Current version
AI FREE Certification Standard
Standards are versioned. Every certificate records the standard version used during its evaluation, and requirements are never silently changed for existing certificates.
What the standard covers
- Definitions
- Certification categories
- Evidence requirements
- Permitted uses of technology
- AI disclosure
- Evaluation methodology
- Human review
- Certification decisions
- Appeals and reassessment
- Suspension
- Revocation
- Expiration and renewal
- Public verification
- Misuse of certification marks
Each certification has its own criteria. The standard makes no promise of absolute AI detection: it defines what evidence is required, how it is evaluated and how a decision can be reviewed, suspended or revoked.
AI FREE Certification Standard (AFCS-1.0)
1. Purpose and scope
AFCS-1.0 is the operational certification standard of AI FREE. It defines how a declared human origin, human intellectual and creative contribution, and any involvement of artificial intelligence are assessed for a specific object within a specific scope. AI FREE does not present itself as an AI detector. No technology can determine authorship with absolute certainty; certification is based on the combination of the applicant declaration, evidence, source files, creation history, metadata, process documentation, automated signals, contextual analysis and human review. The absence of evidence of AI is not by itself proof of human authorship.
2. Normative status and legal notice
AFCS-1.0 is a private certification policy published by AI FREE. It does not claim ISO accreditation, government recognition, regulatory authority or official governmental certification. Where mandatory consumer or sector legislation applies in a given jurisdiction, that legislation prevails and specific legal review may be required.
3. Definitions
Object: the specific work, product, service, content, organization or process submitted. Scope: the boundaries of what is certified. Generative AI: systems that produce new content. Conventional tooling: deterministic or non-generative software. Material involvement: a contribution that affects the certified intellectual or creative core, by function and importance rather than quantity alone. Certification category: what was certified. Certificate status: the current administrative state of an issued certificate. The two must not be confused.
4. Certification categories
AFCS-1.0 works with four independent categories, not hierarchical levels: 100% HUMAN, HUMAN CREATED, HUMAN CREATED - AI ASSISTED and HUMAN VERIFIED. Each answers a different question about the object and its scope. Assessment may also produce administrative outcomes: APPROVED, ADDITIONAL EVIDENCE REQUIRED, INSUFFICIENT EVIDENCE, NOT ELIGIBLE or REJECTED. Issued certificates carry one of the statuses ACTIVE, UNDER REVIEW, SUSPENDED, REVOKED or EXPIRED.
5. 100% HUMAN
Created without material generative AI involvement. The intellectual core, the creative core and the materially creative execution must be human. Conventional digital tooling is permitted: word processors, spreadsheets, digital cameras, traditional editing software, conventional colour adjustment, crop, resize, compression, format conversion, conventional spell checking, accessibility tools, deterministic automation, compilers, IDEs without material content generation, traditional filters and technical processing that does not create materially relevant new content. The mere presence of AI features in a tool does not disqualify a work; what matters is which feature was used and what function it performed. Material generative use inside the certified scope - generated paragraphs, illustrations, scenes, music, voice, material code, generative fill creating new elements, generated characters, or generated arguments, conclusions or intellectually relevant structures substantially incorporated into the work - normally makes the object ineligible for this category. A simplistic rule such as 'used a chatbot, therefore not 100% HUMAN' must not be applied: asking an AI how to operate a tool or where to find documentation is not the same as using AI to create the work.
6. HUMAN CREATED
The intellectual and creative core of the work is demonstrably human: essential conception, central intellectual decisions, creative structure, direction, substantial authorship and fundamental expressive decisions. Advanced technology may exist in the process, but technology must not replace the intellectual or creative core being certified as human. HUMAN CREATED does not necessarily mean that no AI was used at any moment; that stricter condition belongs to 100% HUMAN.
7. HUMAN CREATED - AI ASSISTED
Original intellectual creation is human, with declared and bounded AI assistance. The creation, the direction and the main intellectual and creative decisions are human; the use of AI is declared; the role of AI is identifiable; the contribution of AI has verifiable limits; the human retains substantial editorial and creative control; and AI supports execution without becoming the predominant author of the certified core. Assistance is assessed qualitatively, not only quantitatively. This category is not granted merely because a human wrote the prompts. Prompting alone does not prove human authorship of the result: describing a novel, a painting, an application, a song or a scientific article and having the system produce essentially the whole output does not make the result a certifiable human creation, even when the idea, the prompts and the selection among outputs were human. The reviewer must assess who made the materially relevant intellectual and creative decisions.
8. HUMAN VERIFIED
Independent verification of human identity, origin, responsibility, supervision or involvement associated with a product, service, organization, content or process. This category has a different purpose from the other three and applies when the object of certification is primarily human identity, human responsibility, human supervision, origin, human participation, human control of a defined process, or organizational accountability. It must not be used to bypass a work that does not meet the requirements of HUMAN CREATED or HUMAN CREATED - AI ASSISTED. The certificate scope must state exactly what was verified.
9. Scoped certification
A certification applies only to the object and scope identified in the certificate. Certifying the text of a book does not automatically certify its cover, translation, audiobook, marketing or illustrations. Certifying software does not automatically certify the website, advertising, documentation or promotional images. Certifying a product does not automatically certify its advertising campaign. Scope is mandatory and is published on the certificate.
10. Composite works
A work may contain components of different origins, for example human text with an AI-generated cover. Rather than denying certification outright, AI FREE may certify a defined scope such as 'text content only'. The public registry must state that limitation clearly, and the seal must not be used in a way that suggests the excluded components were also certified.
11. Evaluation model
Every application is assessed against twelve dimensions: type of work; human origin; human intellectual contribution; human creative contribution; AI involvement; role of AI; creative impact; human control; disclosure; evidence quality; provenance; and materiality of the technological contribution to the certified object.
12. Creative impact scale (T0-T3)
T0 technical or non-creative: compression, conversion, organization, crop, resize, compilation, deterministic processing. T1 minor assistance: the tool helps without creating a materially relevant contribution to the core, such as minor suggestions, technical help, certain linguistic corrections or auxiliary organization. T2 material assistance: AI contributes materially to parts of execution or expression while the intellectual and creative core remains demonstrably human; a possible candidate for HUMAN CREATED - AI ASSISTED. T3 generative core: AI produces, determines or replaces a substantial or essential part of the intellectual or creative core, such as central text, a central character, an argument, an intellectual conclusion, the main composition, the main image, essential design, intellectual architecture, a core algorithm or fundamental scenes; normally not eligible for HUMAN CREATED or HUMAN CREATED - AI ASSISTED with respect to the resulting work as a whole. 'Substantial' does not mean quantity alone: a small element may be creatively essential.
13. Materiality principle
Percentages are auxiliary signals, never the main criterion. If AI generated five per cent of a novel but that five per cent is the title, the main character, the plot twist or the conclusion, its impact is materially greater than five per cent. If AI generated five per cent of a codebase but that code is the core algorithm that differentiates the product, its intellectual impact may be high. Assessment combines quantity, function, importance, originality and creative dependency.
14. Evidence classes (E1-E4)
E1 declarative: forms, declarations, process descriptions. E2 supporting: screenshots, documents, references, auxiliary records. E3 provenance: source files, version history, commits, EXIF, project files, drafts, timestamps, raw files. E4 strong provenance: a consistent set of independent evidence demonstrating temporal progression and authorship or process. No single piece of evidence is treated as universally infallible. Where evidence is insufficient to support the requested category, the reviewer requests additional evidence or records INSUFFICIENT EVIDENCE. In material doubt, human authorship is not presumed.
15. Metadata
Metadata are signals, not absolute proof. Metadata can be removed, modified, lost, rewritten, inconsistent or incomplete. Metadata are always assessed together with other evidence.
16. Automated AI detectors
Detector results are automated signals. They are never treated as proof of AI and never as proof of human authorship, and they can never be the only basis for approval, rejection, suspension or revocation. Each use records the tool, the version when available, the date, the result and the known limitations. A single detector result does not produce an automatic rejection.
17. Provenance
Significant weight is given to the evolution of the work: for a book, outline to drafts to revisions to final; for software, requirements to commits to branches to tests to releases; for photography, raw to edits to export; for design, sketch to layers to revisions to final; for music, sessions to tracks to stems to mix to master; for video, raw footage to project timeline to edits to export. Process evidence is usually more valuable than isolated analysis of the final file.
18. Decision matrix and decision tree
The decision matrix combines type of work, AI technology, role of AI, creative impact, human control, disclosure and evidence, and produces an eligible category or an administrative decision. The decision tree is: identify what is being certified; define the certification scope; identify the claimed human contribution; determine whether generative AI was used; if it was, determine its role; classify the AI contribution as T0, T1, T2 or T3; evaluate the human intellectual and creative core; evaluate evidence and provenance; check inconsistencies; determine eligibility; and record a reasoned human decision. No automated rule issues a final certificate where the category policy requires human review.
19. Rules by type of work
Text, books and articles: 100% HUMAN requires text substantially written by the human without material generative text; word processors, conventional spell checking, formatting, references, bibliographic tools and non-generative automation are normally permitted; clarity suggestions, editorial help, auxiliary brainstorming, certain rewordings, research support, organization and assisted translation may fall under AI ASSISTED when declared and when the intellectual core remains human and demonstrable. A text originally produced by AI and later edited by the applicant is not certified as human creation on that basis alone. Journalism: investigation, fact gathering, interviews, source selection, interpretation, writing, editing and automatic generation are assessed separately; AI used for transcription or organization differs from AI used to write the article; certification never implies that the reported facts are true. Academic work: hypothesis, methodology, research, analysis, interpretation, results, writing, code and references are assessed separately; AI FREE does not certify scientific validity, ethical approval, absence of plagiarism, academic correctness or methodological quality. Software: idea, requirements, architecture, UX, logic, algorithms, code, tests and documentation are assessed separately and assistants are judged by function; trivial autocomplete does not carry the same weight as generation of complete functions, modules, algorithms, architecture or business rules; 100% HUMAN excludes material AI code generation in the certified scope, while AI ASSISTED may be eligible when AI supports parts of implementation, debugging or explanation and architecture and essential decisions remain human and demonstrable. Photography: conventional editing such as crop, exposure, contrast, white balance, colour correction, conventional sharpening and limited non-generative technical removal is distinguished from generative editing such as created objects, generative background replacement, generated people, material generative expansion, generative fill and significant creative reconstruction; a human-captured photograph may cease to be eligible for 100% HUMAN for the final image when generative AI creates material elements of the composition. Illustration and design: sketch, concept, composition, shapes, typography, drawing, rendering, generation and post-production are assessed separately; a human sketch followed by almost entirely AI-generated imagery is not automatically HUMAN CREATED; purely technical upscaling may be T0 or T1, while generative upscaling that reconstructs or invents materially relevant detail may be T2. Video: script, filming, direction, performance, editing, effects, voice, music, generated scenes, B-roll, subtitles and translation are assessed separately; automatic subtitles do not carry the same weight as a generated scene; where generative elements are used, the record states what, where, why, how much and the creative role. Music and audio: composition, lyrics, performance, arrangement, recording, mixing, mastering, voice, instruments and effects are assessed separately; AI-assisted technical processing is not automatically equivalent to generated composition, while AI-created melody, lyrics, lead vocal, essential arrangement or material composition is assessed as a material creative contribution. Translation: the original work and the translation are distinct objects; a 100% HUMAN work may have an AI-assisted translation, and this does not make the translation 100% HUMAN; the certificate states the object and scope. Products: concept, engineering, design, manufacturing, packaging, marketing and documentation are assessed separately; a human-created product design is not automatically altered because an advertising campaign uses AI imagery, but the campaign must not use the seal in a way that suggests those AI images are certified as human. Companies and services: an entire company is not certified generically as 100% HUMAN when it uses AI systems across various activities; the certification must have a defined scope, such as a specific advisory or customer process, and HUMAN VERIFIED is often the relevant category for human service, human review, responsibility, supervision, decision-making and professional delivery.
20. AI assistance disclosure
For HUMAN CREATED - AI ASSISTED, the internal record captures the AI tools, the purpose, the stage of the process, the nature of the assistance, the materiality, the human control and the reviewer assessment. A summarised public statement may be shown, such as 'AI assistance was declared and reviewed within the scope of this certification', optionally with a short description of the function where that supports transparency.
21. Inconsistencies and undisclosed AI
Where the declaration and the evidence diverge, the application is not rejected automatically; a clarification is required. If a material inconsistency remains without a satisfactory explanation, the outcome is ADDITIONAL EVIDENCE REQUIRED, INSUFFICIENT EVIDENCE or REJECTED, as appropriate. Where relevant evidence of undisclosed AI use appears, the application moves to review or clarification and the applicant may respond. If material use incompatible with the requested category is confirmed, the reviewer may recommend a different category, deny the requested category, or reject the application in the case of a materially false declaration under the applicable policy. An application is not converted to another category without the applicant knowing and accepting the change when this alters the public claim of the certificate.
22. Human review and reviewer override
Automation may recommend, AI may support the analysis and the system may compute signals, but the final decision allows a reasoned human review. Any override of an automated recommendation records the reviewer, the reason and the date. Every decision must be justifiable and auditable.
23. Explainability
Every negative decision must have an understandable reason. Statements such as 'AI detected' are not acceptable. Decisions are expressed in terms such as 'the submitted evidence was insufficient to establish the claimed human-origin requirements under AFCS-1.0' or 'declared generative AI involvement materially affected the creative scope covered by the requested certification'.
24. No false precision
AI FREE does not publish scores such as '97% human' or '3% AI' as if they were scientifically established measurements. Internal risk and confidence indicators may exist, but they are never presented as an objective measurement of authorship. Public and internal reporting uses documented qualitative scales for evidence strength, AI involvement, creative impact, human control and provenance confidence.
25. Non-endorsement
AI FREE certification verifies only the scope defined by the applicable certification standard. It does not constitute an endorsement of the quality, accuracy, legality, safety, originality, scientific validity, commercial value or ethical merit of the certified work unless explicitly stated otherwise.
26. False declarations, suspension and revocation
A materially false declaration may result in rejection, suspension or revocation, according to the stage and the applicable policy. An isolated technical inconsistency is not automatically classified as fraud; a finding of fraud requires human analysis and sufficient evidence.
27. Public registry and privacy of evidence
The public registry shows the certificate ID, status, certification category, certified object, certificate holder, scope, issue date, expiration or renewal date where applicable, standard version, public AI disclosure where applicable and verification status. Source files, private prompts, confidential documents, private evidence and unnecessary personal information are never published. Submitted evidence is private by default, access-controlled, not enumerable by anonymous users, not indexed by search engines and subject to the applicable retention policy.
28. Versioning and existing certificates
AFCS-1.0 is the initial operational version. Editorial corrections that do not change requirements keep the version and are recorded as a documentary revision. Changes that alter interpretation or requirements result in a new version such as AFCS-1.1 or AFCS-2.0 under the formal versioning policy. Every certificate records the standard version under which it was assessed, and new criteria are never applied retroactively and silently. No existing certificate changes category, loses validity, is suspended or is revoked merely because the standard was deepened; any re-evaluation follows a formal process.
29. Language of the standard
English is the normative version of AFCS-1.0. Portuguese is provided as an official operational translation. In case of conflict between translations, the English version of AFCS-1.0 controls, except where mandatory local legislation requires otherwise, in which case jurisdiction-specific legal review applies. Other languages are served in English until a reviewed translation is published.