AI Transparency Notice

Article 50, Regulation (EU) 2024/1689 (EU AI Act)

1. Purpose

This Notice tells you, plainly and in one place, what artificial-intelligence systems Pompara uses, where they touch your interaction with the Service, what content is AI-generated, what the safety architecture does, and what the EU AI Act gives you the right to know about it.

Pompara takes the position that AI transparency is not just a legal obligation but a product principle. You should always know when you are talking to AI and what AI is doing on your behalf. This Notice is published in full and is referenced from the signup flow so that the disclosures happen at the right moment in your experience.

2. You are interacting with an AI system

Pompara is an AI system. Every character you talk to is an artificial-intelligence persona — fictional, set in ancient Pompeii in 79 CE, and entirely make-believe. None of the characters is a real person, a digital twin of a real person, or a representation of any individual living or dead. Their responses are generated by AI models in real time.

This is true at every level of the Service, including text conversation with the characters, any text-to-speech audio you may receive, and any images of the characters shown to you, which are produced in advance by our own image model and selected to fit the conversation.

Pompara is not, and does not try to be, a substitute for a human therapist, doctor, lawyer, financial advisor, or any real human relationship. If you are in distress, please use the real crisis support resources we link to in the Privacy Policy section 14.

3. The AI components of the Service

Pompara is composed of several AI components working together. They are described below at the level of detail required for transparency, without disclosing specific proprietary architecture.

AI component	What it does	How it appears to you
Conversational language models	Generate the character's text responses to your messages	The character's typed messages
Routing aggregator (OpenRouter)	Chooses which underlying language-model provider answers each request, under contractual zero-data-retention terms	Invisible; same conversational output
Sentiment analysis	Infers the emotional tone of your messages to inform character responses and trigger safety framework	Affects tone; activates safety resources when needed
Behavioural and memory inference	Builds and maintains the character's memory of you across sessions	Characters recall what you've told them
Pre-generation prompt safety filter	Inspects your image-generation prompt and rejects prohibited requests	Some prompts refused with a safety message
Custom image-generation model	Produces images of the Pompara characters. Trained on 100% synthetic data with no real-person imagery, all training subjects 21+.	Watermarked image output
Text-to-speech / audio generation	Where deployed, gives voice to character responses	Audio output
Safety classifier on outputs	Inspects model outputs for prohibited content	Invisible when working correctly

Identified language-model providers in the current routing pool (subject to change): Mistral AI, xAI, Anthropic, AionLabs, NovitaAI, plus models hosted on our own infrastructure. All routes use zero-data-retention terms, and we pass only a chosen display name with each prompt — no email, account ID, or other stable identifier.

Image-generation infrastructure: our own custom-trained image-generation model running on third-party serverless GPU infrastructure (RunPod Serverless and vast.ai Serverless), using datacenter-grade GPUs only.

4. Labelling of AI-generated content

Article 50(2) of the EU AI Act requires that AI-generated or AI-manipulated content be labelled in a way that is detectable. Pompara satisfies this for each output type as follows:

• Text. All conversation in the Service is between you and AI characters. The AI nature is disclosed in this Notice, in the Privacy Policy, in the Terms of Service, and in the signup flow.
• Images. Every generated image carries a visible “Pompara” watermark in the corner. We will add embedded machine-readable provenance metadata (C2PA / IPTC-compatible) as that standard matures.
• Audio. Audio outputs are produced by the Service in the context of an AI conversation and are clearly labelled as such in the user interface where they appear.

Do not remove the labelling. Removing, obscuring or altering the Pompara watermark from any AI-generated content is prohibited by our Terms of Service and undermines the transparency the AI Act protects.

5. Deepfakes and real persons — what Pompara does not do

Article 50(4) of the EU AI Act sets specific transparency obligations for AI-generated or AI-manipulated content that constitutes a “deep fake” — i.e. realistic depictions of real persons.

Pompara does not generate deepfakes.

As a matter of architecture:

• Pompara does not accept user-uploaded reference photos of any kind
• The image-generation model is constrained to produce images of the Pompara characters only and rejects prompts referencing real, identifiable persons, public figures or named individuals
• The training data of the image-generation model is 100% synthetic and contains no real-person imagery
• All training subjects depicted in the synthetic training data are over 21 years of age

Because no real person can be depicted by the Service, the additional Article 50(4) deepfake-disclosure requirements do not arise. This is a design decision, not an oversight.

6. Safety architecture

Pompara's safety architecture is described in detail in the Privacy Policy section 14 and in the Terms of Service section 10. For AI transparency purposes, the relevant points are:

• Strict 18+ access. Image-based third-party age verification gates the entire Service.
• Pre-generation safety pipeline. User prompts for images are inspected before reaching the generative model.
• Character-constrained outputs. The image model is constrained to depict our characters only.
• Crisis detection. Sentiment analysis identifies signals of distress and triggers the AI to exit roleplay, respond with care, and surface real crisis resources. Pompara explicitly does not optimise for continued engagement when distress is detected.
• User-driven escalation. Increases in NSFW intensity are always initiated by you, never by the character.
• Make-believe framing. Characters are explicitly fictional, situated in ancient Rome.

7. Your rights regarding the AI

Article 50 and other parts of the AI Act, together with the GDPR, give you specific rights you can exercise at any time:

• Right to know. You always have the right to ask whether what you are seeing or hearing is AI-generated. This Notice is the standing answer; if any specific output ever leaves you in doubt, you can ask at safety@pompara.com.
• Right to data on the AI's view of you. Through the in-product “what does my character know about me” panel you can see, edit, and delete what the AI has stored as its working memory of you.
• Right to a human alternative for safety matters. Safety-critical communications are routed to a human at safety@pompara.com — not handled by AI in any decision-making capacity.
• Right to withdraw from AI processing of special-category data. You can lower your NSFW level or withdraw consent for sexual-preference memory at any time.

8. Sub-processor and provider transparency

The AI models and the routing aggregator used by Pompara are operated by third parties under contracts with us. The Privacy Policy section 8 sets out the full sub-processor table.

9. Changes to this Notice

We will update this Notice when our AI components change in a way that affects user-relevant transparency. The latest version is always available at https://pompara.com/ai-transparency. Material changes will be notified by email.

10. Contact

For questions about this Notice or the AI components of the Service: privacy@pompara.com

For safety concerns, AI-output reports, and DSA notices: safety@pompara.com