Responsible Use of Artificial Intelligence

Our Approach

PARRIIC uses artificial intelligence tools to enhance the quality, efficiency, and effectiveness of our work. We believe that AI, when used responsibly and transparently, can help us deliver better outcomes for our clients while maintaining the highest professional standards.

Primary Platform: Lumo by Proton.
Secondary Platform: Claude by Anthropic
Translation: DeepL Translator by DeepL SE

Why We Use AI

We use AI tools to:

Draft and refine content: Creating drafts and translations of written materials.
Edit and improve clarity: Refining language for accessibility and readability.
Research and analysis: Gathering information and synthesising complex topics.
Technical problem-solving: Developing and working through solutions, particularly for digital platform challenges.
Efficiency: Completing routine tasks more quickly so we can focus on strategic and specialised work.

AI helps us work better and faster, but it does not replace our professional judgment, expertise, or accountability.

AI Platforms we use

Lumo through Proton

We prioritise Lumo, an AI assistant developed by Proton, for day-to-day operations, content creation, analysis, and decision support.

Data Privacy & Security:

Zero-access encryption: All interactions with Lumo are end-to-end encrypted. Sensitive data is never used in our AI work, however if human error occurs and sensitive data is inadvertently submitted, the content remains encrypted and inaccessible to Proton or any third party.
No training on user data: Conversations are not used to train the underlying models.
GDPR compliance: Built to meet European data protection standards.

European Sovereignty:

EU-based infrastructure: Proton is headquartered in Switzerland with strong European data governance.
Alignment with EU AI Act: Designed to comply with emerging European AI regulations.
Reduced dependency on non-EU providers: Supports digital sovereignty objectives.

Long-Term Stability & Independence: Unlike many AI providers driven by venture capital and market speculation, Proton operates as an independent, sustainable business. This independence shields our AI usage from the volatility of the AI investment market, ensuring that our operations, commitment to safety, privacy, and European sovereignty remains consistent regardless of market trends or external financial pressures.

Operational Capabilities:

Multi-language support (11 languages)
File upload and analysis capabilities
Web search functionality (Lumo Plus)
Writing assistance and content generation
Secure chat history management

DeepL for Translation

For translation needs, we use DeepL, a European translation service developed by German company DeepL SE.

European Infrastructure and Data Sovereignty: DeepL processes all translations on infrastructure located in the European Economic Area, ensuring that data remains within European jurisdiction and benefits from EU data protection standards. This aligns with our commitment to data sovereignty and European digital infrastructure.

Privacy and GDPR Compliance: DeepL adheres to GDPR—one of the strictest data protection laws worldwide—and with their Pro service, texts submitted are processed only temporarily and are deleted after translation completion. Translations are never stored or used to train their systems when using the paid service.

Supporting Digital Diversity: We believe in supporting a diverse digital ecosystem rather than contributing to technology monopolies. By choosing European alternatives where they offer equivalent or superior functionality, we help sustain a more balanced and competitive technology landscape. This matters for long-term innovation, user choice, and maintaining alternatives to dominant platforms.

Quality for Heritage Work: DeepL’s neural network technology delivers high-quality translations, including complex, specialised terminology—which is important when working with cultural heritage materials and documentation that requires accuracy and nuance.

Anthropic/Claude

Limited Use for Specific Purposes:

While Lumo is our primary tool, we retain limited access to Anthropic’s Claude API for specific technical challenges where its current capabilities are superior, specifically:

Complex code generation and debugging
Advanced API integration development
Specialised technical documentation

Governance Framework:

Lumo is used in the first instance for all general AI tasks.
This usage is kept under constant review by the AI Ethics Lead to monitor capability evolution.
Claude API access requires approval from the AI Ethics Lead.
Quarterly reviews assess whether Lumo capabilities have reached parity.
All Claude usage must be documented with business justification.
If Lumo capabilities reach equivalence, Claude access will be discontinued.

Our continued use is justified by Anthropic’s unique approach to safety:

The primary reason for selecting Anthropic to fulfill these business needs is their Constitutional AI methodology. Unlike standard models trained solely on human feedback, CAI trains the system using a set of explicit, written principles (a “constitution”) derived from sources like the UN Declaration of Human Rights and safety best practices.

Transparency: This approach allows us to inspect and understand the specific principles guiding the AI’s responses, offering greater transparency than “black box” training methods.
Safety by Design: CAI helps the model self-correct and refuse harmful requests without needing human exposure to toxic content during training.
Ethical Alignment: The constitutional framework aligns with our values of inclusivity and respect for diverse perspectives.

Third-Party and Embedded AI Tools

Project-Specific Exceptions

We recognise that certain client environments or specific project requirements may necessitate the use of additional AI tools not listed above (e.g., Microsoft Copilot within Office 365, GitHub Copilot for development, or client-specific platforms).

Approval Criteria

The use of any additional AI tool is permitted on a project-by-project basis only if:

Client Requirement or Benefit: The tool is explicitly required by the client’s infrastructure or provides a distinct capability that our primary tools (Lumo/DeepL) cannot currently match.
Security & Safety Compliance: The tool meets PARRIIC’s minimum standards for:
- Data Sovereignty: Preference for EU-based or GDPR-compliant infrastructure.
- Data Privacy: Assurance that client data is not used for model training (or is anonymised/encrypted).
- Transparency: Clear disclosure of how the AI processes data.
Human Oversight: The output is subject to the same rigorous human review and fact-checking as our primary tools.

Process

Client Agreement: Use of such tools must be agreed upon with the client, ensuring they are aware of the data handling implications.
AI Ethics Lead Approval: For tools not pre-approved, the AI Ethics Lead must sign off on the security assessment before deployment.
Documentation: Any exception is recorded in the project file, noting the specific tool, purpose, and duration of use.

Embedded Systems

Where AI is embedded in client-provided software (e.g., Copilot in Microsoft 365), we utilise it strictly within the scope of the client’s licence and data governance policies. We do not input sensitive client data into these systems unless the client’s own security protocols explicitly permit it (which we strongly discourage).

Our Commitments

Human Oversight and Responsibility

We remain accountable: All work produced with AI assistance is reviewed, edited, and approved by us.
Professional judgment: AI suggestions are evaluated against our expertise and professional standards.
Client-specific knowledge: AI cannot replace understanding of your unique collections, context, and requirements.
Quality control: We verify accuracy, appropriateness, and compliance with relevant standards.

Transparency

Disclosure: We are transparent about our use of AI tools with clients.
Clear boundaries: We clearly distinguish between AI-assisted work and specialised expertise that requires human judgment.
Honest limitations: We acknowledge what AI can and cannot do well.

Data Protection and Confidentiality

Client data: We do not input confidential client information, proprietary data, or personal data into AI systems.
Anonymisation: When using AI for analysis or problem-solving, we anonymise or use generic examples.
Secure practices: We follow data protection principles when working with AI tools.
Encryption safeguard: Where Lumo is used, zero-access encryption provides an additional security layer should human error occur.

Quality and Accuracy

Fact-checking: We verify information provided by AI, especially for technical standards, dates, and specific claims.
Professional standards: All outputs meet relevant professional standards (cataloguing standards, accessibility guidelines, publication standards).
No plagiarism: AI-generated content is treated as draft material that we transform into original, professional work.

What We Don’t Use AI For

We do not use AI for:

Final decision-making: Critical project decisions rely on human expertise and client consultation.
Cataloguing judgment: Subject analysis, classification decisions, and archival arrangement require professional expertise.
Client strategy: Recommendations about your collections, systems, and priorities require human understanding of your context.
Automated client communication: All client correspondence is personally written.
Processing confidential data: Unless explicitly requested by the client, with appropriate safeguards in place.

Use in Specific Services

Production Management

Drafting style guides, templates, and documentation.
Research on publishing standards and best practices.
Assistance with project planning documents.
All final materials are reviewed for accuracy and client-specific requirements.

Cataloguing Services

Research on standards and controlled vocabularies.
Template creation and documentation.
Analysis of best practices.
Not used for: Actual cataloguing, subject analysis, or description—these require professional archival and library expertise.

Open Source Solutions

Research on platform features and capabilities.
Documentation and training material development.
Technical problem-solving and troubleshooting.
Code review and optimisation suggestions.
All implementations are tested and verified by human expertise.

Governance and Continuous Improvement

AI Ethics Lead

The designated AI Ethics Lead is responsible for:

Conducting quarterly reviews of all AI tool usage and capabilities.
Assessing whether Claude API access remains necessary.
Monitoring developments in AI safety, ethics, and regulation.
Approving any new AI tools or expanded use cases.
Ensuring alignment with the EU AI Act and GDPR requirements.

Staying Informed

We stay informed about:

Best practices for responsible AI use in professional services.
Developments in AI safety and ethics.
Sector-specific guidance on AI use in cultural heritage.
Client preferences and concerns about AI.

We welcome feedback from clients about our AI use and are happy to discuss how we balance efficiency with quality and professional standards.

Client Choice

If you prefer that we do not use AI tools in your work, please let us know. We can deliver all services without AI assistance, though this may affect project timelines and costs.

Questions and Transparency

We’re committed to being open about our use of AI. If you have questions about:

How we use AI in your specific project?
What safeguards we have in place?
How we ensure quality and accuracy?
Our data protection practices with AI tools?

Please contact us at parriic@parriic.eu — we are happy to discuss our approach in detail.

PARRIIC