AI Transparency

1. Models we run today

Flompt currently runs the following models:

• Recommendation ranking (Kubigorithm Heavy Ranker) — predicts how relevant a prompt is for you. Trained internally on aggregated engagement signals.
• Recommendation candidate retrieval (Kubigorithm Light Ranker) — narrows the candidate pool from thousands to hundreds before heavy ranking.
• Content moderation classifier — flags prompts for human review before they go public.
• Semantic search embeddings — converts prompt text to vectors so we can surface similar prompts.

2. What data goes in

When (and only when) you opt in to ML Training in Settings, the following pseudonymized data may be used:

• Public prompt text and tags you authored (private prompts excluded)
• Pseudonymized engagement signals — likes, copies, saves, dwell time
• Pseudonymized session-level patterns — foreground sessions, locale, app version

Identifiers (userId, email, device IDs) are never sent to the training pipeline; we replace them with daily-rotating HMAC-SHA256 pseudoIds. Pseudonymized data still counts as personal data under KVKK / GDPR.

3. What models output

Models output rankings and labels — they do not output your original prompt text verbatim as their own response. Output filtering is applied to prevent memorization leakage.

4. Third-party AI subprocessors

We train our ranking and recommendation models on our own infrastructure; none of your data is sent to a third-party AI service for training. Some features do, however, run on third-party AI services, which process the data solely to perform the requested operation (inference):

• OpenAI (USA): automated moderation of submitted prompts and comments before publication
• Google Gemini (USA): vector generation from content text for semantic search, processing of the search term you type, invisible automatic topic labels, and image generation and editing in AI Studio
• Replicate (USA): running some of the image generation and editing models in AI Studio
• Under the terms that apply to us, none of these providers use your content to train their own models; exactly what data each one receives is itemized in the Third Party Services section of the Privacy Policy
• If we add a subprocessor for training purposes (for example a fine-tuning provider): the Privacy Policy will be updated with the subprocessor name, jurisdiction, and purpose, your ML Training consent will be re-requested in the app, and we will sign a Data Processing Agreement that prohibits cross-customer training

5. Your control

You can opt out at any time:

• App > Profile > Settings > Data & ML Consent — toggle 'ML training' off
• Email [email protected] to request manual deletion of your contributions to future training runs
• Use the Export Data button to download every event, consent record, and prompt linked to your account (KVKK Art. 11 / GDPR Art. 20)

Withdrawing consent stops future training inclusion. Models already trained are not retrained, which mirrors industry standard practice (a single user's data has marginal effect on a trained model).

6. Questions

AI ethics or transparency questions: [email protected]