Resolving to Truth | An AI Settlement Mechanism for Prediction Markets

Sincere thanks to David Shi (Operator.io) for working with me on this idea.

0. Introduction to Prediction Markets

Origins

Prediction markets have evolved over time both in technical complexity and theoretical possibility. Most prediction markets are resolved by an Optimistic Oracle (OO) protocol, such as UMA Protocol which resolves Polymarket markets. When a market is created, a request for resolution data is sent to the OO. Proposers submit answers, and if undisputed, the data is accepted after a liveness period of about two hours. If disputed, the request is reset and resubmitted. Repeated disputes are escalated and UMA token holders vote on the outcome, resolving disputes within 48-72 hours.

Sources of Truth

1) Information Aggregation

Actors with differing knowledge have a financial incentive to make a forecast based on that information, as it will be profitable for them in the long-run. Different actors will have different information, leading to information aggregation across all the information available. As a market evolves through time, this information can also adapt and change over time, and the market reflects that.

2) Incentivizing Truth

All actors in a prediction market are game-theoretically incentivized to allocate capital to their belief in what is true, based on the information available to them. In this respect, the market incentivizes and reflects what actors know or believe to be true.

More Accurate Than Media

For these reasons, prediction markets can be reasonably trusted to better reflect the true nature of events than any other source of media or knowledge. Traditional sources of media do not have the same incentive or game theoretic alignment to be truthful.

Prediction markets are free from the usual biases of opinion because of their use of market forces. Vitalik Buterin, creator of Ethereum, has heralded prediction markets as a useful social tool that can provide the public with more accurate insight into the likelihood of events than traditional media.

There has even been talk of prediction markets evolving to become more accurate search engines for events.

1. Problems Still Remaining

However, prediction markets in their current design are imperfect in their resolution mechanism, and specifically the resolution dispute system. They are susceptible to two vulnerabilities that impact their function as markets for truth.

1) Economic Manipulation

The potential for economic manipulation attack of the capital volume involved in a prediction market is significant. Due to the nature of UMA governance tokens, one can consider the price of 51% of UMA’s TVL as the minimal viable price for manipulating a prediction market’s outcome. If the TVL of a specific market exceeds the minimum viable price of attack, it will be profitable and feasible to manipulate the resolution of a prediction market.

Take this example scenario: The current price of UMA is $2.76 as of 10/1/24 and the trading volume involved in the US presidential election market on Polymarket is $1.1B. This means, in a perfect world, for an attacker to coordinate a 51% attack on UMA’s protocol and determine the result of the market, it would cost a little over $170 million. In this situation, an attacker could place a maximum bet of $550 million, risking $720 million for a profit of $379 million.

Or this recent occurrence on Polymarket:

In the 2024 “Will Israel invade Lebanon in September” market, more individual voters voted for a “Yes” outcome, but the presence of an abnormally-sized holder of UMA tokens ensured the market resolved to a “No” outcome. This demonstrates the vulnerability of PoS-consensus resolution mechanisms to manipulation by one or more biased individuals who are able to acquire a significant amount of governance tokens.

In order to mitigate this attack threat, it would be effective to remove prediction markets’ dependence on UMA token holders for their dispute resolution. Verifiable AI provides an interesting solution for shifting the responsibility for resolving markets away from the UMA governance token. Verifiable AI can execute comprehension, reasoning and decision-making from within a prediction market’s smart contract. It can provide a way to evaluate evidence and determine a market’s resolution without manipulation, in an autonomous way with the same security as the smart contract underlying each market.

2) Herd Decision-making

Conformity is a fundamental aspect of human psychology. Research such as Asch’s classic line study on conformity demonstrates how pressure from a majority group can lead individuals to conform, despite an obvious correct answer. Humans will have a statistically significant tendency to align with group verdict even in situations where they know the group is incorrect. Proof of Stake Oracle networks are thus susceptible to conformity, especially when they have market resolution discussions collectively in times of market dispute.

Additionally, the incentive structure of this OO resolution mechanism drives tokenholders to seek to submit results that are aligned with the consensus, as this will garner them rewards instead of a slash to their economic stake. However, there is no guarantee that the consensus will be the correct decision, only that if a voter is aligned with consensus, they will receive token rewards. This incentivizes alignment with the group, but not with the absolute truth of a market’s resolution.

In straightforward markets, this may not be an issue. But in a market that has an unclear resolution? Voters are likely to seek to align with the consensus resolution at all costs.

The Venezuelan election prediction market on Polymarket initially showed a strong preference for Maduro, with a 95% probability of his victory based on official results. However, UMA Protocol’s OOs disputed the outcome. Through the resolution dispute process, UMA voters, swayed by claims from the opposition and contested data, ultimately ruled in favor of González. This decision contradicted the Venezuelan electoral commission and global consensus, leading to significant market shifts and heated debates over the reliability of official versus opposition data in the resolution process.

This example leaves several concerns for consistent resolution of prediction markets, including:

• Market manipulation by whales

• PoS dispute resolution creating a herd-mentality which guides market resolution instead of the objective results

• PoS dispute resolution leading to a re-interpretation of the original settlement conditions of a market, rendering that market’s resolution subjective

Together, these forces detract from the objectivity of the resolution process for prediction markets. The advent of smart contracts brought forth trustless enforcement of digitally-prescribed conditions. With the combination of AI and smart contracts which is now possible, we can upgrade the decision-making capabilities of this autonomous code from algorithmic to interpretative and intelligent. The combination of these two technologies provides a way to mitigate the illustrated vulnerabilities of prediction markets, ensuring their resolution is objective and programmatically enforced.

2. Towards Fair, True Outcomes

Can prediction markets act as a more absolute source of truth than even official sources?

The Venezuelan election example was a fascinating controversy, which deserves a revisit. The official verdict of the election was that Maduro won, yet the opposing party claimed election fraud. UMA’s PoS dispute resolution system took it upon itself to make a verdict here that deviated from the official statements from Venezuela.

This is incredibly interesting as it signals how prediction markets truly do have the potential to inform on the true outcome of real events, regardless of manipulation by mainstream sources. No single source of truth can guarantee the result of this election, but perhaps the market aggregation of information is the closest verdict to truth we have.

But are UMA’s voters, and the system itself, optimally designed to make this kind of verdict? Can this system incorporate more objectivity and reduce its possible vectors for manipulation? Perhaps then prediction markets can confidently provide a level of truth that goes above and beyond in impacting the world. As identified earlier, the combination of AI and smart contracts to supplement the market resolution mechanism provides a path forward to achieve this.

It is possible that in the future, a prediction market verdict could be used as verifiable evidence to overturn an election result. It’s a distant reality, but it would be more possible with an upgrade to market resolution architecture.

We suggest that a combined approach of human intelligence, machine intelligence and incentivized consensus systems can provide a system to achieve this kind of outcome for prediction markets.

Trouble For Scaling

To summarize, prediction markets face challenges in:

• Resolving to absolute truth as markets scale

• Expanding to more kinds of events, given the need (and struggles) for objective market outcomes

It will become increasingly important to create solutions that mitigate the aforementioned attack vectors if prediction markets are to successfully scale in in capital involvement, adoption and application of their results.

3. Bringing Objectivity to Subjectivity

Decentralized and verifiable AI, which can act and make decisions autonomously within smart contracts without the risk of centralized manipulation, provides a promising backbone for an AI-driven smart contract resolution mechanism. We shall refer to this as the AI Settlement Oracle in this context.

AI can be integrated in smart contract design with methods such as Zero-knowledge ML (zkML) and Optimistic ML (opML).

Decentralized, verifiable AI can upgrade autonomous smart contract mechanisms with intelligent reasoning and decision-making, providing new opportunities for expanding on the immutability and autonomy of smart contracts to situations with more involved evaluation and decision-making. Until now, smart contracts have been limited in their direct application by the computation capacity of their VM environment. In human or group-led decision-making, it’s near impossible to accurately trace the evidence and reasoning used to come to a decision. In smart contracts, verifiable AI is fully auditable, from its parameters to the inputs used to produce inference. In this regard, any decision made by a verifiable AI can be verified by third parties to ensure trust.

4. Architecture for an AI Settlement Oracle

We propose a combined resolution architecture to realize the benefits of both PoS consensus and smart contract automation. As identified, there are prominent vulnerabilities in the current PoS-based resolution mechanism in the form of

1. Economic Manipulation

2. Herd Decision-making

Our proposed solution serves to upgrade the current resolution mechanism, used by the leading prediction markets, by mitigating these threats to their markets’ resolution objectivity. We believe that objective resolution of prediction markets is critical to their scaling. We suggest building on the game theoretical foundation of current PoS resolution systems, with the implementation of AI specifically in the dispute resolution mechanism.

Polymarket’s Existing Resolution Mechanism

The UmaCtfAdapter plays a central role in managing Polymarket’s binary CTF (Conditional Token Framework) prediction markets, allowing users to propose and dispute results for markets, as part of the resolution process by UMA’s Optimistic Oracle (OO).

Architecture Overview

Markets are initialized with the initialize() function, which sets up key parameters like the question, reward token, and UMA liveness period, while also requesting a ‘price’ from the OO. The ‘price’ in this case refers to the information provided by UMA’s OO users to resolve a market.

The market initializer provides:

• Ancillary Data: The question and its description.

• Reward Token: The ERC-20 token used to pay rewards.

• Reward: The incentive amount for the successful proposer.

• Proposal Bond: A bond posted to ensure good faith in results proposals.

• Liveness: The timeframe in which the OO can accept disputes.

Once initialized, any OO user can propose a result by posting a bond. If another user disagrees, they can escalate the process by disputing the result, moving the issue to a Decentralized Voter Mechanism (DVM) for resolution through which UMA tokenholders vote on the result through a PoS-consensus mechanism.

Resolution Flows

There are several possible flows that dictate how a question is resolved based on the number of proposals and disputes:

1. Business as Usual: The question is proposed and resolved directly by the OO without dispute.

2. Single Challenge Flow: A challenge is made to the first price proposal, leading to a reset and second proposal before ending in resolution.

3. Double Challenge Flow: Multiple challenges escalate the process further, making resolution more rigorous by involving the DVM.

It is this last flow, the Double Challenge Flow, in which the resolution is escalated and which leads to the inherent risks of economic manipulation and herd decision-making. Here we propose upgrading the process with an AI Settlement Oracle.

Protocol Design

We propose maintaining the existing Polymarket resolution mechanism provided by UMA. This system is proven to be efficient for market resolutions in most cases where the resolution is an obvious one. We propose supplementing the DVM with an AI Settlement Oracle.

In the case of the Double Challenge Flow where resolution is escalated to DVM, we suggest a new dispute mechanism.

Tokenholders involved in the PoS dispute resolution should submit verifiable evidence (via zkTLS or other) & their verdict on the true market outcome to the protocol. The protocol collects evidence & verdicts from all participating tokenholders and submits this information with a query for accurate resolution to ORA Protocol’s AI Oracle network.

ORA’s AI Oracle network provides verifiable, decentralized AI inference from any model. Current leading foundational models such as Llama3 can be called in a trustless way using ORA’s network. To achieve the best integration of decentralized AI in this resolution system, we recommend using the fine-tuned version of a powerful open-source model like Llama3 to ensure it is most equipped for reasoning and decision-making in this context.

The specified LLM will evaluate all the verifiable evidence alongside tokenholders’ own personal verdicts. It will come to its own decision based on this aggregated information, and use that decision as the final market resolution. Tokenholders who submitted a verdict aligned with the AI’s decision will be rewarded and those who submitted an opposite verdict will be slashed.

The PoS component of this mechanism serves the purpose of collecting information. The verdicts of each individual tokenholder should be made based on the evidence available to them. The verifiable AI component of this mechanism serves the purpose of mitigating the possibility for economic manipulation and herd decision-making for the resolution of controversial, high profile or high volume markets. Here, the verifiable AI component can act as a final, objective, autonomous decision-maker for the resolution process, eliminating the possibility for a 51% attack and inaccurate herd decision-making.

New DVM Resolution Architecture with AI Settlement Oracle

In summary, the proposed architecture has the following components:

• OO Submission System — optimistic oracle-based submission system for usual resolution of markets.

• PoS Tokenholder Dispute System (DVM-equivalent) — stake and slash system for tokenholders to vote on market resolutions escalated to the dispute process.

• Collection of Verifiable Evidence — tokenholder participants must collect verifiable evidence to support the resolution of each market. zkTLS provides a viable option for verifiable offchain data.

• Submission of Evidence and Resolution Verdict — tokenholder participants must submit their verifiable evidence and a resolution decision for the market outcome (yes or no) in a dispute scenario.

• Query to AI Oracle — protocol smart contracts will send a query with all tokenholders’ submissions (evidence + verdict) to ORA’s AI Oracle: a decentralized and verifiable AI oracle network running on opML.

• AI Settlement Oracle — final dispute resolution decision will be made based on the submitted evidence by a verifiable AI decision-maker, mitigating present vulnerabilities of economic manipulation and herd decision-making.

The intention here is to reduce security risk, and provide an architecture that better allows for scale and achieving consistent market resolution to absolute truth by reducing the shortcomings of PoS consensus resolution systems.

5. Discussion

There are several obvious strengths and weaknesses to this protocol design which are worth discussing.

Strengths

• Mitigation of economic manipulation

The proposed integration of a decentralized AI Oracle eliminates the possibility of economic manipulation in the final resolution outcome by adding a final decision-making layer beyond the PoS consensus voting mechanism, as governed by protocol tokens. This ensures that the process remains free from malicious attacks which would influence the result unfairly. However, it's worth noting that while the resolution outcome is protected, economic manipulation during the live market is still possible.

• Autonomy, transparency, objectivity.

By relying on decentralized AI autonomously operating within smart contracts, we introduce transparency and eliminate subjective biases from decision-making. The AI’s verifiable nature guarantees that its inferences and resolution decisions are objectively derived from analysis of the submitted evidence.

• Balanced human and AI decision-making.

A unique advantage is the balanced implementation, where humans resolve typical markets, and the AI steps in only for disputed markets. This setup reduces incentives for disputes, as the AI, with no inherent bias, handles them objectively when human decision-making stalls.

• Neutral decision-maker in dispute markets.

When human judgment reaches an impasse, the AI, acting as a neutral, autonomous entity, can make decisions independently, avoiding potential biases seen in PoS consensus mechanisms that might result from groupthink or concentrated influence.

• Continued tokenholder involvement.

Tokenholders remain incentivized to submit their evidence and verdicts, even as the AI plays a decisive role in dispute scenarios. This ensures ongoing engagement and participation in the market resolution process.

Weaknesses

• Bias in AI models.

Models are trained on specific datasets, which may harbor inherent biases. While foundational models like Llama3 are robust and trained on vast datasets, they are still susceptible to implicit biases based on the nature of the training data which we may not yet be aware of.

• AI misjudgment risks.

Although AI provides a solution to mitigate human error, it introduces its own vulnerabilities, such as hallucinations or errors in unprecedented situations not covered by its training data. This could result in inaccurate or unexpected market resolutions.

• Tokenholder data breadth.

There does exist a potential attack vector if tokenholders collude to submit the same data, which could compromise the AI’s ability to accurately process the breadth of evidence that directs an accurate resolution. This protocol design could benefit from a mechanism to incentivize the submission of varied data.

• Adoption challenges.

Gaining trust and proving the efficacy of this AI-integrated system will be critical for driving adoption. Demonstrating the reliability of the AI Oracle in dispute resolution will be essential to building confidence in its decision-making capabilities.

Accountability and Protocol Design

A resounding consideration to make is about accountability. An AI model cannot reasonably be held accountable for unintended outcomes. This is an important issue to consider and should guide the specific implementation of the AI in a proposed resolution mechanism design.

In the specified protocol design, the AI Settlement Oracle acts as a final stage of decision-making for a market’s resolution, if it is sufficiently disputed. The addition of this AI serves to mitigate the vulnerabilities of the existing resolution process (economic manipulation and biased human decision-making), but what happens if the AI gives an incorrect resolution?

Another possible protocol design could place the AI Settlement Oracle as the first mechanism for resolving disputes, instead of the last. Once a market’s resolution is disputed to be passed to the DVM it could be submitted to the AI Settlement Oracle immediately for a decision. The market resolves if the PoS voters don’t dispute the AI Settlement Oracle’s decision. If PoS voters dispute the result, the resolution moves for a final decision by PoS consensus. In this situation, the addition of this AI serves to improve the efficiency of the dispute resolution process. It may not achieve the mitigations of the identified problems with market resolution, but it prevents a situation in which AI is culpable for an incorrect market resolution.

It will be important to find the optimal blend of representation and responsibility between the AI Settlement Oracle and PoS system in order to mitigate the flaws of both most effectively.

6. Conclusion

To conclude, we expect this combined resolution architecture will realize immediate benefits for prediction markets.

It will:

• Reduce dispute controversies

• Reduce the possibility for economic manipulation of market resolution

• Strengthen the objectivity of market resolutions

• Provide a more robust architecture to scale value and markets

Together, this will improve the security, scalability and truth of prediction markets, thus driving their evolution to new markets, new users and new applications of their data.

It was relevant to begin this article with a reminder of why prediction markets are interesting: their architecture is intended to render them superior arbiters of truth than any other media. It is this truth that is so important.

As prediction markets scale through time and space, it will eventually become possible to use their market resolution data for prediction models themselves, once enough outcomes are recorded. In order for these models to be accurate it is critical that prediction market data consists of objective, true market resolutions. An AI model that is trained on this data would be incredibly powerful, leveraging the already-established value of prediction market data.

Therefore, solutions that work towards consistently objective, true market resolutions both has value in strengthening the security and scalability of existing prediction markets, but also has incredibly powerful downstream implications.

About ORA

ORA is Ethereum’s trustless AI that makes blockchains AI native.

ORA breaks down the limitations of smart contracts by offering richer data sources and arbitrary compute so developers can innovate freely.

ORA’s work has been trusted by Compound, Ethereum Foundation, Uniswap, Optimism, Arbitrum, Polygon, Manta, and beyond.

www.ora.io | x.com/oraprotocol | discord.gg/ora-io