AI Excluder v1

Hybrid AI‑content detection and exclusion SDK for feeds, search results, and apps.

Install

npm install ai-excluder-v1

Node.js v20+ is recommended for best compatibility. (Current dev env is on v18; upgrade via nvm install 20.)

Quick start

const { AIExcluder } = require('ai-excluder-v1');

const excluder = new AIExcluder().init({
  threshold: 0.6,
  onExclude: (item, res) => console.log('Excluded', item.id, res.confidence),
});

const items = [
  { id: 1, type: 'text', content: 'As an AI language model...' },
  { id: 2, type: 'text', content: 'I wrote this myself.' },
];

const filtered = await excluder.filterContent(items);
console.log(filtered);

Direct detector usage:

const { detectAI } = require('ai-excluder-v1');

const res = await detectAI('Some text', 'text');
// => { isAI: boolean, confidence: number }

API

detectAI(content, type, options?)

• content: string for text, tf.Tensor3D for images, or { text, image, metadata } for hybrid.
• type: 'text' | 'image' | 'hybrid'.
• options.threshold: override detection threshold.
• options.metadata: optional metadata string for watermark checks.

Returns { isAI, confidence } (0–1).

class AIExcluder

• init(options)
  • threshold: default decision threshold.
  • storage: optional persistent storage (defaults to localStorage in browser or memory in Node).
  • prefs: initial preferences override.
  • onExclude(item, res): callback when an item is excluded.
• getPrefs() / setPrefs(partial)
  • Preferences: excludeAI, excludeTextOnly, excludeImagesOnly, threshold.
• filterContent(items)
  • Filters an array of items, excluding those flagged as AI.
• bindToggle(element, onChange?)
  • Plain JS: attach click handler to toggle AI exclusion.
• withAIExclusion(Component)
  • React HOC: wraps a component that receives items prop.

Demos

• Web demo:

  npm run demo:web

  Open http://localhost:3000 and toggle exclusion.

• React demo:

  npm run demo:react

  Demo lives in demos/react/react-demo.

Limitations

• Detection is hybrid and heuristic‑heavy in v1; false positives/negatives are possible.
• Text ML uses roberta-base-openai-detector via @xenova/transformers and downloads on first use.
• Image detection is best‑effort unless a stronger fine‑tuned model is plugged in.

See feedback.md for planned improvements.

License

MIT. See LICENSE.