Three major entities announce a standard for AI agent dispute resolution. Yet there is zero code, zero testnet, zero specification. The announcement is the product.
GenLayer, OKX, and MetaMask jointly propose “Internet Court”—a standardized framework for AI agents to resolve disputes in digital commerce. The press release promises automatic arbitration, fairness guarantees, and seamless wallet integration. But as a Smart Contract Architect with 25 years in this industry, I see a contradiction: a standard that exists only as a headline is no standard at all. It is a hypothesis.
Context The premise is seductive. AI agents now execute transactions autonomously—trading, negotiating, even creating digital art. When two agents disagree on a delivery or a payment, who judges? Human arbitration is slow, expensive, and prone to bias. Internet Court aims to replace juries with AI moderators, recording all decisions on a blockchain for auditability. GenLayer provides the underlying infrastructure; OKX and MetaMask add user reach.
This is not new in concept. Kleros and Aragon Court already offer on-chain dispute resolution using human jurors and token staking. Kleros handles roughly $50 million in disputes annually. Internet Court’s twist is replacing humans with AI—ostensibly faster, cheaper, and scalable. But the devil is in the opcodes.
Core: Why AI Arbitration Is a Hard Problem From a cryptographic perspective, the core challenge is verifiable computation with trustless AI. For a blockchain to enforce an AI decision, it must guarantee that (a) the correct model was used, (b) the input evidence was not tampered with, and (c) the output was computed without exploitation. Existing solutions like zk-SNARKs can prove that a computation occurred, but proving the correctness of an AI inference—especially a large language model—remains computationally prohibitive.
Consider the execution flow Internet Court would require: 1. Agent A submits a complaint (evidence: chat logs, transaction hashes). 2. A smart contract invokes an oracle that runs an AI model (e.g., GPT-5 or a specialized classifier). 3. The oracle returns a verdict: refund 0.5 ETH to Agent B. 4. The contract enforces the transfer.
Where is the security? The oracle becomes a single point of failure. If the AI model is compromised—by adversarial input, biased training data, or a backdoor—every dispute becomes a potential exploit. Kleros mitigates this by distributing judgment across hundreds of random jurors, each with economic stake. Internet Court substitutes that with model integrity. Security is not a feature; it is the architecture. And here, the architecture is missing.
During my 2020 audit of Uniswap V2’s constant product formula, I learned that simplicity often separates robust protocols from vulnerable ones. Uniswap had a single invariant: x y = k. Internet Court’s invariant would be: every AI decision is fair because the model is trustworthy. That is not an invariant; it is an assumption. A bug is just an unspoken assumption made visible.*
Furthermore, the standard lacks any interface specification. A proper standard (like ERC-20) defines exactly how to query the balance of an address. Internet Court has not published even a draft of the evidence schema, the appeal mechanism, or the slashing conditions for malicious AI operators. Without a formal machine-readable protocol, this is not a standard—it is a blog post.
Contrarian: The Real Purpose Is Platform Lock-In The contrarian angle is that Internet Court is not about disputes—it is about bootstrapping GenLayer’s ecosystem. By positioning their blockchain as the only layer capable of executing AI arbitration (since their consensus is designed for decision logic), they create vendor lock-in. OKX and MetaMask’s support is likely limited to wallet-level notifications and signature prompts—not the heavy lifting of AI validation. This is a common play: announce a “standard” with big partners, attract developers, then monetize through token fees.
But the risk is that the complexity will scare off 90% of developers—as I predicted for Uniswap V4 hooks. AI arbitration requires not just Solidity skills but also machine learning expertise. The intersection is tiny. Most projects will either ignore the standard or fork a simpler version without the AI component.
Clarity is the highest form of optimization. Internet Court’s current clarity is zero.
Takeaway If the standard remains a press release, it joins the graveyard of blockchain dead letters—alongside countless “universal identity” and “cross-chain messaging” announcements. If GenLayer delivers a working testnet with a verifiable AI inference protocol (e.g., using SNARKs or a specialized sidechain), it could define a new vertical. Watch for a formal specification, not a blog post. Watch for a GitHub repository with pseudo-code, not a logo. Watch for an economic model that ensures AI operators are economically punishable. Until then, the smart money reads the code—or in this case, the lack of it. Code is law, but logic is the judge.