Human Mark
Scan a draft for the phrasing that reads as AI before you publish it, then issue a certificate your audience can check.
HumanMark
Scan a draft for the phrasing that reads as AI before you publish it, then issue a certificate your audience can check.
Date: 2026-08-20 Form factor: web app Status: Prototype
What it is
HumanMark is a pre-publish scanner for creators and small brands. Paste a draft in and it flags the specific phrasing patterns that make writing read as AI-generated — promotional filler, vague "experts say" attributions, forced rule-of-three lists, and a handful of others — then scores the piece. Below a threshold, the piece can be issued a Human-Made certificate: a badge with an ID, a timestamp, and a public verification page a reader can check against.
Who it serves
Solo creators, newsletter writers, and small marketing teams who've started hearing "this feels like AI" from their own audience and don't have a way to check a draft before it goes out, or to prove after the fact that a piece was actually theirs.
Why it could be profitable
Freemium: a handful of free scans a month, then a subscription for unlimited scans, certificate issuance, and an embeddable badge widget. An agency tier adds seats and a client dashboard, since agencies are the ones fielding "did you use AI for this?" from their own clients.
The demand signal here is real but indirect. Consumer preference for AI-generated creator content dropped from 60% in 2023 to 26% now, and 55% of people say they're more likely to trust a brand that publishes human-generated content (getboldmarketing.com, citing 2025–2026 survey data). By Q4 2025, "more human-generated content" was the top thing consumers said they wanted from brands in 2026. High-profile misses — Coca-Cola's AI holiday ad, McDonald's Netherlands pulling its AI Christmas spot — show what happens when a brand gets caught on the wrong side of that shift.
What's less proven is that creators want a badge for it specifically. The AI-detection tools that already have paying customers — Channel Factory's AI Slop Detection, DoubleVerify's AI SlopStopper — sell to media buyers and ad platforms, not to the people writing the content. Nobody's shown yet that a solo creator will pay for a certificate the way a Fortune 500 ad buyer pays for a compliance report. That's the open bet this prototype is testing, not a settled one.
Form factor & scope
Single-page web app. This prototype demonstrates the scan-score-certify-verify loop end to end, running entirely client-side against a small pattern library and a mock certificate registry — no real AI-detection model, no server, no accounts.
How to run it
- Open
index.htmlin any modern browser. - Paste a draft into the scanner (or click "Load a sample draft") and click Scan for AI tells.
- Review the flagged phrases, then click Issue Human-Made Certificate if the score clears the threshold.
- Switch to the Certificates tab to see previously issued badges, or paste a certificate ID into the verify box to check one.
What's in this prototype
- A pattern-matching scanner covering seven AI-tell categories (promotional language, vague attribution, rule-of-three padding, inflated significance, filler/hedging, negative parallelism, em dash overuse), each with inline highlighting on the draft text.
- A 0–100 AI-tell score with a pass/fail certification threshold.
- A certificate registry seeded with 12 sample brands and past scores, browsable and sortable.
- A certificate verification lookup by ID, returning issuer, score, and issue date.
- A copyable embed snippet for a passing certificate.
Roadmap
- Swap the regex pattern library for a real classifier trained on labeled AI/human text pairs.
- Real accounts, real badge hosting, and a public verification URL instead of a client-side lookup.
- Version history per piece, so a certificate can point to the actual edit trail instead of a single snapshot.
- Slack/CMS integrations so the scan runs where drafts already live (Google Docs, WordPress, Ghost).
- Team plans with per-seat scan quotas and a shared brand style guide the scanner checks against.
Sources
- Human-Made is the New Premium — consumer preference stats (60%→26% AI-content preference drop, 55% trust-human-content figure, Q4 2025 "more human content" demand data)
- The AI Marketing Backlash: Why 'AI-First' Brands Are Starting to Fall Flat — Coca-Cola, McDonald's Netherlands examples and research on how labeling content as AI-generated reduces perceived authenticity
- Channel Factory Announces the Launch of AI Slop Detection — existing paid AI-slop detection product, sold to media buyers rather than creators
Requirements
HumanMark — Requirements
Goals
- Let a user paste a draft and get an immediate, itemized breakdown of AI-tell phrasing rather than a single opaque score.
- Make the certify step feel earned — a certificate only issues once flagged patterns are addressed or the score clears a stated threshold.
- Make a certificate checkable by someone other than the person who issued it.
- Demonstrate the full loop (scan → fix → certify → verify) in one sitting, with no login.
Primary user
A newsletter writer or small brand's content lead who has had a reader or client ask "did you write this or did AI?" and wants a way to check a draft before publishing, and a way to point to something after the fact.
Functional requirements
- FR1: The scanner accepts pasted or typed text in a textarea with no minimum length enforced below a small floor (20 words) needed for a meaningful score.
- FR2: A "Load a sample draft" control fills the textarea with one of at least three preset drafts of varying AI-tell density, so the flow is demonstrable without a paste.
- FR3: Scanning runs entirely in the browser against a defined pattern library; no text leaves the page.
- FR4: The scanner detects at minimum seven categories: promotional language, vague attribution, rule-of-three padding, inflated significance/legacy language, filler and hedging phrases, negative parallelism ("it's not just X, it's Y"), and em dash overuse (rate per 100 words above a set threshold).
- FR5: Each detected match is shown inline in the draft text (highlighted) and listed in a results panel with its category label and the exact matched phrase.
- FR6: An overall AI-tell score (0–100, lower is more human) is computed from a weighted count of matches per 100 words, and shown with a pass/fail line against a certification threshold (score < 25).
- FR7: When the score passes, an "Issue Human-Made Certificate" control becomes available; when it fails, the control is disabled with a note on how many points/matches stand between the draft and passing.
- FR8: Issuing a certificate generates a certificate record (ID, title, score, issue date, issuer name) and adds it to the in-session certificate registry.
- FR9: A Certificates view lists all certificates (seeded sample data plus any issued this session), sortable by score and by date.
- FR10: A certificate detail shows an embeddable badge snippet (copyable HTML) referencing the certificate ID.
- FR11: A verification lookup accepts a certificate ID and returns the matching record's issuer, score, and issue date, or a clear "not found" state.
- FR12: The pattern library and thresholds are defined in one place in
script.jsso categories/weights are easy to see and adjust. - FR13: The app runs from
file://with no build step and no network calls other than loadingsample-data.json.
User stories
- As a newsletter writer, I want to paste my draft before I hit send, so that I can catch phrasing that reads as AI-written before a reader does.
- As a small brand's content lead, I want to see exactly which phrases got flagged and why, so that I can fix the draft instead of guessing at a score.
- As a content lead, I want to issue a certificate once a piece passes, so that I have something to point to if a client or reader questions it.
- As a reader or client, I want to look up a certificate ID, so that I can confirm a piece was actually certified and not just claimed to be.
- As an agency account manager, I want to browse past certificates across pieces, so that I can see which clients' content is trending toward more or fewer AI tells over time.
Non-functional requirements
- No text pasted into the scanner is transmitted anywhere; all matching runs client-side.
- The app is usable on a phone-width viewport (scanner, results, and certificate views all reflow).
- Certificate IDs are generated deterministically enough within a session to be looked up immediately after issuance, but are treated as opaque strings by the UI.
- No external accounts, API keys, or authentication of any kind.
Out of scope (for the prototype)
- A real machine-learned AI-detection classifier — this prototype uses a transparent pattern library, not a trained model.
- Persistent storage across sessions or across browsers — certificates issued in one session don't survive a reload.
- Real badge hosting or a publicly resolvable verification URL — verification here is a lookup against the local sample data plus session-issued certificates.
- Payment, accounts, or team/seat management.
Open questions
- Will creators pay for a certificate directly, or does this only work bundled into a tool they already pay for (a CMS, a scheduling tool, an agency platform)?
- Does a pattern-matching scanner catch enough real AI phrasing to be trustworthy, or does it need a trained classifier before anyone would rely on the score?
- Who is the certificate actually for — the audience reading it, or the client paying the agency that issues it?
