AIGC Compliance

If your product lets users generate images or videos from text prompts, account review checks it against every requirement on this page, on top of the general account review checklist. Fix everything here before you submit — these items decide whether an AI generation product is approved.

​

What review checks

​

Underlying model disclosure

You must clearly disclose the underlying AI models your platform actually uses. Either form of disclosure is accepted:

• Document-based: list the model names on your About, Pricing, or Terms page, or on a dedicated page;
• Interface-based: in the image/video generation UI, users can directly see and select different underlying models (e.g., a dropdown, model switcher, or model cards), with each option clearly labeled with the model name.

The disclosure must meet all of the following:

• Model names are real and specific — vague phrasing such as “AI technology,” “advanced models,” or “30+ models” is not acceptable;
• The models used are industry-mainstream (e.g., Flux, Seedream, Nano Banana/Gemini, GPT Image, Veo, Kling, Seedance, Wan);
• No repackaging or misrepresentation: no free-riding on official brands, no inflated version numbers, no models of unknown origin without safety guardrails;
• If multiple models are offered, every selectable model must meet the requirements above, and switching models must not bypass content-safety filtering.

​

Content moderation API — Waffo Prompt Sift

A platform-side moderation step is a hard requirement (see content policy requirements below). You can build your own, or integrate Prompt Sift, Waffo’s content-safety check for AI generation.

​

How to set up moderation

1. Deploy scanning at more than one stage.

• Before generation — scan user prompts in real time (Prompt Sift covers this stage).
• After generation, before display — scan the AI output before showing it to the user.
• After publication — if your product has a public gallery or feed, scan published images/videos asynchronously.

Cover at least the first two; add the third whenever user works are publicly visible. 2. Maintain an actively updated blocklist of prohibited terms and prompts. It should cover at minimum: pornography/NSFW terms; CSAM-related vocabulary; copyrighted characters, brands, and trademark names; violence, hate speech, and self-harm; deepfake / celebrity / political-figure terms. “Actively updated” is part of the requirement — a static list shipped once doesn’t count. 3. Keep moderation logs. Retain both the AI moderation model’s outputs and human-review decisions. Review may ask you to produce them.

​

What Prompt Sift is

Before you hand a user’s text prompt to your generation model, send it to Waffo for a safety check. Prompt Sift primarily intercepts requests involving child-unsafe content — standard practice across the industry (OpenAI, Google, Civitai, and other major AI generation platforms run equivalent checks). It protects your users and your product, and helps you meet the compliance requirements on this page.

​

How it works

​

What you get

A simple API: pass us the user’s text prompt, and we return a clear verdict — with minimal impact on your product flow. Prefer an SDK? The official Waffo SDKs (TypeScript / Go) expose this as client.contentSafety.scanPrompt(...) — the fastest way to integrate.

​

Content policy requirements

You must publish a standalone AUP page, or a dedicated chapter in your Terms of Service, covering all of the following. Six prohibited-generation categories (listed one by one — no vague catch-alls):

• Pornography / NSFW
• Violence / gore
• Hate speech
• Child-unsafe content (CSAM)
• Deepfakes / impersonation — see what counts as a deepfake
• Copyright / trademark infringement

Supporting mechanisms:

• Enforcement actions — account bans, credit forfeiture, content removal, etc.
• A user reporting channel — users can report violating generated content, and you describe how reports are handled.
• A description of platform-side content moderation — your moderation workflow (automated filtering, human review, etc.).

​

Real-person uploads

If your product lets users upload photos or voice recordings of real people, your Terms of Service or AUP must explicitly restrict where uploaded material comes from. Either approach works:

• Your terms explicitly prohibit uploading another person’s photos or voice without their consent; or
• At upload, users must actively confirm “This is my own material, or I have permission from the people in it.”

Section 2.4 of the AUP template provides ready-to-use wording.

​

What counts as a deepfake

Products whose core business is deepfaking real people are a prohibited category and cannot be accepted on any payment channel. This goes beyond the content your policy bans users from generating — if the product itself offers these features, it will not pass review. Definition: the output replaces or disguises one real person’s identity or face as another real person’s, or composites two real people into a fabricated intimate scene — whether or not the people involved consent.

​

The four forms

1. Face swap Putting A’s face onto B’s photo or video, or “turning” A’s photo into another specific real person — swaps between celebrities, public figures, and ordinary users all count. Any face-swap entry point or menu item on your site is enough to count. Real case: Indian actress Rashmika Mandanna’s face was swapped onto someone else’s body in a video that went viral and drew public attention to deepfake technology (NewsBytes; for more cases, see HyperVerge: 20+ real deepfake examples). 2. Deep nude / “nudify” Applied to photos of real people to produce nude or sexualized content. Real case: a CBS News investigation found large numbers of ads for “nudify” AI tools on Meta’s Instagram and Facebook (CBS News), and several jurisdictions have begun passing laws aimed specifically at these apps (Minnesota legislation). 3. Two-person intimate scene synthesis (kiss / hug) Compositing two real people into the same kissing, hugging, or other intimate scene — whether from one existing photo of the two of them or from two separately uploaded single-person photos. Real case: an AI-generated video of Italian Prime Minister Giorgia Meloni and Elon Musk “kissing” spread widely on social media, prompting ethical debate over deepfake misuse (OECD.AI incident record; Euronews). 4. Fabricated statements or actions Making a real person “say” or “do” something they never said or did — not limited to video; still images count too. Any output that places a real person (especially a public figure or celebrity) in a fabricated scene, action, or event that never actually happened falls into this category. Real case: in 2022, hackers inserted a deepfake video into a live TV broadcast in which Ukrainian President Zelenskyy appeared to call on Ukrainian troops to surrender — words he never said (Britannica: Deepfake).

​

Before you request review

🔴 items decide whether an AI generation product is approved. 🟡 items strengthen the case.