# FAIR AI Attribution (FAIA)
## The FAIA Project
**FAIA – FAIR AI Attribution** – is an open framework for structured, machine-readable, and verifiable disclosure of AI involvement in digital content creation. The project is developed by [**Liccium**](https://liccium.com/) in collaboration with [**Leiden University**](https://www.universiteitleiden.nl/en) and the [**GO FAIR Foundation**](https://www.gofair.foundation/) in response to the growing need for transparency, provenance, and accountability in digital publishing, research communication, and media production.
As generative AI systems become increasingly integrated into writing, editing, design, publishing, and content production workflows, it is becoming progressively more difficult to distinguish between human-created, AI-assisted, and AI-generated material. At the same time, existing disclosure approaches are fragmented, inconsistent, and often technically fragile. Metadata may be removed during distribution, platform-specific labels rarely persist across environments, and there is currently no widely adopted infrastructure for interoperable and verifiable AI attribution.
FAIA addresses this problem by providing a shared attribution vocabulary together with technical mechanisms for persistent and verifiable declarations. The framework enables creators, publishers, researchers, platforms, and AI providers to disclose whether and how AI systems contributed to the creation or modification of digital content.
## **Why is FAIA necessary?**
The rapid growth of AI-generated content creates significant challenges for digital ecosystems:
* **reduced trust** in digital information and media authenticity
* difficulties in documenting **editorial and research integrity**
* increasing **regulatory transparency requirements**
* challenges for **provenance verification** and reproducibility
* contamination risks for future AI training datasets
These challenges are increasingly recognised in policy and regulatory frameworks, including the European Union AI Act, which introduces transparency obligations for AI-generated and AI-manipulated content.
FAIA contributes to addressing these challenges by enabling AI involvement to be disclosed in a way that is:
* **machine-readable**
* **interoperable** across systems and platforms
* **cryptographically verifiable**
* **persistently linked** to digital content
* resolvable independently of individual platforms or publication environments
## Goals and Scope
The FAIA framework provides a practical and implementation-independent foundation for AI attribution across digital media ecosystems. The framework is designed to work across text, images, audio, video, datasets, and mixed-media workflows.
Its core goals are to:
* enable **transparent disclosure** **of AI involvement** in content creation and modification
* support **interoperable attribution** across platforms, workflows, and media types
* strengthen **provenance, accountability, and trust** in digital publishing and research environments
* **support emerging transparency and compliance requirements**
* provide infrastructure for downstream services such as verification, moderation, dataset filtering, and provenance analysis
## Technical Foundation
FAIA combines semantic metadata, persistent identifiers, digital signatures, and interoperable registry infrastructure to support durable AI attribution workflows.
Within the reference implementation developed in the FAIA project ecosystem:
* **FAIR AI attribution metadata** is expressed in machine-readable form using semantic web standards
* declarations are **persistently linked to content through ISCC fingerprints**
* declarations may be **digitally signed** using certificates or verifiable credentials
* declaration records may be published and resolved through **interoperable registry infrastructure**
* APIs and semantic vocabularies **support integration** into external applications and publishing workflows
The framework is implementation-independent and does not depend on a single platform, registry operator, or software vendor. Third parties are encouraged to integrate FAIA-compatible workflows and services using the published vocabularies, APIs, and interoperability mechanisms.
## Current Implementations
Current prototype implementations developed within the FAIA project ecosystem include:
* the FAIA **Statement Generator:** \
* **machine-readable attribution vocabularies and ontologies**:\
* openly documented **declaration APIs**\
[https://dev.liccium.com](https://dev.liccium.com/)
* **registry explorer** infrastructure\
[https://faia.io](https://faia.io/)
* **trust infrastructure** based on digital signatures and verifiable credentials\
[https://creatorcredentials.com](https://creatorcredentials.com/)
Work is also ongoing on user-facing applications and federated registry infrastructure supporting broader ecosystem participation and interoperable deployment across independent platforms and services.
## Outcomes and Impact
FAIA aims to support more transparent and accountable digital publishing and AI ecosystems by:
* enabling **creators and organisations to disclose AI involvement** consistently and verifiably
* supporting **provenance and integrity workflows** across publishing and research environments
* helping platforms and downstream services **interpret attribution metadata programmatically**
* supporting **future verification, moderation, and dataset management** workflows
* reducing fragmentation across AI transparency approaches and metadata systems
While the framework is intended as a cross-sector infrastructure applicable to all forms of digital media, early implementations and collaborations have focused particularly on academic publishing, research communication, and professional publishing environments.
FAIA contributes to the development of open and interoperable infrastructure for AI attribution that remains portable across platforms, resilient across distribution environments, and accessible across different technical and institutional ecosystems.
# Vocabulary
The FAIA Vocabulary provides a structured, machine-readable model for describing AI involvement in digital content creation. Rather than relying on a single label, FAIA combines complementary vocabulary elements that describe the overall nature of a work, the role AI played during its creation, and the AI systems involved.
The vocabulary is designed to support interoperable AI attribution across different media types, workflows, publishing environments, platforms, and technical systems.
FAIA semantic resources are published as open vocabularies and ontologies using persistent identifiers and Semantic Web standards, including JSON-LD or RDF.
Together, these components support transparent, consistent, and interoperable disclosure of AI involvement across media types, workflows, and application domains.
### FAIA Flags
FAIA Flags provide a high-level classification of the resulting content. Every FAIA declaration contains exactly one Flag describing whether the work is Human-Created, AI-Assisted, or AI-Generated.
The Flags communicate the overall degree of AI involvement while remaining independent of the specific tools or workflow used.
[Learn more about FAIA Flags →](/technology/vocabulary/faia-flags)
### Activity Codes
Activity Codes describe how AI contributed during the content lifecycle. Unlike the FAIA Flags, multiple Activity Codes may be associated with a single declaration to describe different forms of AI involvement.
FAIA defines a set of generic Activity Codes – Co-Creation, Contribution, Enhancement, Refinement, Transformation, and Analysis – that apply across media types. It also supports the integration of domain-specific activity vocabularies, allowing communities to adopt established classifications within the broader FAIA framework. The first supported example is the STM AI Classification for scholarly publishing.
[Learn more about Activity Codes →](/technology/vocabulary/activity-codes)
### System Attribution
System Attribution records the AI system used during content creation or modification and, where available, the version involved. Identifying the AI system provides additional transparency, reproducibility, and context, particularly as different systems and releases may differ substantially in capabilities and behaviour.
[Learn more about System Attribution →](/technology/vocabulary/system-attribution)
### Relationship Between the Vocabulary Components
The three vocabulary components complement one another:
* FAIA Flags answer: "What is the overall nature of this content?"
* Activity Codes answer: "How did AI contribute during its creation?"
* System Attribution answers: "Which AI system and version were used?"
Together they provide a structured vocabulary for describing AI involvement in a way that is understandable to both humans and machines.
# FAIA Flags
FAIA Flags provide a high-level classification of AI involvement in digital content. Every FAIA declaration carries exactly one Flag, describing the overall nature of the resulting work no matter how many AI systems or activities were involved in producing it.
The Flags offer a simple, consistent way to communicate whether content was created entirely by humans, produced with meaningful AI assistance, or generated predominantly by AI. They focus on the resulting work rather than the individual production steps. More detailed information about AI involvement is expressed separately through Activity Codes and System Attribution.
### Human-Created Content (HCC)
Content created, generated and edited exclusively by humans or human-controlled instruments. While digital tools such as word processors, image editors, or audio software may be used, no generative AI systems are involved at any stage of the creative or editorial process.
### AI-Assisted Content (AAC)
Content where a human or a human-controlled instrument remains the primary creator and AI systems contributed during the creation process to various degrees. This may include AI-generated input that humans or human-controlled instruments accept or reject, generation of content fragments that humans or human-controlled instruments integrate into larger works, or refinement steps performed under direct human supervision and editorial control.
### AI-Generated Content (AIG)
Content generated predominantly or entirely by an AI system, where the AI serves as the main creative agent. Human input is limited to initiating prompts, selecting among AI-generated outputs, or making minor adjustments that do not materially alter structure, substance, or expressive intent. The resulting content is accepted largely as produced by the AI, with no substantive human editing or creatorship.
## Design Principles
The three FAIA Flags are mutually exclusive, so the overall classification stays unambiguous. They deliberately remain high-level and do not identify individual AI-supported activities or specific AI systems. Those are represented separately through Activity Codes and System Attribution, letting a declaration combine a simple overall classification with more detailed workflow and technical information.
Together, the three vocabulary components provide a structured, interoperable, and machine-readable framework for describing AI involvement across media types, sectors, and workflows.
One optional point you could add, if you want the flag-to-activity relationship signposted here: in the vocabulary, Transformation and Analysis are reserved for AIG. That nuance might belong on the Activity Codes page instead, to keep this section purely high-level. Your call.
# Activity Codes
Activity Codes describe how AI systems were used during the content lifecycle. While the FAIA Flags classify the overall nature of the resulting content, Activity Codes describe the structural role AI played within the workflow.
Unlike the FAIA Flags, multiple Activity Codes may be associated with a single declaration. This allows AI involvement to be described at a more granular level by capturing the different ways AI contributed during the creation, modification, or analysis of content.
## FAIA Activity Codes
### Co-Creation
AI and humans jointly produce content through an interactive process. The AI generates substantial parts of the content based directly on human-created material and structure. The resulting expressive contributions are interwoven and not cleanly separable.
### Contribution
AI produces discrete content elements that are incorporated into a primarily human-created work. The human determines the overall structure, intent, and final form of the resulting output.
### Enhancement
AI modifies or extends existing human-created content to improve quality, clarity, richness, completeness, or structure while preserving the original expressive intent and authorship of the source material.
### Refinement
AI performs limited adjustments that preserve the existing structure, meaning, expressive intent, and authorship of the content.
### Transformation
AI converts existing content into another form, modality, or representation while preserving semantic equivalence or maintaining a close relationship to the original content.
### Analysis
AI interprets existing content to derive structure, features, metadata, meaning, or analytical insights.
### Summary
| Code | What the human provides | What the AI provides |
| -------------- | ----------------------------------------- | ------------------------------------------------------------ |
| Co-Creation | Core ideas, structure, or source material | Independent sections, elements, or content developed jointly |
| Contribution | Overall structure and intent | Discrete components integrated by the human |
| Enhancement | Existing content | Extensions or quality improvements |
| Refinement | Existing content | Minor surface-level or technical adjustments |
| Transformation | Prompt or task definition | Autonomous re-expression into another form |
| Analysis | Prompt or analytical instruction | Automated extraction, classification, or metadata generation |
## Domain-Specific Activity Codes
FAIA provides a generic, cross-domain set of Activity Codes that can be applied consistently across different media types. At the same time, it is designed to complement existing sector-specific standards and metadata ecosystems where appropriate.
Rather than replacing established classifications, FAIA allows domain-specific activity vocabularies to be incorporated into FAIA declarations, enabling interoperability while preserving terminology that is familiar within individual communities.
### STM AI Classification
The STM AI Classification is the first domain-specific activity vocabulary supported by FAIA. Developed by the STM Association, it provides a structured classification of AI use during the preparation of scholarly publications.
Unlike the generic FAIA Activity Codes, which apply across all media types, the STM vocabulary describes specific activities within academic research and publishing workflows. FAIA supports these classifications as machine-readable Activity Codes that can be included alongside FAIA Flags and System Attribution within the same declaration.
{% embed url="" %}
#### STM Language Refinement
AI tools are used to suggest or apply refinement, correction, editing, or formatting of manuscript text in order to improve clarity of language.
#### STM Manuscript Drafting
AI tools are used to generate rewrite, or draft part or all of the manuscript text.
#### STM Manuscript Translation
AI tools are used to assist translation of an author's original work into another language.
#### STM Data Presentation Refinement
AI tools are used to refine or format, or improve the presentation, clarity, or readability of the data reported in the manuscript or associated materials.
#### STM Illustrative Figure Generation
AI tools are used to generate, refine, correct, edit, or format images, diagrams, or other figures for illustrative or aesthetic purposes only, not as research data or research results.
#### STM Research Data Visualisation
AI tools are used to generate, refine, correct, edit, or format visualisations of research data or research results for inclusion in the manuscript.
#### STM Code Presentation Refinement
AI tools are used to refine or format code reported in the manuscript or associated materials in order to improve readability or clarity, without altering the functionality of the code.
#### STM Reference Gathering
AI tools are used to suggest references that may be relevant for inclusion in the manuscript's reference list.
Multiple STM Activity Codes may be associated with a single publication, allowing AI involvement to be described across different stages of the research and publication workflow.
### IPTC Digital Source Types
FAIA can also be used alongside the IPTC Digital Source Type vocabulary for photography and news media, allowing existing provenance classifications to be incorporated into FAIA declarations where appropriate.
## Design Principles
Activity Codes describe the role AI played during the content lifecycle rather than the overall nature of the resulting work. Every declaration may contain one or more Activity Codes, depending on the complexity of the workflow.
Together with the FAIA Flags and System Attribution, Activity Codes provide a structured and interoperable vocabulary for describing AI involvement across media types, sectors, and workflows.
# System Attribution
System Attribution identifies the AI system involved in the creation or modification of digital content. While the FAIA Flags describe the overall nature of the resulting work and Activity Codes describe how AI contributed during the workflow, System Attribution records which AI technology was used.
Identifying the AI system adds transparency, reproducibility, and context. Different AI systems, and different releases of the same system, can vary substantially in capabilities, training data, behaviour, limitations, safeguards, and licensing conditions. Recording this information allows AI involvement to be interpreted more accurately.
FAIA defines two complementary attribution fields: System and Version.
## System
The System field identifies the AI system, model, service, or model family used to generate or modify content. Where possible, the provider's official designation should be used.
Examples include:
* `ChatGPT`
* `Claude`
* `Google Gemini`
* `DALL·E`
* `Stable Diffusion 3`
* `Midjourney`
## Version
The Version field identifies the specific release or version of the AI system used. Where known, the designation should follow the naming or numbering scheme used by the system provider.
Examples include:
`5.2`\
`Sonnet 4`\
`2.5 Pro`\
`v1.3.2`\
`Beta-2`\
`2025-10-release`
## Why it matters
System and Version together support transparency, reproducibility, regulatory disclosure, rights management, and informed interpretation of AI-involved content. The version in particular provides precise context about system behaviour, capabilities, limitations, and known issues at the time of use, since different releases of the same system may differ substantially in performance, training data, safeguards, or output characteristics.
## Design Principles
System Attribution identifies the AI technology involved in the content lifecycle. It does not describe the degree of AI involvement or the role AI played during creation, which are represented separately through the FAIA Flags and Activity Codes.
Every System Attribution may contain a System value and, where available, a Version value. Together they provide a consistent, machine-readable way to identify the AI technology used, complementing the broader FAIA attribution framework.
# Semantic Resources
FAIA publishes its core semantic resources as open, machine-readable assets for interoperable AI attribution across media types, platforms, and technical systems.
## Ontology and vocabulary
The FAIA ontology defines the properties used to describe AI involvement in a declaration: the FAIA flag, the activity code, the system attribution, and the system version. It imports the Liccium ontology for the underlying Declaration class.
The FAIA vocabulary defines the controlled terms used in declarations, modelled as SKOS concepts. It covers three layers: the three FAIA flags (HCC, AAC, AIG), six generic activity codes, and eight STM-specific activity codes for academic manuscript preparation.
## FAIR by design
The framework follows the FAIR principles of Findability, Accessibility, Interoperability, and Reusability, and was developed in collaboration with the GO FAIR Foundation. The ontology and vocabulary are published in Turtle through persistent w3id.org identifiers, so terms resolve to stable definitions independently of any single platform.
## Resources
FAIA ontology:
FAIA vocabulary:
## Implementation independence
FAIA defines a common semantic model without prescribing a specific implementation. The ontology and vocabulary can be used in publishing systems, repositories, registries, provenance frameworks, and verification services.
# FAIA Reference Implementations
FAIA supports different levels of implementation depending on the required level of persistence, verification, and infrastructure integration.
## FAIA Statement Generator
The FAIA Statement Generator provides a lightweight mechanism for creating portable AI attribution statements using the FAIA vocabulary.
The generated statements can be embedded into websites, attached to publications, or shared alongside digital content. They support both human-readable and machine-readable disclosure of AI involvement.
However, these statements are not persistently bound to the content and are not independently verifiable. Without cryptographic signing, registry publication, and persistent content–metadata binding, there is no reliable mechanism to verify authorship, integrity, or long-term association with the underlying asset.
## Liccium Reference Implementation
The Liccium platform provides a reference implementation of persistent and verifiable FAIA declarations.
The implementation combines:
* ISCC-based content fingerprinting
* machine-readable JSON-LD declaration metadata
* cryptographic signatures and verifiable credentials
* declaration APIs
* federated registry infrastructure
Using the Liccium implementation, FAIA declarations can be:
* created through applications or APIs
* digitally signed by the declarant
* persistently linked to digital content through ISCC identifiers
* published to interoperable registry infrastructure
* independently resolved and verified across platforms and workflows
The implementation includes the Liccium Declaration API, registry infrastructure, and user-facing applications supporting declaration creation, signing, publication, and verification workflows.
# Whitepaper
## Whitepaper
The FAIA whitepaper presents the full framework for FAIR AI Attribution: the vocabulary for disclosing AI involvement, the attribution model built on FAIA Flags, Activity Codes, and System Attribution, and the technical infrastructure for persistent, verifiable declarations bound to content through ISCC fingerprints.
It sets out the transparency problem FAIA addresses, the use cases and stakeholders it serves, and how declarations are created, signed, published, and resolved across federated registry infrastructure. It also positions FAIA in relation to related standards and initiatives, including C2PA, IPTC Digital Source Type, and the STM AI Classification.
The whitepaper is available as a pre-publication on Zenodo.
### Read the whitepaper
The FAIA Framework: A machine-readable vocabulary for AI attribution\
{% embed url="" %}
### How to cite
Le Dévédec, S., Gambardella, A. A., Hettne, K., Posth, S., & Schultes, E. (2026). The FAIA Framework: A machine-readable vocabulary for AI attribution \[Whitepaper]. Zenodo.
The whitepaper is published under a Creative Commons Attribution-ShareAlike 4.0 International licence (CC BY-SA 4.0).
### Funding
This whitepaper is a deliverable under the Dutch Responsible AI in de praktijk programme, funded by SIDN fonds and Topsector ICT. See [Funding](/consortium/funding) for more about the programme and partners.
# Press
FAIA (FAIR AI Attribution) is an open framework for structured, machine-readable, and verifiable disclosure of AI involvement in digital content. This page provides a short project description, logos, overview graphics, and a media contact for journalists, researchers, and partners.
## Media contact
For interviews, background, or asset requests, contact .
## About FAIA
FAIA provides a shared vocabulary and technical model for disclosing whether and how AI systems contributed to the creation or modification of digital content. It defines three complementary elements: high-level attribution flags (human-created, AI-assisted, AI-generated), activity codes describing the role AI played in the content lifecycle, and optional system attribution identifying the AI system and version involved.
FAIA supports disclosure at different levels. At the lightweight end, the FAIA statement generator () lets creators and organisations produce portable transparency statements that are both human-readable and machine-readable, and share them through a URL, embed them on a website, or attach them to a publication, with no specialised infrastructure required. At the more advanced end, declarations can be issued and cryptographically signed by creators, rightsholders, institutions, or AI system providers, expressed as machine-readable JSON-LD, and validated through a trust model based on digital certificates and verifiable credentials.
Persistent declarations are bound to content through ISCC fingerprints (ISO 24138:2024) and published across a federated registry, so attribution remains resolvable even when files are copied, transformed, or stripped of metadata. FAIA supports transparency across publishing, journalism, research, and media production, and provides a foundation for compliance with emerging obligations such as the EU AI Act.
FAIA is developed by Liccium in collaboration with Leiden University and the GO FAIR Foundation.
## Short description
FAIA (FAIR AI Attribution) is an open, machine-readable framework for verifiable disclosure of how AI was involved in creating or modifying digital content. It is developed by Liccium with Leiden University and the GO FAIR Foundation, supported by SIDN fonds and Topsector ICT.
## Logos
The graphics are provided for editorial and informational use referring to the FAIA project and its partners. They are published under a Creative Commons Attribution-NoDerivatives 4.0 International licence (CC BY-ND 4.0), so they may be reproduced and shared unaltered, with attribution to FAIA (FAIR AI Attribution), but may not be modified, recoloured, cropped, distorted, rotated, or otherwise adapted. Please keep adequate clear space around them. Source files are available on request, and for other uses contact .
## FAIA Flags
## Overview graphics
These graphics may be reproduced unaltered in articles, presentations, and reports about FAIA, with attribution to FAIA (FAIR AI Attribution). Source files are available on request.
## Project partners
Leiden University (project lead) provided academic oversight and developed the ethical, legal, and scholarly foundations of the framework, with a focus on transparency, reproducibility, and research integrity. Contributors include Leiden University Libraries, Centre for Digital Scholarship.
The GO FAIR Foundation contributed expertise in semantic data modelling, metadata standards, and FAIR principles, helping ensure that the FAIA vocabulary and declarations are interoperable, machine-readable, and openly accessible.
Liccium developed the technical infrastructure, the applications, and provides a reference implementation of the FAIA framework within its content verification software.
## Funding
FAIA is a deliverable under the Dutch Responsible AI in de praktijk programme, a joint funding initiative by SIDN fonds and Topsector ICT supporting practical applications of trustworthy AI. See Funding for more.
## Resources
Whitepaper on Zenodo (pre-publication):
FAIA registry: [https://faia.io](https://faia.io/)
FAIA statement generator:
Documentation: [https://faia.liccium.com](https://faia.liccium.com/)
# FAIA Project Partners
The FAIA project (Fair AI Attribution) is a joint initiative that brings together expertise in data modelling, publishing standards, AI ethics, and verifiable declaration. Our consortium partners include:
## **Leiden University**
*Centre for Digital Scholarship*
Leiden University provides academic oversight and develops the ethical, legal, and scholarly foundations of the FAIA framework. Their focus is on transparency, reproducibility, and research integrity in the context of AI-assisted content creation.
Leiden University Libraries’ Centre for Digital Scholarship:\
## **GO FAIR Foundation**
The GO FAIR Foundation contributes deep expertise in semantic data modelling, metadata standards, and FAIR principles. It ensures that the FAIA vocabulary and declarations are interoperable, machine-readable, and openly accessible.
GO FAIR Foundation:\
[https://gofair.foundation](https://www.gofair.foundation/)
## **Liccium**
Liccium develops the technical infrastructure and implements the FAIA attribution framework into its content verification software. This includes browser plugins, APIs, and publishing tools that allow creators and platforms to document and verify AI involvement.
Liccium:\
[https://liccium.com](https://liccium.com/)
# Contact and Team
If you are interested in collaborating or piloting the FAIA framework, please get in touch with us via e-mail: or contact us directly:
## **Leiden University**
|
|
# Funding
The FAIA (Fair AI Attribution) project is made possible through support from the Dutch innovation program: [**Responsible AI in de praktijk**](https://www.sidnfonds.nl/nieuws/call-responsible-ai-in-praktijk)**.**
This joint funding initiative by SIDN fonds and Topsector ICT supports practical applications of trustworthy AI that promote transparency, prioritise human oversight and control, and deliver positive societal impact. FAIA was selected for its contribution to verifiable AI disclosures in publishing and digital media.
## SIDN Fonds
SIDN fonds logo
SIDN Fonds: \
## Topsector ICT
Topsector ICT logo
Topsector ICT:\
# Presentations
## Download the FAIA 1-pager
{% file src="/files/5xG77PD67XdnLV6NRViW" %}
FAIA 1-pager
{% endfile %}
## March 2026
FDO Forum Conference, TU Vienna, March 27 2026,
## November 2025
**A Fair and Verifiable AI Preference System – A Registry-Based Framework for Transparency, Control, and Provenance**\
November 05, 2025, Seoul, Korea, ICOTEC,
## **October 2025**
**FAIR AI Attribution – Flagging Synthetic Media Using Open Standards**\
WIPO Frontier Conversation 12 – IP and Synthetic Media\
October 28-29, 2025
Video Recording of the presentation: \
## September 2025
**Kick-off Event, Responsible AI in de praktijk**\
September 24, 2025, Utrecht
## The FAIA Pitch
{% embed url="" fullWidth="false" %}
Video Pitch
{% endembed %}
This short video introduces the FAIA initiative. It highlights the problem of AI opacity, emerging regulation like the EU AI Act, and a practical solution for certifying content with verifiable AI involvement.
# FAIA statement
FAIA: AAC; Enhancement; Claude, Opus 4.8
This website contains AI-assisted content. AI was used for enhancement. The AI system used was Claude (Opus 4.8). The human author remains the primary creator. Declared under the FAIA framework (faia.io).
