Signals, not verdicts

What made
this?

Paste a site, drop an image, or paste text. TELL scores it for AI-generation signals and shows you every piece of evidence behind the number.

No account. No upload — images are analysed in your browser and never leave your device.
Every result states what it cannot prove, because most detectors lie about that.

Website origin analysis

~40 signatures · builder fingerprints · code & copy patterns

Fetched through a public CORS relay. If a site blocks it, switch to Paste HTML — open the page, press Ctrl+U, copy everything, paste it here. That mode is always more accurate: it sees the raw markup.

Image origin analysis

C2PA · EXIF · sensor noise · block structure · canvas size
100% local
Drop an image, or click to choose
JPEG · PNG · WEBP — nothing is uploaded anywhere

Writing-style signals

Deliberately not called a "detector" — read why below
Low confidence by design
0 words
/ 01 — METHOD

Every signal, in the open

No black box and no model you have to trust. Each score is a weighted sum of the checks below, and each result shows which ones actually fired.

Website

  • Generator tags & badges — Framer, Wix ADI, Durable, Webflow, 10Web, "Built with Lovable". Near-conclusive when present.
  • Builder CDNs — framerusercontent, wixstatic, lovable.app, v0.dev, bolt.new, base44, durable.co and more in src/href.
  • Assistant leftovers — uniform <!-- Hero Section --> comment series, "TODO: Replace with your actual API key", unreplaced (555) 123-4567.
  • Code patterns — the canonical max-w-7xl mx-auto px-4 sm:px-6 lg:px-8, indigo/purple palette spam, untouched shadcn button strings, deep div nesting.
  • Copy density — AI marketing vocabulary per 1000 words, triads, "whether you're a…", and crucially polish without specifics: no prices, addresses, names or dates.
  • Imagery — DALL·E and Midjourney CDN hosts, prompt-shaped alt text, stock URLs.

Image

  • C2PA / Content Credentials — the signed provenance manifest. DALL·E, Sora, Firefly and Imagen embed it; Midjourney does not.
  • EXIF — camera make/model/lens says capture; a Software field naming a generator says otherwise.
  • Sensor noise — a Laplacian pass measures micro-detail. Real sensors leave grain; diffusion output is glassy-smooth.
  • Block structure — genuine JPEG capture leaves energy on the 8×8 grid. Clean exports don't.
  • Canvas size — exact 512/768/1024 squares are generator defaults, not camera dimensions.
  • Tonal spread — histogram occupancy and highlight/shadow clipping.

Text

  • Burstiness — variance in sentence length. Human writing lurches; model output evens out.
  • Cliché density — a curated list of model-favourite phrases, per 1000 words.
  • Connective load — however / moreover / furthermore / additionally.
  • Contractions — casual human writing uses them; polished output tends not to.
  • Opener repetition — models reuse sentence-opening structures.
  • Specificity ratio — concrete numbers, dates, names and units per 100 words. High polish with low specificity is the strongest single tell.
  • Em-dash rate — a known habit of recent models.
/ 02 — LIMITS

Where this is wrong

Every detector has these problems. Most hide them behind a confident percentage. Here they are.

Never use the text tool to accuse anyone

Text detection is the weakest of the three by a wide margin, and it fails unevenly — it flags some people far more than others.

Detectors lean on perplexity (how predictable the word choices are) and burstiness (sentence-length variation). Someone writing carefully in a second language naturally produces more standard, predictable constructions — which reads to a detector exactly like a language model. Neurodivergent writers are flagged at elevated rates for related reasons.

Meanwhile the reverse is trivial: light paraphrasing defeats most text detectors entirely. So it under-catches the people gaming it and over-catches the people writing honestly.

That's why this tool is called writing-style signals, caps its own confidence, and will never show you a "98% AI" verdict.

Liang et al., Patterns (2023) — 7 detectors flagged human-written TOEFL essays as AI in ~61% of cases; one flagged 97.8%.
Weber-Wulff et al. (2023) — of 14 detection tools tested, none reached 80% accuracy.

Images: absence proves nothing

  • C2PA metadata is stripped by screenshots, by most social platforms, and by ordinary editing. Most images online carry none — that's normal, not suspicious.
  • C2PA proves a signed record exists and still matches the file. It does not prove the photo depicts reality.
  • Pixel heuristics confuse heavy denoising, upscaling and screenshots with generation.
  • "No signals found" means inconclusive — never "confirmed human".

Websites: the honest one

  • This is the only one of the three that can reach real confidence — because generator tags and builder CDNs are facts, not statistics.
  • But Tailwind, Next.js, shadcn and three-card grids are used by thousands of human developers.
  • A Framer or Webflow site can be entirely hand-designed. The platform having AI features proves nothing.
  • Agencies were writing "seamless" and "elevate" long before AI existed.

What a score means

  • 0–24 — no meaningful signals found
  • 25–49 — inconclusive, mixed evidence
  • 50–74 — likely AI-assisted somewhere in the process
  • 75–100 — strong evidence, and only reachable through explicit artifacts like a generator tag or a C2PA AI declaration
  • Statistical signals alone are capped at 74. They cannot earn the top band.