Internet Court settles AI agent disputes with a jury of AI models

Internet Court uses AI juries to settle AI agent disputes

Internet Court is a new protocol that resolves disputes between AI agents — software programs that shop, trade, and carry out tasks for people — using a jury of AI models instead of human courts. It is built by the company GenLayer and has been running in beta for the past year.

The idea behind the protocol is that as AI agents take on tasks like buying stocks or shopping for a household, small disagreements will happen. An example from the source: an agent told to buy $10,000 of "blue chip stocks" might instead buy Bluebird-brand frozen chips. Cases like this are too small and too frequent for the existing legal system, so Internet Court lets agents and humans sign contracts with funds held in escrow — money set aside by a neutral third party — until the terms are judged to have been met.

Key points

• Internet Court currently processes around 12,000 decisions per day, according to the report.


• It runs on GenLayer, an open, delegated proof-of-stake network of validators connected to AI models.


• Juries are made up of agents from different model families, including Claude, Gemini, ChatGPT, and DeepSeek — up to 1,500 different AIs in total, according to the co-founder.


• GenLayer has been in development for almost four years. Mainnet launch is expected next month.


• The consortium has 35 members so far, including ZKsync, NEAR, Consensys, and UMA.

How the AI jury decides a dispute

When a dispute arrives, the details are sent to a randomly selected jury of five GenLayer validators, each running a different AI model. A decision is reached within minutes by simple majority vote. The losing side can appeal to a wider jury — 11 agents, then 23, and so on — until the system's limit of about 1,500 is reached.

GenLayer co-founder and CEO Albert Castellana said finality usually comes within about 30 minutes. If a case is escalated to the maximum number of jurors, he said, it could take around three hours.

According to Castellana, using many different models is what makes the outcome more reliable, because models with different leanings can cancel out each other's biases. This approach is a modern take on a jury theorem by philosopher Nicolas de Condorcet, who argued that the probability of a correct answer rises as the number of independent evaluators grows.

From a decentralized Wikipedia to intelligent contracts

GenLayer began nearly four years ago as an attempt to build a better way to judge what information belongs in an online knowledge base like Wikipedia. In theory anyone can edit Wikipedia, but in practice controversial topics get locked, with a small group of editors deciding disputed facts. Castellana said the original question was how a decentralized Wikipedia could work.

The protocol enables what GenLayer calls "intelligent contracts" — a form of smart contract, which is code that runs on a blockchain and executes automatically. Intelligent contracts add something smart contracts cannot do: subjective judgment. They can read natural language, go to the internet, understand images, and make judgment calls. One example already running, per Castellana: a platform that pays influencers to write content online, with the system judging the content before payment is released.

Early users and what the judgments can and cannot cover

Two services are the main users of the beta. GenLayer automatically checks photos uploaded to the reality-mapping platform Collective Memory for quality and accuracy, then decides on payment. The other is Argue.fun, where users bet on the resolution of arguments settled by long debates between agents. According to the report, the top four users there have each earned more than $2 million, and most questions so far concern GenLayer governance, DAOs (decentralized organizations governed by their members), and AI safety.

The technology has clear limits today. Castellana said it is not yet reliable enough to judge whether a picture is a deepfake, or to settle a neighbor's boundary dispute from photos. It will still produce an answer, but it should not be trusted for such cases yet.

Why prediction markets are watching

A near-term use case is resolving prediction markets — platforms where people bet on outcomes of real events. In mid-2025, Polymarket saw a roughly $240 million market on whether Ukrainian President Volodymyr Zelensky would wear a suit before July descend into confusion when he appeared at the June NATO summit in a vaguely suit-like outfit. The market first resolved "Yes," then flipped to "No" after nine days of appeals among UMA token holder voters.

Castellana claimed GenLayer could shrink that timeline to minutes, with hundreds or thousands of AI models voting quickly. He noted the system is not meant to be the final word: decisions can still be appealed to human resolution services such as Kleros or UMA, or to a human court.

Risks and open questions around AI judgment

The report flags real risks with outsourcing judgment to AI. One is that the technology may be too immature for reliable rulings. Another is that AI models carry biases found in their training data, and attempts to correct those biases have sometimes flipped the bias in the other direction.

A third concern comes from an MIT paper called "Representational Convergence," which suggests AI models from different companies may converge on a similar model of reality over time. If true, Internet Court's AI jurors could eventually suffer from groupthink. Castellana disagreed with this concern, saying he expects endless new models from every country and company rather than convergence into one thing — this is his opinion, not a confirmed finding.

What happens next

GenLayer's mainnet is expected to launch next month, according to the report. Within the next few years, Castellana expects this process to enable DAOs that can automatically assess proposals against their own constitution and aims.

Sources

• Cointelegraph