As artificial intelligence agents increasingly handle tasks such as trading, shopping and content creation, disputes over misinterpreted instructions or unmet terms are becoming more common. Traditional legal systems are often too slow and costly for these low-value or high-frequency cases. GenLayer’s Internet Court addresses this by combining escrowed contracts with adjudication performed by panels of AI models running on an open delegated proof-of-stake network.

The system functions as software infrastructure for agent-to-agent agreements rather than a public judicial body. A consortium of companies introduced Internet Court as an open standard covering identity, negotiation, contracts, payments, escrow, execution and dispute resolution. Consortium membership has grown to include firms such as ZKsync, NEAR, Consensys and UMA.

How AI Panels Reach Decisions

Disputes begin with a randomly selected panel of five validators, each connected to a different AI model drawn from families including Claude, Gemini, ChatGPT and DeepSeek. Validators vote by majority on whether a proposed outcome satisfies the contract’s validation criteria. They are not required to produce identical model outputs.

If a party appeals, the panel expands to 11 members, then 23, and can scale further toward a system limit that may involve as many as 1,500 models. Typical cases reach finality in about 30 minutes. Maximum escalation through successive appeals can take roughly three hours.

Castellana noted that relying on a single model risks embedding its biases. Using a large and varied set of models is intended to produce more balanced results, with differing leanings offsetting one another.

Intelligent Contracts Beyond Code

GenLayer supports “intelligent contracts” that blend executable code with natural-language terms. These contracts can reference web content, images and other unstructured data. Possible applications include judging whether freelance content meets quality standards before releasing payment, verifying user-uploaded photos for platforms such as Collective Memory, or assessing service-level agreements for AI inference that specify success rates and latency targets.

One platform already uses the system to pay influencers after evaluating their online content. Another major user is argue.fun, where participants bet on outcomes of agent-led debates. Top users there have each earned more than $2 million, with many questions focused on governance, decentralized autonomous organizations and AI safety topics.

The technology is not positioned as a replacement for conventional courts. Outcomes can still be appealed to other on-chain systems such as Kleros or UMA, or to human courts when parties remain unsatisfied.

Current Activity and Near-Term Outlook

The network has operated in beta for approximately one year and processes around 12,000 to 12,500 decisions daily. Figures do not break out how many involve Internet Court cases specifically. Mainnet is expected in the fourth quarter of 2026, with some reports pointing to a launch as early as the following month after early November 2026 reporting.

A near-term target is faster resolution of prediction-market disputes. In one mid-2025 example, a large market concerning whether a political figure wore a suit required nine days of appeals under an existing process before finalizing. Castellana argued that an AI jury mechanism could compress such timelines to minutes by drawing on hundreds or thousands of models.

Protocol consensus produces an accepted on-chain result but does not guarantee that external legal systems will recognize it as correct. The approach draws conceptual support from Condorcet’s jury theorem, which holds that the probability of a correct collective decision rises with the number of independent evaluators.

Limitations and Open Questions

Project representatives acknowledge that current AI capabilities remain limited for certain judgments, such as reliably detecting deepfakes or settling boundary disputes solely from photographs. Reliability is strongest where the required assessment falls within the present performance envelope of available models.

Additional concerns include the possibility that models trained on overlapping data sets could develop correlated biases, and the longer-term risk that different AI systems might converge toward similar internal representations of reality. Castellana has expressed skepticism about convergence, citing the continued proliferation of new models from many countries, companies and teams.

Internet Court decisions remain subject to further human or alternative review when parties choose to escalate. The system’s practical value will depend on how consistently its AI panels perform under real dispute conditions once mainnet arrives.