detect·deepfakesby Resemble AI
Text detection

Is this text AI-generated?

Paste any passage — article, email, essay, message — and get a verdict in under a second. Works on GPT, Claude, Gemini, Llama, and most open-weight LLMs.

  • · Privacy mode on by default — text never stored
  • · Three depths of analysis: low (fast) → high (most accurate)
  • · Best on passages of 250+ words
0 / 100,000 chars · 50 more needed

How the text detector works

Every LLM leaves a statistical fingerprint. Token probabilities, n-gram distributions, punctuation rhythm, and stylometric variance all differ between human prose and machine-generated output. We combine these signals with model-specific features to attribute a passage to AI or human origin.

The `depth` control changes how much reasoning we apply: low for fast triage, high for forensic-quality analysis on disputed text.

Where this detector falls short

No text detector is 100%. Short passages (under 50 words), non-English text, heavily edited AI output, and domain-specific jargon (legal, medical, code) are all harder. For high-stakes decisions, treat the score as one input alongside context and human judgment — never the only one.

Frequently asked questions

How accurate is the text detector?
Accuracy scales with passage length. On 250+ word samples we observe >95% accuracy across major LLM families. Short passages (under 50 words) are genuinely hard for any detector — treat them as inconclusive.
Which LLMs does it catch?
GPT (OpenAI), Claude (Anthropic), Gemini (Google), Llama (Meta), plus most open-weight derivatives. It also flags light human-edited AI output — full rewrites are harder.
Is my text stored?
No. detectdeepfakes.com sends every request with privacy_mode enabled, so Resemble AI does not persist the text content. We do not store it either.
Can I call this from my app?
Yes — the same detector is available via the Resemble AI API. See app.resemble.ai for production access and rate limits.
Why do I need 50+ characters?
Short strings are statistically indistinguishable between human and AI text. The minimum keeps false positives out of the public tool.