The Unprompters standard
How scoring works
Every listing carries an AI-involvement estimate: roughly how much of the creative work came from AI rather than from the person selling it. It comes with a range, a stage-by-stage breakdown, and the reasoning behind it. Here is exactly how it's built, and where it falls short.
What the tiers mean
1. The creator discloses, stage by stage
“Did you use AI?” is too blunt a question. A novelist who ran spellcheck and one who had a chatbot write chapter three both used AI. So creators answer separately for four stages:
- Idea & concept · Coming up with what to make, research, outlining
- Making it · Writing, coding, drawing, composing, shooting
- Editing & polish · Revising, fixing, retouching, mixing
- Supporting assets · Cover images, descriptions, sample data, extras
For each stage they pick a level:
- No AI (0% of that stage): Spellcheck and basic grammar tools are fine.
- Light help (15% of that stage): AI suggested or checked; I made every creative choice.
- Substantial (50% of that stage): AI produced parts that I edited and kept.
- Generated (100% of that stage): AI produced this; I directed it.
Stages are weighted by what the buyer is paying for. In a poem the writing matters most. In an idea, the idea does.
The disclosure is the floor. Nothing we detect can bring a score below what the creator admits to. New listings must disclose 50% AI or less.
2. We scan the work
Text and poetry are split into passages and checked for statistical habits of language models: stock vocabulary, formulaic constructions, very even sentence rhythm, stacked transition words. Each passage gets a probability, and the score is the share of the text that reads as machine-made. You can see the passage-level result on every listing.
Code is checked for tutorial habits: comments that narrate every line, docstrings on everything, “example usage” blocks, emoji in log strings. Human tells like TODOs and commented-out experiments count the other way.
Images are checked for what the file says about itself: generator names, IPTC “trained algorithmic media” tags, Stable Diffusion and ComfyUI parameters, and C2PA Content Credentials. A generator tag is strong evidence. No metadata means nothing; it gets stripped all the time.
The private deliverable is scanned along with the public preview, but only the preview's results are shown. Scans can only raise a score above the disclosure, and only in proportion to how confident they are.
3. Evidence narrows the range
The best proof that a person made something is the process. Creators can attach evidence, which is shown publicly so anyone can check it:
- Early drafts+25
- Time-lapse or screen recording+35
- Version history (commits, doc history)+30
- Sketches, raw files or project files+25
- Live stream or video of the process+35
- Content Credentials (C2PA) in the file+35
- Other+10
Only evidence with a file or link counts, because a note alone gives nobody anything to check. Strengths combine with diminishing returns, up to 90%. At 50% or more a listing is marked Evidence attached. That means the creator supplied it, not that we checked it, so open it before you rely on it. Evidence also softens detector suspicion. That's deliberate: a time-lapse of someone painting outweighs a pattern-matcher.
What detection can't do
We want to be blunt about this, because a marketplace built on trust can't oversell its tools.
AI detectors are unreliable. OpenAI withdrew its own text classifier in 2023 because of its low accuracy. A peer-reviewed comparison of 14 detectors (Weber-Wulff et al., 2023) found none above 80% accuracy. Independent tests of commercial tools keep finding false-positive rates well above what vendors advertise, and higher still for people writing in a second language.
They are easy to fool. Light paraphrasing, editing, or simply asking a model to write casually defeats most of them. Metadata can be stripped with a screenshot.
So we never treat a scan as a verdict. Our built-in analyzer is a transparent baseline, and its confidence is capped. Every estimate is shown as a range. Disagreements between a scan and a disclosure go to a person.
Reviews & disputes
- If automated checks find substantially more AI than disclosed, or a file's metadata says AI was used at a stage disclosed as “No AI”, the listing is held for review instead of going live. The creator sees why and can add evidence.
- Anyone signed in can report a listing. Reports go to moderators, never straight to removal.
- Moderators look at the evidence, not just the number. Misrepresenting AI use is grounds for removal; an honest disclosure never is.
- Creators can appeal any decision with additional process evidence.