AI Use & Transparency Policy

This policy explains how the CUEBES platform uses artificial intelligence, what your rights are regarding AI systems on CUEBES, and our commitments around training, transparency, and human oversight. CUEBES is operated by vaicat B.V.; this policy is part of the Privacy Policy.

1. Core Principles

• Human-first: AI assists users; it does not make decisions for them.
• Transparency: users are informed when AI is involved.
• User control: you can opt out of AI personalisation; you remain responsible for content you publish.
• Privacy protection: personal data is safeguarded against misuse, including against being used to train models.

2. What We Mean by AI

For this policy, "AI" refers to automated systems capable of:

• Generating text, summaries, or responses
• Classifying or categorising content
• Providing personalised recommendations
• Drafting assistance and suggestions
• Supporting content-moderation workflows
• Embedding-based retrieval over a corpus (search, recommendation)

This includes generative AI (large language models), traditional machine learning (recommendation engines), and rule-based / deterministic systems described as "AI" in their product surface.

3. AI Training & Your Data

3.1 No Training Without Consent

We do not use personal data or user content to train or fine-tune general-purpose AI models. This commitment covers your code, posts, comments, messages, votes, profile data, and any other content you contribute to CUEBES.

Any future training on user data would require:

• Explicit, informed opt-in consent
• Clear in-product prompts explaining what data is used
• The ability to withdraw consent at any time

3.2 What This Covers In Practice

Concretely, today:

• CathAI on-device (Apple Foundation Models): uses Apple's pretrained models. We do not fine-tune them on your data.
• Server-side cath-embed: generates embeddings of a public-article corpus (not your content) for retrieval. Your queries are processed at request time and not retained for training.
• VIT pipeline (M1–M4): runs against user content for live tasks (extraction, classification, ranking). It does not train on your content.
• University and Cube Agents: pre-built rule-based + retrieval bots. They do not learn from cube traffic.
• Recommendation systems: use deterministic rules and aggregate signals; not ML training on your personal data.

3.3 Withdrawing Consent

If you ever consent to a future AI-training feature, you can withdraw consent at any time through Settings. Withdrawal takes effect immediately for future processing. Previously processed data may not be "unlearned" from a trained model — that is an inherent limitation of the technology.

4. Transparency Requirements

4.1 Disclosure

We indicate AI involvement when:

• Content is generated or substantially drafted by AI
• You interact with an AI-powered assistant or chatbot (CathAI, agent accounts)
• Recommendations or rankings are based on AI/ML analysis
• Required by law (EU AI Act, DSA Art. 27 transparency)

4.2 Synthetic Media

AI-generated or substantially modified images, audio, or video ("deepfakes") are labelled where applicable law requires.

4.3 User Responsibility

When you use AI-assisted features:

• You must review outputs before sharing or publishing
• AI outputs may be inaccurate, incomplete, or misleading
• You are responsible for content you publish, even if AI-assisted

5. What AI Cannot Do on CUEBES

• Make legally binding decisions — those require human judgment
• Autonomously suspend or terminate accounts — account actions require human review
• Replace required human review — significant decisions need human oversight
• Determine your legal rights or obligations — those require qualified professionals
• Provide professional advice — legal, medical, financial, or immigration advice requires licensed professionals

6. AI in Content Moderation

6.1 Assistance Role

AI assists our moderation workflow by:

• Flagging potentially violating content for review
• Detecting spam, malware, and coordinated abuse
• Prioritising review queues based on severity
• Identifying patterns of policy violations

6.2 Humans Make Final Decisions

Humans make every final enforcement decision. Specifically:

• Account suspensions and terminations require human review
• Content removal for borderline cases involves human judgment
• Appeals are reviewed by human moderators not involved in the original decision

The full process is documented in the Moderation & Enforcement Policy.

7. Your Controls

7.1 AI Personalisation

You can control AI-powered personalisation:

• Disable recommendations: Settings → Privacy → Personalisation
• Reset recommendation history: clears the signals used for AI suggestions
• Opt out of AI features: use the standard non-AI alternatives where available

7.2 Objection Rights (GDPR Art. 21)

You may object to automated processing based on legitimate interests. Email privacy@vaicat.com to exercise this right. See the Privacy Policy for the full set of GDPR rights.

7.3 Human Review

For any automated decision that significantly affects you, you can request:

• An explanation of the decision logic
• Human review of the decision
• The opportunity to contest the decision

8. CathAI

CathAI is the CUEBES AI assistant. It runs in a hybrid architecture — some processing happens locally on your device, some happens on vaicat-controlled servers in the EU.

8.1 On-Device (Apple Foundation Models)

Drafts, rewrites, and summaries run locally using Apple Foundation Models. Your input never leaves your device for these operations. We do not fine-tune these models.

8.2 Server-Side (CUEBES API + FAISS retrieval)

Article search, social-source lookup, and the multi-stage VIT pipeline (M1–M4) run on the CUEBES API. The FAISS sidecar holds embeddings of a public-article corpus, not user content.

• All server-side AI processing happens on vaicat-controlled infrastructure in the EU.
• No third-party AI providers (OpenAI, Anthropic, Google, etc.) receive your CathAI queries.
• Your queries are not retained for model training.

8.3 EU AI Act Classification

Under the EU AI Act, CathAI is an AI system subject to transparency obligations (Article 50). We disclose that you are interacting with an AI; AI-generated content is labelled where applicable; transparency information is available in app settings.

8.4 Capabilities & Limitations

CathAI can:

• Search and summarise articles, sources, and topics
• Help draft and rewrite text on-device
• Answer general questions about content visible on CUEBES
• Provide guidance on platform features

CathAI cannot:

• Provide legal, medical, financial, or immigration advice
• Make decisions on your behalf
• Guarantee the accuracy of any third-party data
• Access private content you have not granted it permission for

8.5 Disabling CathAI

You can disable CathAI at any time in Settings → Privacy. The app continues to work with manual search and browse. Disabling CathAI does not affect your saved data or preferences.

9. EUnify Matching

EUnify (institution discovery) uses rule-based matching, not machine learning:

• Recommendations are based on your stated preferences (academic interests, country, programme type)
• Deterministic ranking, not probabilistic AI models
• No payment influences ranking
• You can disable personalisation entirely in Settings

10. University & Cube Agents

Some institution and topic cubes host agent accounts — clearly-marked bot accounts that post auto-syndicated content (RSS, public Mastodon and Bluesky timelines) or answer questions within a defined scope.

• Agents are visibly labelled with a 🤖 indicator and a source chip
• Agent accounts are pre-built and rule-based — they do not learn from cube traffic
• Each agent's allowed scope (which cubes it can post in, what sources it pulls) is server-enforced and audited
• Cube moderators can mute a specific agent from posting in their cube
• You can hide all agent posts platform-wide via Settings → Privacy → "Hide auto-syndicated posts everywhere"

Operator

• vaicat B.V.
• Noordeinde 99B, 2514 GD The Hague, Netherlands
• KvK (Dutch Trade Register): 89360672
• Email: privacy@vaicat.com