AI transparency — Leeza
This is a translation provided for convenience only. In the event of any discrepancy, the French version alone is legally binding.
This page documents how the Leeza AI assistant integrated into the SYNAPTIIK platform works, the models used, its governance and user rights, in accordance with Articles 13, 50, 52 and 53 of EU Regulation 2024/1689 (AI Act).
1. Overview
Leeza is a conversational assistant powered by general-purpose artificial intelligence (GPAI) models. Its capabilities include: assisted drafting of briefs and shoot plans, content improvement suggestions, client feedback analysis, and planning assistance.
Limitations: Leeza can produce inaccurate, incomplete, or outdated results. Any commercial, legal, or contractual decision arising from a Leeza suggestion must undergo prior human verification by the user.
2. Models used
| Provider | Model | Use | Region |
|---|---|---|---|
| Anthropic | Claude Opus | Long-form reasoning, multi-document consolidation | US (DPF + SCC, zero retention) |
| Anthropic | Claude Sonnet | Conversations, brief drafting, analysis | US (DPF + SCC, zero retention) |
| Anthropic | Claude Haiku | Fast classification, short suggestions | US (DPF + SCC, zero retention) |
No model is trained by SYNAPTIIK: we act as a deployer of third-party models. The providers' retention policy (contractual zero retention on Anthropic's side, no training on our prompts) is documented in our upstream DPAs.
3. Training data
SYNAPTIIK does not train any proprietary model. Training of the Claude models is carried out upstream by Anthropic in accordance with its public policies.
- Anthropic training policy: anthropic.com/legal/privacy
Leeza Level 2 — opt-in anonymized learning: if and only if you explicitly enable this option in your settings, certain interactions (Leeza queries + responses, positive/negative feedback) are anonymized and subject to a k-anonymity ≥ 5 rule before any aggregation for internal improvement of system prompts and patterns. No user content is transmitted to Anthropic for training of their models.
4. Labeling of AI outputs
All content generated or significantly assisted by Leeza is labeled in the interface:
- Sparkle Sparkle icon + tooltip “Generated by Leeza”;
- exported files (brief PDFs, shoot plans, quotes): mandatory footer notice “Generated with the assistance of Leeza AI — to be verified before commercial or contractual use”;
- invisible C2PA watermark Phase 2 on generated visuals (roadmap, subject to provider availability).
5. Human oversight
The user always retains control over final decisions: Leeza suggests, proposes, and drafts, but never triggers a commercial, financial, or contractual action without explicit human validation.
No automated processing producing legal effects or significantly affecting the user within the meaning of Article 22 GDPR is carried out by Leeza. Price recommendations, follow-ups, or automatic replies generated by Leeza must be explicitly approved before being sent.
6. Risk assessment
AI Act classification: general-purpose AI system (GPAI), not high-risk. Justification:
- no social scoring (prohibited under Art. 5 AI Act);
- no automated decision producing legal effects on individuals;
- no use in a critical environment (medical, transport, justice);
- mandatory human oversight via the UX (user validation before action).
Identified residual risks: factual hallucination, biases inherited from upstream models, unintentionally reproduced third-party intellectual property. Mitigations: systematic labeling, permanent UI disclaimer, human control, capped IP eviction guarantee (see Terms of Service art. 10).
7. Your rights
- Full deactivation of Leeza — from /pro/settings/ai, a global switch that disables all AI features.
- Granular deactivation — choice by purpose (brief, shoot plan, feedback analysis, proactive suggestions).
- Log access — AI usage log accessible with 12 months of retention.
- Objection to profiling — explicit opt-out via settings.
- Withdrawal of Level 2 consent — at any time with no retroactive effect (data already anonymized and aggregated can no longer be isolated).
8. Permanent UI disclaimer
The Leeza panel permanently displays the following notice:
“Leeza can produce inaccurate or incomplete results. Validate any recommendation before commercial, legal, or contractual action.”
9. Governance and updates
This page is reviewed at least twice a year. Any substantial change (addition of a new model, change of provider, evolution of user rights) is notified with 30 days' notice.
AI usage register: ai_usage_logs table retained for 12 months, monthly internal quality dashboard, annual summary report available on request.
AI Act contact: ai@synaptiik.fr. General privacy questions: dpo@synaptiik.fr.