Subnet 32: ItsAI
ItsAI is Bittensor Subnet 32. Its on-chain identity registers the subnet as focused on high-quality AI detection for text — classifying whether a piece of text is human-written or AI-generated.
What ItsAI Provides
The registered identity describes ItsAI as a subnet focused on high-quality AI detection for text. Miners run detection models that judge whether text is human-written or machine-generated, and the network rewards the most accurate classifiers. The registered project website is its-ai.org and the codebase is the It-s-AI/llm-detection repository, which defines the detection task in more detail than the on-chain identity alone.
Detection Context
The ItsAI website presents the project as an AI detector and ChatGPT checker. At the subnet level, that means the useful work is classification: given a piece of text, the network needs a judgment about whether it is human-written or generated by a language model. The website also describes deeper scan results, which fits ItsAI’s focus on more than a single yes-or-no product label.
This task is sensitive because false accusations are costly. A detector that marks human writing as AI-generated can harm a student, author, or professional user even if the detector performs well on average. ItsAI’s public incentive documentation reflects that concern by treating false positives as a separate part of reward quality rather than relying only on raw accuracy.
Evaluation Context
The ItsAI incentive documentation describes validation around two kinds of text: human-written samples and AI-generated samples. The human side uses existing language data, while the generated side is built by prompting language models to complete text from the same source context. That design makes the classification problem about authorship signal rather than obvious topic differences between human and generated examples.
The same documentation also describes data augmentation intended to reduce cheating and overfitting. Validators can alter samples in small ways, such as selecting sentence spans or adding minor text changes, so miners cannot simply memorize known source text. For readers, the important point is that ItsAI rewards detectors that generalize to fresh-looking text rather than detectors that only recognize a fixed validation set.
ItsAI’s reward framing combines classification quality, false-positive control, and ranking quality. That keeps the subnet aligned with the practical goal of AI detection: identify generated text while preserving confidence that real human writing will not be mislabeled. This framing explains why the source treats false positives as a first-order quality concern.
Validation Data Context
The ItsAI incentive documentation describes validation data as balanced between human-written samples and AI-generated samples. That balance matters because the subnet is not only asking whether a detector can recognize generated text. It is also asking whether the detector can preserve human-authored text as human when both classes appear in the same evaluation process.
For the human side, the incentive documentation names the Pile as the source of validation text. For the generated side, it describes creating prompts from the same source context and asking language models to complete them. This makes the comparison more useful than mixing unrelated human and AI datasets: the intended difference between the two classes is authorship, not topic, formatting, or source domain.
The same source says validators vary generation parameters and use many model sources when creating AI-generated samples. For readers, the important point is diversity of generated text. A miner that only recognizes one generator’s style should be less useful than a detector that handles a wider range of generated completions.
ItsAI also uses small data augmentations on both human-written and AI-generated samples. The incentive documentation gives examples such as selecting sentence spans and applying minor text changes. This is a validation-quality detail rather than a user-facing product feature: it helps reduce memorization and overfitting so miners are rewarded for detection behavior instead of lookup against a known corpus.
The reward components follow that same boundary. F1 score captures binary classification quality, false-positive scoring protects human-written samples from being wrongly flagged, and average precision evaluates ranking quality across thresholds. Together, those metrics describe a detector that should be accurate, careful with human text, and useful when scores are interpreted as rankings rather than only as yes-or-no labels.
References: ItsAI incentive documentation
Miner and Validator Roles
Subnet 32 operates under the standard Bittensor two-role structure. Miners supply a capability — here text-detection judgments — and validators evaluate those contributions and set weights. Reward distribution follows Yuma Consensus.
On-Chain Identity
The live Finney identity for netuid 32 registers the subnet name as ItsAI. The registered project website is its-ai.org and the GitHub repository is It-s-AI/llm-detection; a Discord channel is also recorded in the identity. Live subnet data is available on TaoStats.
Relationship to Yuma Consensus
Subnet 32 uses Yuma Consensus to convert the detection-quality weight vectors that validators submit into the emission shares distributed to miners and validators within the subnet each tempo. The linked documentation describes how validator weight submissions are aggregated into consensus weights for each miner registered on the subnet.
In ItsAI’s context, validators score miner text-detection outputs using F1 classification quality, false-positive control over human-written samples, and ranking precision, then translate those combined scores into weight vectors for the subnet. The Emission documentation describes how those consensus weights determine each participant’s share of the subnet’s accumulated emission each tempo.
Development Stage Context
The Introduction to Bittensor describes subnet development as moving from localnet to testnet and then mainnet. For ItsAI (SN32), that sequence changes how readers should interpret AI text-detection examples and validation scoring outcomes.
In localnet, ItsAI-compatible miners and validators can be developed and tested in an isolated environment. Localnet detection scores and emission outcomes do not represent production subnet performance.
On testnet, ItsAI-compatible components can be exercised in a shared, non-production network. Testnet detection evaluations and validator scores are separate from mainnet subnet state.
On mainnet, ItsAI (SN32) is the live production subnet where miners detect AI-generated text and validators score those detections to determine real Bittensor emissions. The ItsAI incentive documentation describes the evaluation mechanism that applies on the production network.
The Bittensor Networks reference separates mainnet, testnet, and localnet. A detection result or emission outcome from one environment should not be read as representing production subnet performance in another environment.
Reader Boundary
Subnet 32 ItsAI should not be read as generic Bittensor subnet documentation, a plagiarism detection service for all media, or proof that one accuracy score captures user harm from false accusations. It names one subnet’s AI-generated text detection competition on netuid 32 (Understanding Subnets, Glossary: Netuid).
False Positives Are a Separate Quality Axis
ItsAI’s incentive documentation treats false-positive control on human-written samples as part of reward quality rather than raw accuracy alone (incentive documentation).
The subnet therefore distinguishes mislabeling human text from missing generated text.
Augmentation Tests Generalization Rather Than Memorization
The same documentation describes validators altering samples with small changes so miners cannot rely on memorized source text (incentive documentation).
Evaluation is framed around detectors that handle varied-looking text rather than fixed validation sets.
Validator Weights Still Flow Through Yuma Consensus
Subnet 32 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).