10% off with code plagiarism-scan10 in the shop!
All LLM Families Covered

LLM Detector: Identify Text from ChatGPT, Claude, Gemini & More

One detector for all major large language models — GPT-4, Claude, Gemini, Llama, Mistral, Phi and beyond. Every sentence scored from 25 (human) to 75 (LLM-generated) · Results in 15 minutes · from $0.29/page.

All major LLMs Results in 15 min. German servers No registration
The Technical Basis

Why All LLMs Produce Detectable Text Patterns

Despite differences in architecture and training, all large language models share the same fundamental text generation mechanism — and this creates universal detectable patterns.

Autoregressive Generation

All LLMs generate text by repeatedly predicting the next token (word or word-piece) based on what came before. This "greedy" or "sampled" prediction always favours high-probability tokens, creating low perplexity — the primary detection signal.

Temperature & Sampling

LLMs use a "temperature" parameter to control randomness. Even at high temperatures, models avoid genuinely improbable words. Human writers regularly choose "improbable" words that are nonetheless semantically perfect — models consistently cannot replicate this.

Training Data Patterns

All major LLMs were trained on similar corpora of internet text and books. This means they all learned the same "safe" academic and formal register — the uniformly fluent, hedge-filled style that detection systems have learned to recognise.

Full Coverage

All Major LLMs Covered by the Detector

Detailed coverage of every major large language model family used in academic and professional writing.

OpenAI: GPT-3.5, GPT-4, GPT-4o

The most widely used LLMs in academic contexts. GPT-4 produces more varied output than GPT-3.5 but remains readily detectable. GPT-4o coverage is included.

Anthropic: Claude 2, 3 Haiku, Sonnet, Opus

Claude models are known for producing especially polished academic prose. All Claude 3 variants (Haiku, Sonnet and Opus) are covered alongside the older Claude 2.

Google: Gemini Pro & Ultra

Google Gemini is increasingly used for research assistance. Both the Pro and Ultra tiers are covered. Bard (now Gemini) output is also detected.

Meta: Llama 2 & Llama 3

Meta's open-source Llama models power many third-party writing assistants and chatbots. Students may use Llama-based tools without knowing it — the detector covers them all.

Mistral & Phi

Mistral (7B, Mixtral) and Microsoft Phi are efficient open-source models that power a wide range of AI writing tools. Both are covered by the detection engine.

Emerging Models

As new LLMs enter the market — from major labs or the open-source community — the detection model is updated to include them. Coverage stays current.

How It Works

How to Use the LLM Detector

1

Upload Your Document

Submit .docx, .pdf, .txt, .odt or .rtf. Up to 50 MB. No account required. Processed on German servers; never added to a public database.

2

Multi-Model LLM Analysis

The PlagAware engine analyses perplexity, burstiness and token distributions for every sentence. The detection model is trained on text from all major LLM families simultaneously for broad coverage.

3

Scored Report Delivered

Within 15 minutes you receive a PDF with each sentence scored 25–75. LLM-generated sentences are highlighted in red. You can identify which passages came from which model family based on context and scoring patterns.

Score Scale

The LLM Detection Score: 25 to 75

The score anchors on two statistical baselines: the typical perplexity and burstiness of human-written academic text (scoring near 25) and the typical perplexity and burstiness of LLM-generated text (scoring near 75).

Because different LLMs have different default generation settings, some models (like GPT-4 with high temperature) tend to score slightly lower than others (like heavily constrained GPT-3.5). The scale captures the full range.

  • Score 25: Maximum human writing signal
  • Score 75: Maximum LLM generation signal
  • Scores 40–55: Ambiguous — review in context
  • Each sentence scored independently
LLM Detection Report 4 LLM sentences

93%

Human

4

LLM sentences

44

Pages

Sentence analysis

29

The sample included 47 participants recruited from three different departments.

68

It is crucial to acknowledge that the multifaceted nature of this phenomenon necessitates a comprehensive and nuanced approach.

33

Two participants withdrew before completing the final questionnaire.

Pricing

LLM Detection Pricing

Pay per page. No subscription. Minimum order $0.90.

Plagiarism Scan

Source plagiarism check only

$0.29 / page
  • 70+ billion sources
  • Results in 15 minutes
  • PDF report with sources
  • PlagAware technology
Order Plagiarism Scan

AI Scan

LLM detection only

$0.29 / page
  • All major LLMs covered
  • Score 25–75 per sentence
  • Results in 15 minutes
  • PDF by email
Order AI Scan
FAQ

LLM Detector — Frequently Asked Questions

An LLM (Large Language Model) detector is a tool that identifies text generated by AI language models such as ChatGPT, GPT-4, Claude, Gemini, Llama and Mistral. Unlike a plagiarism checker — which compares text to existing sources — an LLM detector analyses the statistical properties of the text itself to determine whether it was generated by an AI model.

The detector covers ChatGPT (GPT-3.5, GPT-4, GPT-4o), Claude (Anthropic — all versions), Gemini (Google — Pro and Ultra), Llama (Meta — 2 and 3), Mistral, Phi (Microsoft), Falcon and other open-weight models. The detection model is trained on text from all major LLM families and is updated as new models emerge.

All LLMs share a fundamental text generation mechanism: they predict the next word by sampling from a probability distribution. This causes LLM output to have characteristically low perplexity (predictable word choices) and low burstiness (uniform sentence length variation). The detector measures these properties for every sentence and assigns a score from 25 (human) to 75 (LLM-generated).

The primary report shows the AI detection score per sentence, not the specific model that generated it. The detector confirms LLM involvement reliably, regardless of the specific model. Attributing text to a specific model (GPT-4 vs Claude vs Gemini) is a more complex task beyond standard detection.

Yes. Open-source models including Llama 2, Llama 3, Mistral and Falcon produce text with the same low-perplexity, low-burstiness statistical properties as proprietary models. Many writing assistant tools and chatbots are built on these models, so coverage is essential for comprehensive LLM detection.

Detect LLM-Generated Text — Regardless of Which Model Was Used

Upload your document and receive a sentence-level LLM detection report in 15 minutes. All major models covered.

Use code plagiarism-scan10 for 10% off.