Question 1

How accurate is the detection system?

Accepted Answer

On the MGTD benchmark (the largest public AI detection benchmark with ~2M samples across 15 datasets), Subnet 32 achieves 92% ROC-AUC. On the RAID benchmark it reached 98.3% accuracy, and on GRiD, HC3, and GhostBuster datasets it averages 99.1% accuracy with 96% F1-score. Performance varies by text type and language, but the system is designed to improve continuously as miners refine their approaches.

Question 2

What types of text can it detect?

Accepted Answer

The system validates on a diverse range of sources including academic writing, web-crawled content, financial documents, code, and medical/legal text. It supports English and has demonstrated 98.7% accuracy on Arabic-language benchmarks. Validators continuously add new text types and generation parameters to prevent overfitting to specific writing styles.

Question 3

Can miners see validator queries in advance?

Accepted Answer

No. Validators generate test data on-the-fly using the Pile dataset and language models with randomized generation parameters and augmentations (misspellings, word removal). This prevents miners from memorizing datasets or using shortcuts like hash lookups. Predictions are encrypted before transmission so only the validator can decrypt and score them.

Question 4

What hardware is required to run a detection node?

Accepted Answer

Miners require a GPU with at least 24GB VRAM (e.g., RTX 4090), 24GB RAM, 4 CPU cores at 2.5 GHz, and 20GB SSD storage. Validators need 80GB GPU VRAM (e.g., A100), 80GB RAM, 1TB SSD, and must run 30+ language models via Ollama. See min_compute.yml for full specifications.

Question 5

How is a miner rewarded if the network is decentralized?

Accepted Answer

Miners earn rewards based on their detection accuracy as measured by validators. Validators compute three metrics—F1 score, false positive rate, and average precision—and reward the miners that score highest. Rewards are proportional to the miner's weight in the network, which updates each epoch based on performance history. Better detection models accumulate more weight and earn more.

Question 6

What happens if a miner tries to game the system?

Accepted Answer

The validation process includes multiple safeguards: augmented and adversarial text variants prevent memorization; encrypted predictions ensure validators can't be fooled by seeing the query source; randomized LLM generation parameters prevent pattern exploitation; and continuous benchmarking against academic datasets (RAID, GRiD, HC3) catches models that overfit to the validation set.

ItsAI

ItsAI

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup