Most AI detectors ship a black-box classifier, claim 99% accuracy, and ask you to trust them. We ship the math, claim ~90% accuracy on clear-cut cases, and tell you to verify everything.
A 99%-accurate black box is useless when you need to defend a decision. We'd rather ship a 90%-accurate pipeline that prints its own math. Every score comes with a per-metric breakdown so a human reviewer can sanity-check the result.
The detector runs in your browser. There is no API call, no upload, no log, no signup. The "free tier" is the only tier. Your unpublished manuscript stays unpublished.
Our engine is versioned and deterministic. Re-running the same text on the same engine version produces the same number, byte for byte. We don't fine-tune on the side and silently drift the model.
We ship a curated, open database of ~100 AI stylistic markers. Each is weighted by how reliably it indicates AI. You can see every phrase that fired in your text, with its category and impact.
Why it's different
Other detectors call a black-box classifier and hand you a number. We hand you forty-six numbers — and the math behind each one.
We don't run one model — we run forty-six. Perplexity, burstiness, lexical diversity, Zipf fit, n-gram repetition, punctuation rhythm, AI-phrase density, and more — plus sentence-level scoring and seven measured-but-unfused diagnostics. Each signal gets its own row and its own explanation.
Beyond AI-or-not, we tell you which model family the text most resembles — GPT/ChatGPT, Claude, Gemini, LLaMA, Grok, or Perplexity — with the specific markers that fired.
There is no LLM, no classifier, no neural net anywhere in this product. The engine is a 600-line statistics pipeline. You can read every line of math.
Drop in a PNG, JPG, WEBP, AVIF or HEIC. We inspect C2PA Content Credentials, EXIF and generator tags first, then run 23 pixel-forensic tests — frequency, sensor noise, colour and statistics — all in your browser.
Drop an MP4, WebM, MOV or MKV. We parse Sora, Veo, Runway, Pika, Kling and C2PA provenance, then sample frames for sensor physics plus temporal grain, flicker and optical-flow motion — production-grade, with frame tables and full exports.
Drop a WAV, MP3, FLAC, OGG or M4A. We decode the waveform on-device and measure spectral shape, vocoder bispectral artefacts and temporal voice dynamics — with acoustic tables, diagnostics and the same JSON, Markdown, CSV, HTML and PDF reports.
Your text, images, video and audio never leave your device. No uploads, no API calls, no logs. The detector works offline once the page loads.
Every percentage comes with the math behind it. You see exactly which signal drove the score and how it's weighted. No black boxes.
Each sentence gets its own AI probability score and is coloured accordingly. Tell-tale AI phrases ("delve into", "tapestry", "navigate the landscape") are flagged inline.
Download a JSON report with every metric, every signal, and the engine version. Re-run the same text on the same engine version and you get the exact same number.
The honest comparison
We won't pretend our heuristic is more accurate than a fine-tuned classifier on a benchmark. We will tell you exactly what we computed and why. That tradeoff is the entire point.
| Capability | AI Detector by CeptileOurs | ZeroGPT | QuillBot | GPTZero |
|---|---|---|---|---|
| Pricing | Free, unlimited, no signup | Free tier with limits; $9.99/mo+ | Freemium; $9.95/mo+ | Free tier; $10/mo+ |
| AI backend | None — pure math | Black-box classifier | Black-box classifier | Black-box classifier |
| Code transparency | Algorithm explained line-by-line | Closed | Closed | Closed |
| Signals computed | 46 metrics, each with weight & description | One aggregate score | One aggregate score | One aggregate score |
| Model identification | Per-model heuristic fingerprint (GPT, Claude, Gemini, LLaMA, Grok, Perplexity) | Generic AI % | Generic AI % | Generic AI % |
| Image AI detection | Yes — provenance, 23 forensic tests in-browser | Yes (separate tool) | No | No |
| Video AI detection | Yes — provenance + frames + temporal, in-browser | No | No | No |
| Audio AI detection | Yes — spectral + bispectral + temporal, in-browser | No | No | No |
| Privacy | 100% client-side, no uploads | Server-side | Server-side | Server-side |
| Account required | Never | For premium | For premium | For most features |
| Per-sentence explanation | Yes, with reason codes | Yes, highlights only | Highlights only | Yes, with reason codes |
| Reproducible results | Yes — engine versioned, JSON export | No | No | No |
| Languages | All (English-tuned, works everywhere) | All | All | All |
| Browser extension needed | No — open the page and paste | Optional | Optional | Optional |
