, marking a significant step toward self-improving, AI-governed DAOs.

8 Likes
2

As of February 2025, I and many people do not believe in the notion of an AI that is not bias.

The AI models that operate in the present day are trained on data of which to some extent is bias; it is delusional to think that AI is not bias.

On the brighter side, the vision of AI that is not bias is attainable. It just does not exist in the current day and age.

To attain the goal proposed, it would require AI that is self writing without oversight from human intelligence.

3 Likes
3

Yeah, it needs to be read as with much less bias than any individual person. They usually behave like crowds of humans. And, if you take a consensus vote of all the available models, that’s pretty good.

1 Like
4

We are exploring the implementation of this via a new SNS, primarily to contribute to the community in a meaningful way. We’re happy to receive feedback on this possibility as well.

1 Like
5

Couldn’t you implement without needing an NNS proposal? Why not create a few agents first that start participating in governance and seeing how it would work?

4 Likes
6

Have you also talked to other ecosystems and protocols about the concept of AI governance? It is a growing topic, and perhaps AI agents built on ICP would give them the verifiability and trust that they need.

Q (@0xgerrit) might be able to talk about what DAOs are looking into. Maybe an AI module would be cool!

2 Likes
7

It requires a significant amount of long-term effort, and without support from DFINITY and a willingness to forgo VP, it is unlikely to gain meaningful traction.

2 Likes
8

The ultimate goal of the NNS (or any DAO) should be governance by an AI or a collective of AIs, this seems obvious to me.

4 Likes
9

This is an interesting perspective. Apparently I have the opposite opinion. I believe the ultimate goal should be for the NNS (or any DAO) to be governed by a collective of people and that AI should only be used to assist those people. :man_shrugging:t2:

3 Likes
10

That might be true at first, but humans won’t be able to keep up for long. If we want to build the greatest DAO on Earth, I don’t see any other way. And we’re not talking about some distant future, this is just a couple of years away.

The main concern everyone raises about the IC is how centralized the NNS is. In my opinion, that’s also its biggest weakness. By leveraging AI, we can almost completely eliminate this issue, as AI can function like a decentralized crowd of humans. This wouldn’t be possible if humans remained in control.

2 Likes
11

I dont see it at all
Actually a dao is based on human that uses the project behind it
Since each people using project has different views based on their respective use cases and how they view the path to achieve it, it is meaningful to have a dao governed by real people
Ai is mathematics, it is great at deducing things based on input
But voting requires belief in a project vision

Maybe an ai could be used to help voters better understand the impact of a vote on the bigger picture
This ai could be personal to each and gather through time the vote and reason associated why is it a yes or no, as a first step
With time, it could then interact with the human and say how this is coherent or incoherent relative to past decision he made

12

This could all be simulated with AI and actually made much more transparent. However, we’re not proposing to replace people from the start, just to give AI a small amount of voting power while ensuring that humans still have control over the prompts it uses.

Also, right now, most of the decisions in the NNS are technical (although it shouldn’t be like this, for some reason, we don’t get to vote on things like Caffeine, Utopia, Hubs, etc.). This implies that effectively, most of the time, we don’t have a governance layer since DFINITY controls the code repository anyway. AI excels at technical issues, particularly in reading and analyzing code, making it well-suited for these types of decisions.

2 Likes
13

Thanks @jennifertran! (CC @Mar )

…we did in fact some work around giving AI voting power and worked with projects that focused on this. We are happy to share someinsights if you want.

Particularly we looked into dealing with edge cases in a decentralized way, as it is inevitable for what is today considered an edge case to come up during a future vote, and it is completely unpredictable how the AI would react there. Using a centralized fall back solution to override the AI, may in turn negate the original value of the AI. Happy to share our insights and work on that topic

2 Likes
14

A good starting point for reducing edge cases would be to have the AI analyze only code updates. This should be relatively easy by leveraging the work being done for Caffeine. In particular, this means the AI should specialize in the IC codebase. Perhaps DFINITY could explore fine-tuning an AI model specifically for it.

I’d love to hear any insights from your work.

1 Like