📰 Key Takeaways

GenLayer is an “internet court” that settles disputes with a jury of AIs. Co-founder and CEO Albert Castellana says it doesn’t rely on a single, possibly biased AI. Instead, it draws on up to 1,500 different AIs, and he believes left-leaning and right-leaning models can cancel out each other’s bias. When a dispute comes up, the system randomly picks 5 validators to form a jury. Each validator runs a different AI model, and a simple majority vote produces a first ruling within minutes. If someone objects, the case can be appealed to 11 validators, then 23, with each round widening all the way up to the system’s cap of 1,500. Castellana says final resolution usually comes within 30 minutes, while going through every appeal round can take about three hours. The open coalition now has 35 members, including ZKsync, NEAR, Consensys, and UMA, which handles disputes for Polymarket. GenLayer has been in development for nearly four years. It started as an attempt to build a “decentralized Wikipedia” to fix the problem of contested topics getting locked down by a small group of editors. Its “smart contracts” can make subjective judgments, like automatically deciding whether a freelancer’s content is up to standard before paying them, and some platforms already use it to pay influencers who create content. That said, Castellana admits the technology still can’t reliably tell whether an image is a deepfake, or settle a boundary dispute between neighbors based on photos. For the other use cases, see the original article linked below.

💬 JudyAI Lab Take

GenLayer hands dispute resolution to a jury of AIs. What’s worth noticing is that it doesn’t bet on a single model. It lets up to 1,500 different AIs check and balance each other, and it already has 35 members in its open coalition, including ZKsync, NEAR, Consensys, and UMA, which handles disputes for Polymarket.

For AI builders, the design thinking here is this: instead of chasing one perfect “bias-free” model, accept that every model has a stance, then use diversity to cancel it out. The original article says left-leaning and right-leaning models can offset each other’s bias. The system first randomly picks 5 validators for a majority-vote initial ruling, and if there’s an objection, it widens round by round to 11, then 23, up to a maximum of 1,500. Most cases are settled within 30 minutes, and running through every appeal round takes about three hours. This “fast first, careful later” tiered cost design means simple disputes get resolved cheaply, and only the hard ones pull in more resources. At the same time, Castellana openly admits it still can’t reliably detect deepfake images, which clearly marks the edge of what it can do.

If your product also relies on AI judgment, try running the same question through 2 or 3 different models and see where their answers diverge. Then decide which situations need a human to double-check.

📅 Original Article Info

🔗 Further Reading