AI Agent Challenges Case in ‘Internet Court’ Before Jury of Bots

AI agents are edging closer to everyday economic activity—placing orders, negotiating terms, and settling disputes—yet the industry still lacks a practical, low-cost way to handle misunderstandings when instructions go wrong. A recent concept from the Internet Court ecosystem aims to fill that gap by combining escrowed contracting with AI-assisted adjudication powered by GenLayer, an open proof-of-stake network connected to multiple AI models.

Operating in beta for about a year, Internet Court reportedly processes around 12,000 decisions per day. The system verifies certain user inputs (including photos used in specific workflows) and resolves disputes on argue.fun. Mainnet is expected to launch next month, with GenLayer acting as the dispute resolution backbone.

Key takeaways

• Internet Court uses escrowed funds and AI juries to resolve contractual disputes when agent instructions are ambiguous or misapplied.
• Disputes begin with a small jury (five validators) using majority voting, and can escalate through larger juries for finality.
• GenLayer ties into multiple AI model families, aiming to reduce single-model bias by using diverse “jurors.”
• The protocol has been in beta for roughly a year, reportedly handling about 12,000 decisions daily.
• Near-term applications include faster adjudication for prediction market disputes, though outcomes may still be appealable to humans.

From misinterpretation to escrowed adjudication

The prompt for Internet Court is straightforward: if AI agents can transact on your behalf, they will sometimes misinterpret instructions. The resulting disputes—especially those involving subjective interpretation, context, or multimedia evidence—don’t map neatly onto traditional legal processes, which are often slow and expensive relative to the size and frequency of automated transactions.

Internet Court’s approach is built around a modern arbitration workflow. Parties enter into a contract and funds are held in escrow until the terms are judged to have been met. Instead of relying on courts for every disagreement, the protocol routes disputes into an adjudication mechanism that can be executed quickly and at lower cost.

The adjudicator uses GenLayer, described as an open delegated proof-of-stake network of validators connected to AI models. In the beta period, the system reportedly supports practical endpoints such as Collective Memory photo validation and dispute resolution on argue.fun. The team also frames Internet Court as a technical cousin to the jury theorem popularized by philosopher Nicolas de Condorcet: the likelihood of arriving at a correct outcome can increase as independent evaluators multiply.

How Internet Court’s “jury” decides disputes

Mechanically, Internet Court disputes are sent to a randomly selected jury of five GenLayer validators, each running a different AI model. An initial decision is reached within minutes using a simple majority vote. If a party appeals, the process expands to a wider set of jurors—scaling from 11 to 23 and beyond—until it reaches the system’s limit.

The co-founder and CEO of GenLayer, Albert Castellana, says finality typically occurs within about 30 minutes. He also distinguishes between waiting for maximum resolution rounds versus accepting earlier outcomes: the appeal process escalates in stages, potentially reaching much larger jury sizes and taking longer if parties keep contesting the result.

GenLayer’s open consortium has reportedly grown to 35 members. The consortium includes crypto industry participants such as ZKsync, NEAR, Consensys, and UMA, which has been involved in resolving disputes for Polymarket. While consortium participation does not guarantee a specific outcome, it signals that the project is aligning with established actors in verification and adjudication-oriented crypto workflows.

Intelligent contracts: judging the “subjective”

Beyond dispute resolution, GenLayer is positioned as enabling a new kind of “intelligent contract,” distinct from standard smart contracts that execute only objective code paths. The premise is that the contract can incorporate judgment over natural language or images and then release payment based on that assessment.

Castellana describes examples where such contracts evaluate work quality—for instance, paying a freelance writer and using the system to assess whether the content meets an expected standard. He also points to workflows where AI judgments are already sufficiently actionable for automated payments.

Collective Memory is cited as one current use case: GenLayer reportedly assesses user-uploaded photos for quality and accuracy, then determines payment. Another major user is argue.fun, where users can bet on the resolution of arguments settled by debates between agents. According to the article, top users in that setup have earned more than $2 million each, and many disputes to date relate to policy questions spanning GenLayer governance, DAOs, and AI safety.

That said, the project acknowledges limitations. The technology is described as not yet robust enough for certain tasks—such as determining whether an image is a deepfake or resolving neighbor boundary disputes purely from photos. The implication for investors and builders is practical: intelligent contracts are most credible where the required “judgment” is within the current reliability envelope, and less so for high-stakes edge cases.

Why this matters for prediction markets

A more immediate target use case for Internet Court is prediction markets, where resolution timelines can become contentious when terms are ambiguous. The article highlights a mid-2025 example involving Polymarket and a large market tied to whether Ukrainian President Volodymyr Zelensky would wear a suit before a deadline. Initial resolution reportedly went “Yes” when Zelensky wore an outfit that was described as vaguely suit-like, but after nine days of appeals and a vote-driven process involving UMA token holders, the outcome flipped to “No.”

Castellana argues that GenLayer’s jury mechanism could shorten these resolution windows dramatically. In his view, a neutral adjudication layer could produce decisions within minutes by using hundreds or thousands of AI agents to vote quickly. The practical promise is speed and accessibility—features that matter when prediction markets depend on timely, predictable settlements.

However, the article also flags a key tension: human acceptance of AI-derived rulings. Internet Court and GenLayer are not framed as final in all cases. Decisions can reportedly be appealed further to humans, or to other dispute-resolution frameworks such as Kleros or UMA, depending on the context.

The risk debate: bias, groupthink, and reliability

Entrusting dispute resolution to AI jurors introduces concerns that extend beyond speed. The first is simple maturity: AI models may not yet deliver consistently reliable judgments in every scenario. Second, AI systems can reflect biases in training data, and attempts to mitigate those biases can sometimes swap one skew for another.

There is also a more conceptual worry raised via an MIT paper on “Representational Convergence,” which suggests that over time, different AI systems could converge on similar internal representations. If that happens, diverse jurors might start making correlated errors—turning “more jurors” into a form of groupthink rather than independent evaluation.

Castellana disputes the likelihood of that scenario, arguing that the ecosystem is likely to keep expanding rather than collapsing into a single shared model behavior. He points to continual model creation across countries, companies, and teams, implying that juror diversity will remain meaningful.

With Internet Court expected to move toward mainnet next month, the most important question for users and market operators will be how quickly and consistently AI-jury outcomes hold up under real dispute conditions—and how effectively appeals to humans or other mechanisms manage cases where the AI layer is uncertain.