Pipegen must not create or deliver prohibited content — in particular no sexual depictions of minors, no sexual content involving real people and no pornographic content (see the Terms). To ensure this we check requests and results automatically. The legal basis is our legitimate interest in not creating or spreading illegal content and in protecting the rights of others (Art. 6(1)(f) GDPR) and — where we are legally obliged to act against illegal content — Art. 6(1)(c) GDPR.
a) Checking your request (prompt check). Before a generation starts, a word filter on our server checks your texts (prompt, instructions of the marker edit, change requests for panoramas, dialogue lines, texts of the time windows in productions, text descriptions for 3D sets, texts for speech). If the text is clearly harmless, the check ends here; no data is passed on. Only if the text contains signals that call for a closer look do we send to the language model Claude (Anthropic) through our service provider Runware: your texts, the @ names of the Characters used with the flags “minor” or “real person” and the type of result. Pictures, your e-mail address or your user ID are not sent. The model only describes what the text asks for; whether the generation is refused is decided by our server by fixed rules.
b) Checking the results (output check). We first place a finished image or video in an intermediate store that only we can access, check it and only then release it into your account. Images are checked by an image classifier on our own server; for videos we extract single frames on our server (at most 20). For image generations through Runware we also use its yes/no signal on whether the image is not safe for work. Only in borderline cases do we send up to four reduced single frames (at most 768 pixels on the longer side) together with the flags “minor”/“real person” to Claude through Runware. We delete the file in the intermediate store after the check, at the latest after about one hour. We do not keep a blocked result; we store only a hash (digital fingerprint) of the file.
c) Face detection on uploaded pictures. When you upload your own pictures, a detection model checks exclusively on our server whether the picture shows a human face. We store only the result “face yes/no” and the time of the check. We create no facial features or templates, do not match faces against each other or against databases and identify no one; no data is passed to third parties. If a picture shows a face, you state whom it shows — a real or an invented person (d); only then can the picture be used for a generation. Legal basis: Art. 6(1)(f) GDPR (protecting the people depicted and enforcing the content rules). Since we only detect whether a face can be seen and no technical processing for unique identification takes place, in our assessment this is not biometric data within the meaning of Art. 4(14), Art. 9(1) GDPR (cf. recital 51 sentence 3 GDPR).
d) Confirming your rights to pictures of people. Before pictures with a face are used for a generation — and always when uploading into a Character — you confirm that you hold the rights to the pictures and the consent of the people depicted, or that they show an invented person. We store this confirmation as proof: your user ID and project, the time, the version of the confirmed text, the occasion, the files and Characters concerned and your statement “real person” or “invented person”. We do not store details about the person depicted (name, contact). You cannot change a statement “real person” to “invented” yourself. We also store whether you have marked a Character as “under 18”; this mark can only be removed through support. Legal basis: Art. 6(1)(f) GDPR (proof towards the people concerned and authorities, defence of legal claims). Retention: as long as your account exists and then until the end of the third year after its deletion (regular limitation period, §§ 195, 199 BGB); then we delete the proofs automatically.
e) Log of the content check. If a request or result is refused, the check cannot be completed or the language model decided, we store an entry: time, user and project ID, the generation where applicable, type and stage of the check, result and category, classifier values, Characters involved, an excerpt of your text (at most 2,000 characters) or, for results, only the hash of the file, and the cost of the check. If you publish an App, we check its cover and example pictures in the same way before publication. Harmless requests passed by the word filter or the local classifier are not logged. Only administrators have access. Purpose: proof, correcting wrong decisions, recognising repeated violations (Art. 6(1)(f) GDPR). If you delete your account, the link to your profile is removed. Retention: 12 months, then we delete the entries automatically — unless an open report about content of the same account or a pending review still needs them.
f) Individual decisions. If the check refuses a request, the generation is not carried out and you are not charged. If the check is unavailable at the moment, the request is not carried out either, for safety. We do not block your account automatically. If you think a refusal is wrong, write to info@pipegen.app; a human will then review it (section 23).
g) Reports of content (Art. 16 Digital Services Act). Anyone — even without an account — can report content to us at pipegen.app/report. We process: the details of the content, the reason and your explanation, your name and e-mail address (optional for reports of child sexual abuse material), your confirmation that the report is made in good faith and — if you are signed in — your user ID; also the ID of the account concerned, our decision with its reasons, the person who decided, any note about informing law enforcement authorities and the times of the e-mails sent. We do not store an IP address; to prevent abuse we count reports per IP address briefly in memory. The legal basis is Art. 6(1)(c) GDPR in conjunction with Art. 16, 17 and 18 of Regulation (EU) 2022/2065 (DSA). Reports are reviewed by a human. We do not tell the person concerned who reported. We send the related e-mails through Resend (section 10). Retention: 2 years after our decision, then we delete the report automatically; open reports stay until the decision. E-mails to you are sent in the language of the form (German or English), which we store with the report for this.
h) Recipients. Only in a) and b) in borderline cases does data go to Runware and Anthropic (section 18). Face detection, image classifier and frame extraction run on our server; the log, proofs and reports are stored with Supabase (section 5).