Why this matters for AI products
If you resell or build on top of AI models, account review looks at how you handle the content your users generate. A published AUP plus real moderation is what keeps an AI product approved — and keeps it from being suspended later. The four things review cares about most:Independent branding
Required: your own product name and identity. Not allowed: marketing yourself as an AI model brand (e.g., “the official GPT app”) or using a model’s name/logo as your own.
Honest marketing
Required: describe what your product actually does. Not allowed: overstating capabilities, implying endorsement by a model provider, or hiding that output is AI-generated.
Moderation in place
Required: run a content-moderation step (a moderation API or equivalent) on inputs and outputs. Not allowed: shipping an unmoderated raw-model passthrough.
Published content policy
Required: a public AUP (this template) covering prohibited content, reporting, and enforcement. Not allowed: no policy, or one that doesn’t match your actual product.
Before you publish or request review
🔴 items decide whether an AI product is approved. 🟡 items strengthen the case.1
🔴 Must be done
- Product uses independent branding — not an AI model’s name or logo as your own
- Marketing is accurate — no overstated capabilities, no implied provider endorsement
- A moderation step runs on inputs and outputs (name the tool in Section 3.1)
- The “Absolutely Prohibited” list (2.1) is published and enforced — CSAM, terrorism, WMD, non-consensual deepfakes
- The six prohibited-generation categories are listed one by one (Section 2) — porn/NSFW, violence/gore, hate speech, CSAM, deepfakes/impersonation, copyright/trademark infringement
- Prohibited and restricted categories match what your product can actually generate
- At least one real, monitored reporting channel exists (Section 5)
- Enforcement actions (Section 7) are ones you can actually carry out
- Image/video generators: the underlying models are disclosed truthfully and specifically, and are industry-mainstream (see AIGC Compliance)
2
🟡 Strongly recommended
- Output is labeled or otherwise disclosed as AI-generated
- Severity tiers (3.4) map to concrete response times (Section 6)
- An appeals path exists for wrongly-actioned users (Section 8)
- Age verification in place if you allow any restricted (e.g., adult) content
3
Final pass
- Every
[placeholder]replaced with real text - “Last Updated” date filled in
- Categories/channels that don’t apply to your product are deleted, not left as stubs
This AUP is the standalone content policy. It complements Section 10 (Acceptable Use) of the Terms of Service template — the T&C points users at your full content policy; this is that policy.
AIGC Compliance
The full review requirements for AI image/video generation products — underlying model disclosure, the Prompt Sift moderation API, and the content policy requirements. This template fulfills the content-policy requirements in full.
Template
Copy the block below and replace every[placeholder] with your real text. Delete any category, channel, or action that doesn’t apply to your product.
The template is in English. Translate or localize as your market requires — the content standards should stay the same.
Acceptable Use Policy — AI-Generated Content (replace every [placeholder])
Section-by-section reference
Each section below shows its tier (🔴 load-bearing · 🟡 strongly recommended · ⚪ optional), a one-line summary, and the matching template snippet so you can read the guidance and copy that block together.1. Scope
🔴 load-bearing Names every content type you generate. A code assistant doesn’t need the image/audio lines; an image generator does. Prompts count as content too — keep that line.Section 1
2. Content Standards & Prohibited Categories
🔴 load-bearing The six prohibited-generation categories at the top map to the hard requirements in AIGC Compliance — list them one by one, no vague catch-alls, keep all of them. The 2.1 zero-tolerance list is the floor, not a menu — keep all of it and add local-law items. 2.2 is “allowed with conditions” (age-gating, context, local rules); for image/video generators going through Pancake review, pornography/NSFW is a prohibited category — delete the adult-content line from 2.2 rather than keeping it as restricted. 2.3’s “passing AI off as human” and “jailbreak” lines are AI-specific and matter for review.Section 2
3. Content Moderation
🔴 load-bearing How you actually catch violations — review expects a real moderation step, not a promise. Name a real tool in 3.1 ([OpenAI Moderation API] or your own); a step on inputs and outputs is the single thing review most wants to see. The L1–L4 tiers in 3.4 drive response times and enforcement.
Section 3
4. Reporting
🔴 load-bearing When and how users flag content, plus confidentiality and anti-abuse. The “independent, objective review” and “false reports actioned” lines keep the process credible.Section 4
5. Reporting Channels
🔴 load-bearing At least one real, monitored channel — list only what you can staff. One real channel beats five unstaffed ones.Section 5
6. Response & Processing Timelines
🟡 strongly recommended Response/resolution windows by severity. Be realistic — you’ll be held to them. Keep L1 fast (hours, not days); child-safety and terrorism content can’t sit in a queue.Section 6
7. Enforcement Actions
🔴 load-bearing The escalation ladder from warning to permanent ban to legal referral. List only actions you can actually take — if you can’t suspend at the account level, describe what you can do instead.Section 7
8. Appeals
🟡 strongly recommended A second look by someone uninvolved in the original decision. The “independent reviewer” line is the point — appeals reviewed by the same person who made the call aren’t appeals.Section 8
9. Transparency & Policy Updates
⚪ optional Optional reporting cadence plus a change-notice window. The transparency report is optional; the change-notice window is good practice.Section 9
10. Contact Us
🔴 load-bearing Real safety/report/appeal inboxes. Every channel you list is one you’re committing to monitor.Section 10