AI Content Detector

Detect AI-generated text from modern LLMs using statistical analysis. Analyzes perplexity patterns, sentence burstiness, vocabulary diversity, repetition patterns, and stylistic uniformity — all running locally with zero external API calls.

Why Use This Actor

Existing AI detection tools require expensive subscriptions ($15+/month for GPTZero, Originality.ai) and send your text to external AI services. This actor uses proven statistical indicators running entirely on Apify's infrastructure.

Feature	Ours	GPTZero	Originality.ai	Copyleaks
Price per 1K texts	$6.50	~$10-15/sub	~$15/sub	~$10/sub
Batch processing	✅ 100K+	Limited	Limited	Yes
External API needed	None	Proprietary	Proprietary	Proprietary
Speed	<10ms/text	~500ms	~500ms	~500ms
Deterministic results	✅	❌	❌	❌
Open methodology	✅	❌	❌	❌

Honest limitation: Statistical AI detection is not perfect. No detector achieves 100% accuracy. Short texts (<100 words), heavily edited AI text, and formal/technical writing may produce less accurate results. Use probability scores as signals, not verdicts.

How to Use

Zero-config: just provide an array of text strings.

{
  "texts": [
    "In today's rapidly evolving technological landscape, it is important to note that artificial intelligence has demonstrated remarkable capabilities across a wide variety of domains.",
    "I can't believe how weird yesterday was. My dog somehow got into the kitchen and ate half a pizza while I was on a call. The whole thing was chaos."
  ]
}

With metadata and IDs:

{
  "texts": [
    {
      "id": "submission-456",
      "text": "The article text to analyze...",
      "metadata": { "author": "John Doe", "source": "blog" }
    }
  ]
}

Input Configuration

Parameter	Type	Default	Description
texts	array	Demo texts	Plain strings or objects with text/id/metadata. Max: 100,000
analyzePerplexity	boolean	true	Word predictability analysis
analyzeBurstiness	boolean	true	Sentence length variation
analyzeVocabulary	boolean	true	Type-token ratio, vocabulary diversity
analyzeRepetition	boolean	true	Filler phrases, transitional word density
analyzeStyle	boolean	true	Paragraph uniformity, sentence starters
paragraphLevel	boolean	true	Per-paragraph analysis for mixed content
aiThreshold	number	0.65	AI classification threshold (0.50–0.95)
includeSignalDetails	boolean	true	Full signal breakdown in output
onlyAiDetected	boolean	false	Filter to AI-detected texts only

Output Format

{
  "id": null,
  "classification": "ai_generated",
  "aiProbability": 0.78,
  "humanProbability": 0.22,
  "confidence": 0.82,
  "wordCount": 245,
  "signals": {
    "burstiness": {
      "score": 0.71,
      "sentenceLengthVariance": 8.2,
      "sentenceLengthMean": 22.4,
      "interpretation": "Sentence lengths are suspiciously uniform — AI-like pattern",
      "weight": 0.25
    },
    "repetition": {
      "score": 0.69,
      "fillerPhraseCount": 4,
      "transitionalWordDensity": 0.062,
      "interpretation": "High density of AI-characteristic phrases"
    }
  },
  "flaggedPatterns": [
    {
      "pattern": "uniform_sentence_length",
      "description": "Sentences are unusually similar in length",
      "severity": "strong"
    },
    {
      "pattern": "filler_phrase_density",
      "description": "4 AI-typical filler phrases detected",
      "severity": "moderate"
    }
  ],
  "paragraphAnalysis": [
    {
      "paragraphIndex": 1,
      "aiProbability": 0.82,
      "classification": "ai_generated",
      "dominantSignal": "repetition"
    }
  ],
  "analyzedAt": "2026-03-11T12:00:00.000Z"
}

Signal Score Interpretation

Range	Meaning
0.0 – 0.3	Strongly suggests human writing
0.3 – 0.5	Mildly suggests human writing
0.5 – 0.7	Uncertain
0.7 – 1.0	Strongly suggests AI writing

Detection Signals

• Perplexity: AI text uses consistently "safe" word choices — high predictability, low variance
• Burstiness: Human text has highly variable sentence lengths; AI text is uniform
• Vocabulary: AI text often repeats vocabulary at higher rates (lower type-token ratio)
• Repetition: AI text overuses transitional words ("furthermore", "moreover") and filler phrases
• Style: AI text tends to have uniform paragraph lengths and repetitive sentence starters

Tips and Advanced Usage

• Paragraph-level analysis: Enable paragraphLevel: true to detect mixed content (human intro + AI body)
• Adjusting threshold: Lower threshold (0.50) = more aggressive, higher false positive rate; Higher (0.85) = only flag clear AI text
• False positives: Formal academic and technical writing may score as AI-like — check flaggedPatterns for context
• Batch audits: Use onlyAiDetected: true to filter a large content set to flagged items for manual review

Pricing

$6.50 per 1,000 texts analyzed (PPE — pay only for successfully analyzed texts)

Pricing includes all platform compute costs — no hidden fees.

Use Case	Texts	Cost
Student assignment check	30	$0.20
Blog content audit	500	$3.25
Monthly submission screening	5,000	$32.50
Platform content moderation	100,000	$650.00

Texts below 50 characters and empty inputs are NOT billed.

FAQ

How accurate is this detector?

Statistical detection typically achieves 70-85% accuracy on unmodified AI text. Heavily edited AI text and formal human writing reduce accuracy. Always treat results as probability signals, not proof.

Can it identify which AI model wrote the text?

No — it detects AI-like statistical patterns, not model fingerprints. Outputs from modern LLMs tend to score similarly.

What's the minimum text length?

50 characters minimum. For reliable results, 200+ words recommended.

Does it work on non-English text?

Most signals work across languages (burstiness, vocabulary diversity). Filler phrase detection is English-only.

Is this suitable for academic integrity checking?

As one signal among many, yes. Do not use as sole evidence of misconduct — human review is always recommended.

How does it differ from GPTZero?

GPTZero uses a proprietary AI model (which itself has biases). This actor uses transparent, deterministic statistical methods — same input always produces same output, and you can understand exactly why a text was flagged.

---

Related actors by Junipr: Address Validator | Temporary Email Generator