A chat with a fictional character is a different kind of data than a chat with a coding model. A voice clone is a different kind of data than a typed prompt. A generated video carries your face if you put yourself in it, and your scene description if you do not. The privacy guides on this site cover text chat (How to delete your data from ChatGPT, Claude, and Gemini, How to read an AI privacy policy) and image upload to text chat (AI chats that accept image uploads without training on them). This page covers the multimodal services that go further: character chat, voice cloning, AI music, and text-to-video, where the upload is not just text plus a thumbnail.
TL;DR
- Every major multimodal AI service reserves the right to use your uploads for model improvement, but the carve-outs differ: Character.AI offers an opt-out only inside the EEA and UK, Pika splits personal data from API inputs, ElevenLabs caps voice retention at three years, and Suno’s policy explicitly names User Activity Information and Submissions as training inputs.
- “Training” and “retention” are different things. Replika says it uses small, anonymised portions of chat to train its “proprietary safety algorithms” and explicitly states the data is not used to train third-party models. Suno and Pika use the broader phrase “improve our Services, product offerings, promotional activity, and business, including to train and enhance the models that power our Services.” The two phrases describe different scopes; the second is much wider.
- Sora (OpenAI) does not publish a Sora-specific data controls toggle. Sora inputs fall under OpenAI’s general Consumer Content policy, which lets users turn off training by toggling off “Improve the model for everyone” and lets users delete Personal Data within 30 days of a deletion request.
- The hosted audio / video services (ElevenLabs voice clones, Suno tracks, Runway generations) all keep a record of your inputs and outputs for safety review even after you delete your account, and they reserve the right to keep biometric data for up to three years after your last interaction. Treat the original script and the recorded voice sample as the most sensitive pieces in the workflow.
- The cleanest answer for most multimodal workflows is still self-host: Ollama for character chat and chat with vision, OpenVoice or a local TTS for voice, Stable Diffusion or Wan for images and video. None of those local stacks phone home with your prompts by default.
Why multimodal services need their own checklist
Text-only assistants answer a narrow question: what happens to the words you typed? Multimodal services answer three at once.
The first is what the service does with the script: the dialogue you typed into Character.AI, the lyrics you pasted into Suno, the prompt you wrote for Sora. That is plain text and falls under whatever text-policy the vendor publishes.
The second is what it does with the upload: the photo you fed into a face-swap feature, the voice recording you used to clone yourself, the source video you uploaded to Runway as a style reference. That is regulated under biometric-and-content law, which gives vendors a different (and often more permissive) policy footing. Several vendors explicitly carve out biometric data from training while keeping the rest.
The third is what it does with the output: the video Sora generated, the song Suno rendered, the avatar Replika animates. Outputs may be used for product improvement under clauses that text inputs would not be. Pika’s policy spells this out: Inputs and Outputs go into “improving and training,” where the law permits it or you have consented, and API inputs generally do by default.
A checklist built only on the text-policy column misses the other two.
Character.AI
Character.AI’s policy page states that the service “retain[s] certain data that you share with us to help train and improve our AI models” and that “user-generated content can help us improve the creative writing qualities and functionality of our models and improve the AI-driven conversational experience” (Character.AI - About our Model Training).
The opt-out exists, but it is regional. The same page reads: “If you are in the European Economic Area (‘EEA’) or the United Kingdom (‘UK’), and you would like to opt out of us using your chat content to train our generative AI models… Once you opt out, new content will not be used to train our generative AI models.” Users outside the EEA and UK are not given an opt-out path on this page.
The Character.AI Privacy Policy supports tailored advertising and does not publish a fixed retention schedule; data is retained “for the period necessary to fulfill the purposes for which it was processed” (Character.AI Privacy Policy).
Practical takeaway for a Character.AI user: assume your dialogue is by default, do not paste sensitive personal history into a character chat unless you have explicitly opted out, and treat the script as something the company can still use.
Replika
Replika’s privacy policy describes a narrower training scope. It states that the service is “collecting and immediately anonymizing user feedback, and small portions of Messages and Content data to train our proprietary safety algorithms, enhance chatbot performance, prevent inappropriate outputs, and ensure compliance with safety standards” and adds that the anonymised data is “not used to train third-party large language models or other AI systems” (Replika Privacy Policy).
Replika also gives a specific retention window: it processes profile information, messages, content, interests, and preferences “for up to 60 days after termination of the contract.” It also addresses the model-routing question: data sent to third-party AI language model providers for response generation is “processed transiently and is not retained by those providers beyond the duration necessary to generate the response.”
The right way to remove Replika’s data is to ask for it; the policy says “you can request the deletion of your personal data” and that the right extends to data held by third-party AI language model providers.
Sora and OpenAI
Sora does not have a Sora-only privacy page. Sora inputs are treated as part of “Content” under OpenAI’s general Privacy Policy, which says it collects “Personal Data that you provide in the input to our Services (‘Content’), including your prompts and other content you upload, such as files, images, audio and video, Sora characters, and data from connected services, depending on the features you use” (OpenAI Privacy Policy).
The same policy describes a 30-day deletion window: “Once you choose to delete Personal Data, we will remove it from our systems within 30 days unless we need to retain it for longer as described below.” It also lets you share Sora content with others (“You can share content like ChatGPT conversations or Sora videos and characters”), which means a shared Sora URL may be reachable to anyone with the link.
Training is governed by OpenAI’s general consumer data control. ChatGPT and Sora users can switch off training use through the data-controls settings, and OpenAI states that “By default, data from ChatGPT Business, ChatGPT Enterprise, ChatGPT for Healthcare, ChatGPT Edu, ChatGPT for Teachers, and the API Platform (after March 1, 2023) isn’t used for training our models, unless you have explicitly opted in to share your data with us to improve the services” (OpenAI Enterprise Privacy). That enterprise carve-out does not name Sora, so for sensitive video work the consumer toggle is the only published path.
Runway
Runway collects “Biometric data, such as voice data and scans of faces that you submit when you use certain product features to create videos and audio using characteristics like voice and face” alongside prompts, photos, images, music, videos, audio, and other user content (Runway Privacy Policy).
Runway’s policy does not contain a sentence saying it uses uploaded videos or generated outputs to train its models. The closest statement is the general retention clause (“We will retain your information for only as long as necessary to provide the Service and for the other legitimate business purposes set out in this Policy such as safety, security, compliance, and dispute resolution”). Biometrics get a tighter cap: Runway says it will “retain face scans and voiceprints for no longer than the earlier of the date the initial purpose for collecting the data is satisfied or within three years of your last interaction with us.”
Deletion is via a documented account-deletion flow (Runway Help Center - Deleting your account). Runway also recognises browser-based opt-out preference signals such as Global Privacy Control “to the extent required by applicable law.”
Practical takeaway: even if Runway does not train on your uploads, the three-year biometric retention applies, so a face scan or voiceprint used once for a test render still sits in their system for three years from your last interaction.
Pika
Pika’s policy is the most explicit about how inputs and outputs flow into training. It collects “Inputs and Outputs, including text, image, and video inputs you provide to the Service, including through our APIs, and the associated generated content or metadata” and “Biometric information, such as scans of facial features or voice data, where applicable, when you use features that allow you to create content using such inputs” (Pika Privacy Policy).
On training, the policy says: “Where permitted by applicable law or with your consent, we may also use content you submit - including Inputs and Outputs - to improve our artificial intelligence models and related technologies.” It then carves out AI Self personal data (“any personal information included in the Inputs used to create an AI Self… will only be used to develop, operate, improve and train your AI Self and will not be used to train other artificial intelligence models”) and adds a default for the API: “personal information that you include in Inputs through the API platform will generally be used to improve and train artificial intelligence models.” API users who want to opt out have to do it in writing.
Biometric retention is capped: “If you submit biometric information… we retain that information only as long as necessary to fulfill the purpose for which it was collected or, at most, for three years after you delete your account.” Deletion goes through [email protected], and the policy warns that “We cannot delete your personal information except by also deleting your user account.”
Suno
Suno’s policy names training inputs directly. It collects “User Activity Information, Submissions, Interactive Chat Information, and other Content to improve our Services, product offerings, promotional activity, and business, including to train and enhance the models that power our Services…” (Suno Privacy Policy).
Retention is open-ended (“for as long as is reasonably necessary for the purposes specified in this Privacy Notice”) but voice data, if Suno ever handles it, gets the standard three-year-from-last-interaction biometric cap. Deletion requires an email request.
The combination is the worst-case for an evergreen guide: by default, your Suno prompts and outputs train Suno, there is no documented consumer opt-out in the privacy policy itself, and retention is “as long as reasonably necessary.” Treat every Suno submission as public-by-default.
ElevenLabs
ElevenLabs’ policy defines voice data clearly: “One kind of input is voice data. This could be an audio recording or live narration of your voice… We use these voices to create the output of a synthetic copy of the voice called a ‘voice model’” (ElevenLabs Privacy Policy).
On training, the policy says it uses personal information to “research, build, train and improve our AI models. This includes our models learning patterns in speech and audio.” The opt-out is concrete: “You can opt out of our use of your personal information at any time by navigating to the ‘Data use’ menu in the ‘Terms and Privacy’ section.”
Retention is capped for voice data specifically: “We won’t keep data we generate about your voice for longer than 3 years after your last interaction with us.” Deletion is via account deletion plus an opt-out toggle for use-in-training.
The local fallback
Most of these services have a self-host equivalent that does not phone home with your inputs by default:
- Character and roleplay chat: Ollama with Mistral-Small or Qwen, plus a local frontend like SillyTavern or Open WebUI. The roleplay guide covers this in depth.
- Voice cloning: OpenVoice (MIT), Fish Speech (Apache 2.0), or the Kokoro TTS server (covered in our self-host Kokoro guide).
- Text-to-music: Suno and Udio are closed-weight. The closest self-host is ACE-Step (MIT) for music generation, covered in our self-host ACE-Step guide.
- Text-to-video: Wan 2.2 (Apache 2.0) for diffusion, LTX-Video for fast generations, covered in our self-host AI video generation guide.
None of these local stacks include the hosted service’s account and biometric systems. Your prompts and outputs stay on your disk.
The pre-flight checklist
Before you upload to any multimodal AI service:
- Decide whether the upload is biometric. If the service takes your face, your voice, or your fingerprint to clone or animate, biometric law applies and most vendors cap retention at three years. Treat the upload as the most sensitive piece in the workflow.
- Check the policy for the word “Inputs and Outputs,” not just “uploads.” Services that name both (Pika, Suno) use a wider training scope than services that only name one or the other.
- Check whether the policy publishes an opt-out path for your region. Character.AI’s opt-out is EEA/UK only; Pika’s API opt-out is in writing; Suno’s policy does not publish one. Absence of a written opt-out usually means the default applies.
- Strip metadata from the upload. Photos carry GPS and device info in EXIF. Audio files carry device tags. Use ExifTool before uploading, or take the photo with airplane mode and a fresh camera app that strips location.
- Anonymise the script. Replace real names, employer names, addresses, account numbers, and internal URLs before pasting. The redaction guide walks through the tool.
- Confirm what “delete” means. Account deletion usually also drops the training carve-out. Pika’s policy is explicit that account deletion is the only path to personal-data deletion. Treat “delete this generation” as a UI action and “delete my account” as the only real removal.
The Bottom Line
Multimodal AI services reserve the right to use your uploads and your outputs to improve their models. The clauses differ: Character.AI’s opt-out is regional, Replika’s training scope is narrow and self-contained, OpenAI’s data controls cover Sora through the general consumer toggle, Runway’s biometric cap is three years, Pika’s API defaults to training, Suno’s policy explicitly names its training inputs, and ElevenLabs publishes an in-product opt-out. A checklist that only reads the text-chat column misses the upload and the output. A checklist that reads all three columns, strips metadata, anonymises the script, and chooses the local fallback when the workflow allows keeps the sensitive pieces of a multimodal workflow on your side of the wire.