AI Detector by Ceptile is a free, transparent multi-modal AI detector and AI checker created by Ceptile(ceptile.com) and engineered to run entirely inside your browser. We publish the complete mathematical formulas, benchmark evaluation curves, and inherent failure modes — because any instrument that labels someone’s work owes them the exact arithmetic behind that verdict.
We created AI Detector by Ceptile after closely auditing the dominant commercial tools across the AI detection market: ZeroGPT with its proprietary DeepAnalyse™ branding and restrictive character paywalls; QuillBot with its subscription-gated highlights; GPTZero with its enterprise server pipeline; Originality.ai with its closed-source classifier and credit tokens; and Copyleaks with its enterprise writing-replay cloud. While technically impressive, all of them share a fatal architectural compromise:you paste proprietary writing or upload confidential media to someone else's remote server and receive a percentage score you cannot independently verify.
In practice, this cloud-dependent model incurs three severe penalties:
We engineered the complete antithesis: an open-box statistical instrument with zero server-side inference and zero data uploads. Our client-side engine computes 46 published linguistic signals alongside forensic image, video, and audio discriminators. Every score includes an honest prior assumption (“assume human authorship until empirical evidence proves otherwise”) and a bootstrapped 90% confidence interval. If an analysis result seems erroneous, you can inspect the exact signal row responsible in your browser’s developer tools.
A note on scientific honesty: Our prototype calibration yielded a 0.989 AUC when tested against hand-crafted synthetic prompts. However, when evaluated against real-world LLM outputs (ChatGPT, Claude, Gemini, Mistral), accuracy plummeted below chance. Rather than masking this failure, we discarded the synthetic samples and refitted all weights across a rigorous benchmark of 1,153 authentic human texts and 1,590 model generations. Accuracy initially dropped to an honest 0.909, which we subsequently elevated to 0.966 through multi-dimensional vocabulary diversity (MATTR) and cross-validated logistic regression. We publish our real-world limitations openly rather than hiding behind unscientific marketing claims.
AI Detector by Ceptile provides specialized, auditable client-side forensics across all four major digital media formats:
Executes 46 closed-form natural language metrics inside the browser's JavaScript engine: Zipf law word rank distributions, Moving-Average Type-Token Ratio (MATTR), sentence length burstiness, n-gram repetition entropy, and token-transition cadence. Includes built-in ESL-safe weighting to prevent false positives on non-native English speakers.
Decodes image files into in-memory canvas buffers. Evaluates Error Level Analysis (ELA) compression differentials, Bayer Color Filter Array (CFA) demosaicing residuals, high-frequency discrete cosine transform (DCT) noise grids, and parses C2PA Content Credentials and EXIF metadata without uploading file bytes.
Extracts video frames using HTML5 video decoders to evaluate temporal optical flow continuity. Measures inter-frame shot noise variance, cadence consistency, and temporal blurring artifacts characteristic of modern diffusion video generators (including OpenAI Sora, Google Veo, and Kling AI).
Utilizes the browser's Web Audio API (AudioContext) to compute fast Fourier transforms (FFT). Scans audio waveforms for unnatural high-frequency cutoffs, spectral centroid dynamics, neural vocoder phase coupling (Bark, ElevenLabs), and synthetic silence markers common to AI voice cloning and generated music (Suno, Udio).
If software accuses someone of AI plagiarism, it must provide the exact proof. A percentage score without derivation is arbitrary conjecture. We publish every mathematical formula, weight coefficient, and fusion calculation so that anyone can replicate our conclusions.
Your intellectual property belongs exclusively to you. We never upload your text or media to cloud servers. There is no premium tier to purchase privacy — privacy is guaranteed to all users by physical architectural design.
We openly reject dishonest “99% accuracy” marketing banners. We publish our complete receiver operating characteristic (ROC) curves, per-genre false-positive rates, and failure modes. Our engine is versioned (2026.2) and fully deterministic: identical input produces identical output every time.
Neural classifiers trained on specific LLMs rapidly degrade when newer foundation models launch. Mathematical heuristics that capture fundamental properties of token decoding and image synthesis provide durable, long-term diagnostic utility.
| Industry SaaS Pattern | AI Detector by Ceptile |
|---|---|
| Server-Side Ingestion: Content uploaded, stored in remote database logs, and potentially mined for model training unless opted out. | 100% On-Device Processing: Zero uploads, zero retention, zero training ingestion — enforced by browser architecture. |
| Opaque Single-Percentage Score: Output provided without explanation, confidence intervals, or mathematical derivation. | Exhaustive Signal Table: Full transparency with 46 linguistic signals, z-scores, bootstrapped 90% confidence intervals, and reliability grades. |
| Binary Forced Verdict: Forces a rigid "AI vs. Human" claim even on short or ambiguous text samples. | Three-Tier Output: Probabilistic classification (Likely Human, Mixed / Paraphrased, Likely AI-Generated) with explicit "Inconclusive" status on short samples. |
| Paywalled Reports & History: Requires paid credits, subscription plans, or mandatory account creation to export findings. | 100% Free & Open: Unrestricted PDF, JSON, Markdown, and CSV exports, with private local history stored on your device. |
| Unpublished Heuristic Weights: Arbitrary proprietary scoring logic hidden behind marketing buzzwords. | Fully Published Math: Every metric, coefficient, and logistic fusion equation is documented on our How it works page. |
Competitor descriptions reflect publicly available features and marketing claims at the time of writing. We claim no affiliation with or endorsement by ZeroGPT, QuillBot, GPTZero, Originality.ai, Turnitin, or Copyleaks.
Principles we uphold regardless of commercial temptation:
AI Detector by Ceptile (aidetector.ceptile.com) is a core open-science initiative ofCeptile (ceptile.com), an innovative AI search engine, deep research workspace, and knowledge platform founded and developed by Shaurya Agarwal (originally prototyped as Trexinity).
Ceptile was engineered from first principles to resolve the fundamental flaw of modern search engines: the proliferation of SEO link-farming, intrusive advertisements, and unverified AI hallucinations. Ceptile retrieves verified multi-source web evidence and synthesizes transparent answers with reproducible, click-to-source citations.
Combines real-time web discovery with an autonomous Deep Research engine that recursively investigates multi-layered topics, cross-references scholarly sources, and synthesizes cited research briefs.
Organize personal knowledge with Ceptile Notes — a local-first Markdown notebook and citation repository stored securely on your device, mirroring this detector's privacy design.
Enables natural multi-language research via Ceptile Translate, visual knowledge search, voice queries, and collaborative research workspaces designed around user sovereignty.
As generative models flood online search indexes and academic literature, trustworthy research demands knowing the provenance of information. Building an auditable, client-side AI detector empowers researchers, writers, and students to verify content authenticity without sacrificing their personal privacy.
AI Detector by Ceptile is operated directly by Ceptile and hosted at aidetector.ceptile.com. We do not pretend to be a bureaucratic accreditation body or commercial certifier. Instead, we offer concrete engineering proofs: verifiable client-side source code, published mathematical formulations, deterministic engine versioning (2026.2), and an uncompromising zero-upload privacy posture you can verify in your own browser.
There are no user accounts, no subscriptions, and no commercial data tracking mechanisms on aidetector.ceptile.com. Direct technical correspondence is managed by our engineering team via email athello@ceptile.com and our parent portal atceptile.com.
Discovered a false positive on human writing? Identified an edge-case bug? Want to suggest model calibration adjustments? Send us a short excerpt (1–2 paragraphs) alongside the engine version from your report. We review all correspondence carefully.
Explore related pages: AI Detector Tool · Privacy Policy · Terms & Conditions · Accuracy & Limits
