Artificial intelligence can serve as an embedded judge in prediction markets, according to Andrew Hall, a professor of political economy at Stanford University's Graduate School of Business.

https://t.co/wQfka1JDA6

— a16z crypto (@a16zcrypto) January 22, 2026

He illustrated the issue of "fair" dispute resolution using the example of the presidential elections in Venezuela.

Last year, contracts worth over $6 million were placed on the event's outcome. However, after the campaign, the market was left in confusion:

  • The government declared Nicolás Maduro the winner;
  • The opposition and international observers reported fraud.

"Should the resolution of contracts in the prediction market follow the 'official' information (Maduro's victory) or the 'consensus of credible reports' (opposition's victory)?" Hall questioned.

This is not an isolated case, the expert noted. In another instance, someone allegedly manipulated a map of Ukraine regarding a territorial dispute.

Hall emphasizes the need to create a fair contract resolution system that people can trust. In such a case, prices would become significant signals for society.

Not Just a Prediction Market Problem

Similar issues plague financial markets. The International Swaps and Derivatives Association has been grappling with resolution problems in the credit default swap market for years—contracts that pay out in the event of a company's or country's bankruptcy.

Decision-making committees vote on whether credit events have occurred. However, the process has been criticized for its lack of transparency, potential conflicts of interest, and inconsistent outcomes.

"The fundamental problem remains the same: when large sums depend on determining what happened in an ambiguous situation, any resolution mechanism becomes a target for manipulation, and ambiguity becomes a potential point of debate," Hall stated.

Characteristics of a Good Solution

The expert outlined several key characteristics that any viable solution must possess:

  • Resistance to manipulation—if the verdict can be influenced by editing Wikipedia, spreading fake news, bribing oracles, or exploiting loopholes, the market turns into a game where the best manipulator wins;
  • Reasonable accuracy—the mechanism must make the correct decision in most cases. While perfect accuracy is impossible, it is crucial to eliminate systematic errors and blatant misses;
  • Transparency—traders must clearly understand how the mechanism operates;
  • Neutrality—participants should be assured that the system does not favor any specific user or outcome.

Human committees can meet some of these properties, but they are susceptible to manipulation and cannot maintain neutrality.

AI as a Solution

Hall proposes using large language models (LLMs) as judges, with each model and prompt recorded on the blockchain at the time of contract creation.

The basic architecture looks like this:

  1. When creating a contract, the market maker specifies not only the criteria for dispute resolution in natural language but also the LLM and the exact prompt that will be used to determine the outcome.
  2. The specification is recorded on the blockchain using cryptography.
  3. When trading begins, participants can review the entire mechanism of the contract—they know exactly how the model accesses the specified sources of information and makes a decision.

This method addresses several key issues:

  • AI resists manipulation (though not absolutely). The results of a large LLM are not easily edited. To alter a decision, a fraudster would need to change the information sources the model relies on;
  • Accuracy is ensured—neural networks can quickly navigate the web and seek new information;
  • Transparency—the entire dispute resolution mechanism is available for analysis and verification. There can be no changes to the rules during the process or subjective decisions;
  • Reliability—LLMs have no financial interest in the outcome and cannot be bribed.

However, there are drawbacks: AI can make mistakes. It might misinterpret a news article or fabricate a fact.

Manipulations are not impossible, but they are harder to execute. Fraudsters could pay to place specific information in major media outlets. This is expensive but feasible.

There is also a risk of attacking the LLM's training data. However, this would require action long before the contract is signed.

Conclusion

AI-based solutions replace one set of problems with another, more manageable set. Platforms should experiment with different LLMs to gain experience, Hall believes.


As best practices emerge, the community needs to work on standardizing combinations of AI programs. This will help concentrate liquidity, according to the author.

Recall that in January, analysts at a16z crypto predicted growth in prediction markets and ZK proofs.