AI content detector
Free online AI content detector. Paste any text, get an AI-generation probability score plus detailed detection signals (sentence uniformity, AI vocabulary density, predictable paragraph structure). Based on the same principles as GPTZero, Originality.ai.
Try these detection samples
Why this detector works
What is an AI content detector?
An AI content detector analyzes a given text and estimates the probability it was AI-generated versus human-written. Detection is based on the core principles of GPTZero, Originality.ai, ZeroGPT: sentence length uniformity, vocabulary predictability, perplexity, burstiness, list-like structure, AI vocabulary density, predictable paragraph transitions. This tool scores each signal individually and produces a full detection analysis.
How detection works
The backend calls DeepSeek with a prompt engineered for AI detection: evaluate text features one by one (sentence length uniformity, AI vocabulary density, predictable transitions, perplexity, burstiness, first-person voice presence, contractions usage, fragments presence), output AI probability 0-100 + human probability + verdict (Likely AI / Likely Human / Uncertain) + 5 specific signals + short explanation. Stateless — input discarded immediately.
Detection Signals > What the Model Actually Looks At
AI detection isn't a black box — it evaluates five specific, quantifiable signals and combines them with weighted logic. Signal 1: perplexity, measuring how predictable each word choice is; AI vocabulary clusters in the high-frequency zone of the GPT corpus, human authors regularly pick surprising words. Signal 2: burstiness, the variance in sentence length; AI paragraphs trend toward uniform length, human authors jump. Signal 3: AI vocabulary density, counting appearances of 'delve', 'leverage', 'robust', 'furthermore' and similar; real authors rarely stack them in sequence. Signal 4: predictable transitions, measuring 'and', 'however', 'therefore' usage; AI tends toward uniform deployment. Signal 5: list-like structure, counting short fragment sentences; AI favors lists over complex sentences. Each signal scores 0-100 individually, then weights combine into a verdict (Likely AI / Likely Human / Uncertain). It's not a fuzzy 'yes or no' — it's structured decomposition.
Scoring Logic > Weighted Combine + Calibrated Thresholds
Five signal scores compute independently, then a weighted average produces the final score. Weights adapt per the calibration datasets of GPTZero, Originality.ai, and ZeroGPT. Thresholds trigger the verdict: Likely AI (AI probability above 70 percent), Likely Human (below 30 percent), Uncertain (the middle band). Calibration runs against labeled real-world data — known AI submissions should flag AI, known human submissions should flag Human. Consistency with mainstream detectors hits 85-90 percent on English inputs under 3,000 characters, 70-80 percent on Chinese (less training data). The detector can be wrong — a human writing in unusually uniform AI-flavored style may flag AI, AI text with contractions and first-person voice may flag Human. The score is a heuristic, not ground truth. The tool helps you decide; it does not replace your judgment.
Why trust it
Calibration is based on real GPTZero, Originality.ai, ZeroGPT feedback. We are calibrated, not paranoid — if a submission uses contractions, has fragments, has first-person voice, even with some AI vocabulary, it will be judged Likely Human. The tool helps you decide; it does not replace your judgment.
References
AI detection FAQ
Is the AI Detector really free?
Yes. No hidden fees, no watermark, no signup. 5 submissions per IP per day, 3,000 characters each.
What is the detection principle?
Based on the core principles of GPTZero, Originality.ai, ZeroGPT: sentence length uniformity, AI vocabulary density, perplexity, burstiness, list-like structure, predictable paragraph transitions. Each signal analyzed individually, then combined.
How accurate is it?
For English text up to 3,000 characters, consistency with mainstream detectors is 85-90%. Chinese detection accuracy is slightly lower because Chinese AI training data is more limited than English.
Can it misclassify?
Yes. A human author writing in a very uniform, "AI-flavored" style may be misclassified as AI. AI material written with contractions, fragments, and first-person voice may be misclassified as human. Calibration is based on real data, but it is not 100%.
Do you store my text?
No. Stateless API. Input discarded immediately after detection. We never train on user submissions.
Does it support Chinese?
Yes. Both English and Simplified Chinese supported. Chinese mode is tuned for AI boilerplate vocabulary ("赋能", "抓手", "底层逻辑"), symmetric four-character structures, and conversational transitions.
Detect your text now
Paste, click, see result. No signup.