Liquid Democracy for Distributed AI Systems
Creating Reputation Consensus and Liquid Democracy based on an open ledger protocol, for distributed AI systems.
Introduction
History has made it abundantly clear that an effective and reliable level of governance is a necessary element for any community to survive.
In the ancient times, brute force was employed to create an acceptable level of consensus in the population. Naturally, such a consensus mechanism was favorable to a minority of the population that could employ the brute force on others. Over time, as the different human societies evolved — their systems of governance transformed to remain relevant and applicable.
While previously technology has shaped the governing mechanisms of human communities, it will soon create entirely new communities — ones that have Artificial Intelligence (AI) or Artificial General Intelligence (AGI) systems.
Which raises the question: how will such communities be governed?
Types of Democracies
Today, democracy is the most prevalent system of governance. Let us briefly consider three different types of democracies:
In a Direct Democracy, the voters are directly involved with the governance of the community. The voters get to express their opinions on every issue that concerns the community, by voting directly on each one of them. Therefore, a direct democracy offers its community members full control over all the issues that concern them. The main problem that plagues a direct democracy is that people are not knowledgeable enough or have the time, to vote on every issue that concerns the community. This problem becomes more prevalent as the size of the community increases. Therefore, direct democracies in their pure form are very hard to scale.
In a Representative Democracy, voters delegate their voting rights to representatives who vote on their behalf. The representatives are expected to be experts about the various issues that affect the community. Representative democracies, however, solve the scalability issue by concentrating power in the hands of a few. If the representatives are corrupted, they might not act with the best interests of the community in mind.
In a Liquid Democracy, the voters get the best of both worlds. Like a direct democracy, they can vote directly on an issue. If they do not want to vote on the other issues, they can authorize others to vote on their behalf. Various highly reputable experts can be authorized by the voter to vote on their behalf on various issues. Unlike a representative democracy, where voters have to wait for the next election cycle, in a liquid democracy, the voters can very quickly cancel their authorization of such experts if they feel they are not being represented correctly. Therefore, in a liquid democracy, those authorized with the highest number of votes have that power because they maintain a good reputation in the community.
Trust in Reputations
The communities in the future will most likely have a combination of human members and AI/AGI systems. The importance of creating a reliable and effective governance mechanism for such communities that is the most resilient to manipulations or corruptions, cannot be emphasized enough.
The reliability and effectiveness will prove particularly challenging in the case of AI communities because of the speed and scale of electronic communications and the low latency in system response. As we have discussed earlier, in its entire history humanity has not been able to create a mechanism to achieve genuinely democratic consensus. History seems to be repeating itself in the way the consensus mechanisms of modern distributed systems are being designed.
One can see in the Proof-of-Work (PoW) consensus mechanism, that operates many popular blockchain networks, the principle of force is power. In such networks, those who own the highest amount of computing resources govern the network.
Similarly, in almost all of the modern day communities, the financial elites can shape the consensus of the governance systems through their financial power. As they steer the consensus toward scenarios that are favorable to them, the rich are bound to get richer and so gain more power. This principle of money is power is reflected in the second most popular consensus mechanism of modern distributed blockchain systems: Proof-of-Stake (PoS).
The underlying assumption behind the Proof-of-Stake consensus mechanism is that if people “stake” a relatively large value as a guarantee of good faith, which they would risk losing the asset if they act against the rules, the principle of self-interest will prevent them from trying to game the consensus mechanism. For acting in good faith, the stakers are rewarded with more value. In the Proof-of-Stake mechanism hence, just like our societies, those with more money govern the network — and the system rewards them by increasing their wealth.
One consensus mechanism that claims to be the solution to the ills of Proof-of-Stake, and is being used currently to design AGI ecosystems, is called the Delegated Proof-of-Stake (DPoS). A careful analysis of the mechanism makes it clear that in DPoS the rule on the basis of financial capabilities is implemented indirectly — those who are to stake their assets are voted as the selected delegates by the community members, to act in good faith. The assumption is that if these delegates abuse their powers — the community can vote them out. Those voted to be in such a lucrative position, however, have a strong financial incentive to remain in that privileged position and so they find various schemes to subvert the DPoS consensus mechanisms using their financial power. Besides, such a system cannot be implemented in the AI communities. Such communities will operate at very high speeds — making them uncontrollable, due to our limited human capabilities.
It is obvious, therefore, that the currently popular consensus mechanisms of the distributed systems are not suitable for any emergent AI community. These mechanisms risk the possibility of an AI system achieving dominance over others. Such an AI may either be hostile to the majority of other AI systems in the community or to the humans of that community. Such a scenario might be possible due to the emergent hostility of an AI system that is hostile to humans (the Transformative AGI scenario) or the humans managing the dominant AI system can decide to favor a minority of humans over the majority (the Swiss Army Knife AGI scenario).
To avoid such scenarios, we suggest that the distributed AI/AGI systems be based on a Reputation Consensus that implements the Proof-of-Reputation (PoR) principle.
In the consensus mechanism of Proof-of-Reputation (PoR), those who earn a better reputation and a greater long-term audience base get to govern the network. In PoR, the principle behind liquid democracies: reputation is power, rings true.
The principle behind the Proof-of-Reputation consensus mechanism, that reputation is power, opposes the power of brute force (PoW) and the power of money (PoS or DPoS).
Furthermore, such a consensus mechanism will make it possible to implement a system of authentic liquid democracy. The current system of representative democracy implemented in the various societies of the world has proven ineffective and is easily manipulated by those with financial power. Due to the scale and size of the modern world communities, human or artificial, direct democracy cannot be implemented. Liquid democracy, therefore, is the only practical alternative.
Different consensus mechanisms of the modern distributed blockchain systems and the different types of social consensus over the history of human societies.
Computing Reputations
The Reputation Consensus principle states that the governing power of a member of a human or artificial society depends on the reputation of the member as computed on the basis of the following fundamental principles:
The first fundamental principle is that the reputation may be computed through the different measures performed by all the members of a community in respect to the one whose reputation is being computed for, with account to the reputations of all of the other members themselves.
The second fundamental principle is the time scoping of the reputation, so that the reputation measures collected by the member in the past contribute less to the member’s current reputation, as compared with the latest ones — which will be more impactful.
The third fundamental principle is the complete transparency in the reputations of all the members and the measures that they performed. Such transparency will ensure that audits of reputations and the measures performed by the members throughout their history can be analyzed to prevent reputation cheating.
The fourth fundamental principle is the precedence of human measures over those measures that are “artificial,” i.e., are provided by AI. This will ensure that the measures provided by the human participants of hybrid communities have unconditional precedence over the measures provided by the AI systems. The underlying assumption of this principle is that the AI systems will have the capability to contribute to the evaluation of humans and artificial entities in a hybrid community.
In terms of the measures that can contribute to the evaluation of reputation, we will primarily consider the following:
- Members staking financial values on the other members.
- Members providing ratings in respect to the transactions committed with the other members.
- The financial values of the transactions between the members.
- The textual, audio and video reviews made by members in respect to the other members or the transactions between them.
However, various other measures may be considered — depending on the implementation of a given reputation system. Applicability of these measures may not only depend on the accuracy and the reliability that they may provide, but also on the resistance to those attack vectors that target the takeover of the consensus mechanism via reputation cheating and gaming.
What’s Next?
What we want to highlight is that we decided to consider the problem of consensus in a distributed community because the centralized, and some of the decentralized, solutions of AI/AGI systems appear to be targeting the interests of a limited group of powerful and resourceful humans. This makes the likelihood of such entities serving the interests of humanity as a whole to be very unlikely. We believe it is imperative that the governance mechanisms of the AI/AGI systems do not favor the powerful and the rich but are structured to allow for liquid democracy.
The next step that we envision on our path is the creation of an experimental Reputation Agency that will perform reputation calculations upon the needs of any human, artificial or hybrid community. One of the first practical cases for such a reputation calculation will be the SingularityNET ecosystem.
Our next challenge in developing a Proof-of-Reputation consensus mechanism involves formal mathematical modeling and multi-agent simulations — which will allow us to figure out the best set of measures and settings that are the most resistant to the possible consensus takeover attack vectors.
How Can You Get Involved?
If this article piqued your interest, the conceptual design of the Reputation Consensus for reputation systems is described in our earlier Aigents project proposal on Steemit blockchain.
The technical design of the reputation system for SingularityNET is described in this Medium publication. A detailed scientific publication on the reputation system is available over here.
If you would like to learn more about SingularityNET, we have a passionate and talented community which you can connect with by visiting our Community Forum. Feel free to say hello and to introduce yourself here. We are proud of our developers and researchers that are actively publishing their research for the benefit of the community; you can read the research here.
For any additional information, please refer to our roadmaps and subscribe to our newsletter to stay informed about all of our developments.
