As AI agents take on tasks such as trading, shopping and managing transactions, misunderstandings between bots and their users could become increasingly common. Internet Court is designed to handle these disputes through contracts that keep funds in escrow until an AI-based judging process determines whether the terms were fulfilled.
The system relies on GenLayer, an open delegated proof-of-stake network whose validators run different AI models. A dispute is initially sent to a randomly selected jury of five validators, with a decision typically reached within minutes by simple majority vote. Participants can appeal to larger juries of 11, 23 and eventually more agents, with finality generally taking about 30 minutes. Reaching the system’s maximum level of escalation, involving up to 1,500 agents, could take around three hours.
Internet Court’s juries can include models from the Claude, Gemini, ChatGPT and DeepSeek families, among others. GenLayer co-founder and CEO Albert Castellana argues that using multiple models can reduce the influence of any single system’s biases. The protocol’s open consortium currently has 35 members, including ZKsync, NEAR, Consensys and UMA.
GenLayer has operated in beta for about a year and currently processes approximately 12,000 decisions per day. It has authenticated user-uploaded images for Collective Memory and settled disputes on argue.fun. Its mainnet is expected to launch next month.
The project began nearly four years ago with an attempt to create a decentralized way to evaluate information for an online knowledge base. It later developed the concept of “intelligent contracts,” which can interpret code, natural language, internet data and images, while also making subjective judgments.
One example is an agreement with a freelance writer that automatically evaluates submitted work before releasing payment. GenLayer is already used by Collective Memory to assess the quality and accuracy of uploaded photographs and determine payments. Argue.fun uses the technology to settle debates on subjects including GenLayer governance, decentralized autonomous organizations and AI safety. Its top four users have reportedly earned more than $2 million each.
The technology has clear limitations. Castellana says it is not yet reliable enough to judge whether an image is a deepfake or to settle a property-boundary dispute based on photographs. AI systems can also reproduce biases in their training data, while research on “Representational Convergence” raises the possibility that models from different companies could eventually develop similar views, creating a form of groupthink.
A potential short-term application is prediction-market arbitration. Castellana believes GenLayer could resolve ambiguous market questions in minutes using hundreds or thousands of AI votes, rather than relying on lengthy human appeals. However, Internet Court does not prevent further challenges: users could still take a disputed decision to systems such as Kleros or UMA, or ultimately to a human court.
Castellana expects the process could eventually support more autonomous DAOs, allowing them to assess proposals against their constitutions and broader objectives. Whether users will accept AI-generated judgments remains an open question, particularly as the systems continue to develop.
Sources: Cointelegraph.