Beyond AI detectors and self-declared labels: a process-based certification for human creators

Human Origin: a process-review label for human-made creative work Concept, method and pilot design Camille G.ORCID: 0009-0009-0364-6396Loch Ness Version 1.0 — 30 September 2026 Pilot website: human-origin-label.lochness-paris.com Abstract Generative AI has made it inexpensive to produce text, images, music and video that appear human-made, and inexpensive to label such work as human-made. Automated AI detectors are not sufficiently reliable to settle questions of authorship, while several existing human-made labels rely primarily on the creator's own declaration. This paper presents Human Origin, a label designed to assess how a work was made rather than what the finished work looks like. A human reviewer examines evidence of the creative process, either during a live video call or through a screen recording containing a one-time code. The reviewer records the outcome on a public, numbered verification page. The proposed system has two visually distinct levels—reviewed and self-declared—as well as revoked status. Badges are served dynamically so that revocation can propagate to displayed badges, and the system monitors the domains on which reviewed badges appear. A discipline-specific Creator Charter defines the permitted and prohibited uses of AI for writing, illustration, photography, music and video. This paper describes the design principles, review procedure, evidence requirements, badge architecture, limitations, threat model and pilot roadmap of Human Origin. Keywords: human-made content; generative AI; authorship; certification label; provenance; process evidence; trust marks; creators 1. Introduction and problem statement Creators who make their work themselves face a new form of uncertainty. A human-written essay, hand-drawn illustration or recorded track can be questioned simply because a machine could have produced something similar. At the same time, generated content can be presented as human-made at very low cost. Automated AI detection does not resolve this problem reliably. An independent test of 14 detection tools found that all scored below 80% accuracy and only five exceeded 70%, with performance declining further when AI-generated text was edited or paraphrased [1]. OpenAI withdrew its own AI text classifier in July 2023, citing its low accuracy [2]. Detector performance can also be uneven across populations: a Stanford study reported that seven detectors incorrectly classified, on average, 61% of essays written by non-native English speakers as AI-generated [3]. A detector score is therefore a weak basis for decisions concerning individual creators. Labels have emerged in response. Several widely used labels are primarily declarative: creators state that their work is human-made without an independent review of the production process. Not By AI, for example, explicitly describes itself as not being an AI-detection tool and relies on creators following its 90% human rule [4]. The Authors Guild's Human Authored certification operates on an honor system supported by a signed license and public database [5][6]. Declarative labels can be useful statements of intent and are comparatively easy to adopt. However, they cannot independently distinguish a genuine declaration from a false one. Human Origin explores a different approach: review the evidence of the creative process rather than attempt to infer authorship from the finished artifact. 2. Related initiatives Several services now provide forms of evidence-based or declarative certification, often using numbered certificates or publicly verifiable records. Their approaches differ principally in who assesses the evidence, what evidence is considered and what information is retained. Initiative Assessment model Notable features Not By AI [4] Self-declared Widely used badges; creators follow a 90% human rule Authors Guild Human Authored [5][6] Honor system; signed license Numbered mark for books; public database Officially Human Art [7] Self-declaration registry with uploaded evidence Certificate ID, QR code, embeddable badges; tier based on amount of evidence Humanmade [8] Hand verification by staff Earlier work, online presence and timelapses; free pledge level and paid certification options HUMA Certificate [9] Expert review and automated forensic tools Four categories of AI involvement; one-year validity The Human Made Mark [10] Review of production materials Film and television; unique number shown in credits Human Origin belongs to the broader evidence-based certification family. Its distinguishing design choices are: review of the creative process itself; no routine retention of creators' source files; dynamically served badges whose revoked status can propagate; monitoring of where reviewed badges are displayed; and discipline-specific standards defining what the certification does and does not cover. 3. Design principles 3.1 Process over product A finished work cannot, by itself, establish how it was produced. The review therefore focuses on traces of the creative process, including: drafts; version histories; layered source files; raw captures; project files; notes and outlines; and other discipline-specific production evidence. The objective is not to identify whether a finished artifact looks human-made, but to examine whether the available process evidence is consistent with the creator's account. 3.2 Claims no stronger than the evidence Human Origin does not claim that no AI tool was ever used. Instead, the label records that an independent reviewer examined the creative process associated with a specified work, using a specified method, on a specified date. The scope of the certification is therefore explicitly limited to the work and process described on the verification page. 3.3 Two levels that cannot be confused Human Origin distinguishes between: Reviewed — the creative process has been examined under the Human Origin review procedure; Self-declared — the creator makes the declaration, but the process has not been independently reviewed; and Revoked — a previously issued reviewed status is no longer valid. The badge system uses visually distinct colours: green for reviewed; orange for self-declared; and red for revoked. Status is also expressed through text and icons so that the distinction does not depend on colour perception. 3.4 Public verifiability Every reviewed badge has a unique number resolving to a public verification page. The verification page records, as applicable: creator; discipline(s); work covered; review date; review method; evidence examined; official website; official social accounts; current status; and an explanation of what the badge does and does not certify. 3.5 Trust that can be lost A badge may be revoked when the applicable standard is breached. Because badges are served dynamically by the Human Origin server, a revoked badge can change to a revoked state wherever it is displayed, subject to caching behaviour. 3.6 Data minimisation Human Origin does not routinely collect or retain creators' source files. During live reviews, the creator shares their screen and files remain on the creator's device. No recording is made by the reviewer. Where a creator submits a screen recording, the recording is transmitted through an accepted encrypted or expiring channel and is deleted within seven days of the decision. 3.7 A standard for each discipline A Creator Charter defines permitted and prohibited uses of AI for each covered discipline. Examples include: Discipline Example of permitted use Example of prohibited use Writing Spelling correction AI-drafted text Illustration Non-generative production tools Generative image creation Photography Noise reduction Generative fill Music Conventional production/editing tools Generative composition Video Conventional editing Generative creation of substantive footage The complete discipline-specific standards are part of the Human Origin pilot documentation. 4. Review methodology 4.1 Application A creator applies by email and provides: a link to the work to be reviewed; and the websites and accounts on which the creator intends to display the badge. This information determines the certification scope, including the work covered, recorded on the public verification page. 4.2 Review modes Two review modes are proposed. 4.2.1 Live review The creator participates in a video call of approximately 30 minutes. During the call, the creator shares their screen and opens relevant files and production evidence. Nothing is sent to the reviewer and the session is not recorded by the reviewer. 4.2.2 Screen-recording review The creator makes a screen recording showing the relevant files and production evidence. Immediately before recording, the reviewer sends a one-time code. The creator displays or enters the code during the recording, providing evidence that the recording was produced specifically for the review. Accepted transmission channels include: Signal disappearing messages; WhatsApp disappearing messages; Threema disappearing messages; Telegram Secret Chats; expiring Proton Drive links; and expiring SwissTransfer links. The review method used is recorded on the public verification page. 4.3 Evidence by discipline Discipline Typical evidence examined Writing Document version history, drafts, outlines and notes Illustration Layered source files such as PSD or Procreate files, sketches and timelapse recordings Photography RAW files with EXIF data and contact sheets of the series Music DAW project files, separate stems and raw recordings Video Raw footage and editing project files The evidence examined is recorded in summary form on the verification page rather than published in full. 4.4 Deci