Imagine your app with user-generated content (UGC) getting blocked in the App Store due to lack of moderation. Or users complaining about harassment in chats. Without a reliable content moderation system, you cannot release a product that passes review and remains safe. UGC moderation is a key element of any social app. We are a team of mobile developers with 5+ years of experience, and we have implemented dozens of moderation systems for iOS and Android. In this article, we'll show you how to build an AI text moderation pipeline using the OpenAI Moderation API that ensures compliance with App Store Review Guideline 1.2 and protects users from unwanted content. One of our projects—a fintech app with chats—we implemented multi-level moderation that reduced complaints by 80%. The client-side filter catches 40% of violations before they reach the server, reducing latency and load. Final moderation accuracy reached 99.5%, with a median check time of 180 milliseconds. This approach guarantees speed and precision. Savings on manual moderation amounted to 60% per moderator cost.
Problems We Solve
The main technical challenges when implementing moderation:
API Key Leakage
If the OpenAI Moderation API is called directly from the client, the key ends up in the binary. Even obfuscation does not help—attackers extract it. Solution: all requests go through a backend-proxy. Only the server knows the key.
False Positives
OpenAI returns probabilities, not a binary answer. Without proper thresholds, up to 30% of legitimate content gets blocked. We adjust thresholds for each category and add manual verification for the "gray area." In one project, this reduced false positives by 60%.
Multilingual Support
Non-standard forms (translit, leetspeak, deliberate misspellings) reduce accuracy. We apply text normalization before the check—this increases detection by 20%.
How AI Text Moderation Improves Your App's Security?
The architecture includes four levels:
User enters text ↓ [Client] Local check (instant) ↓ passed [Backend] OpenAI Moderation API (100–300 ms) ↓ passed [Backend] Custom rules (regex, domain-specific) ↓ passed Content published ↓ parallel [Backend] Async re-check (more expensive model) According to OpenAI documentation, Moderation API is designed to detect harmful content across several categories. The client-side filter catches obvious violations before sending them to the server. This reduces load and protects the user from delays. For one fintech app, we implemented such a pipeline: 99.5% accuracy with a median delay of 180 ms.
Why Client-Side Moderation?
On the client, speed is crucial. We use the NaturalLanguage framework on iOS and similar on Android. A simple example—a local list of banned words compiled into regex:
import NaturalLanguage class LocalTextModerator { private let forbiddenPatterns: NSRegularExpression init() { let patterns = ["word1", "word2"].joined(separator: "|") forbiddenPatterns = try! NSRegularExpression( pattern: "\\b(\(patterns))\\b", options: [.caseInsensitive] ) } func quickCheck(_ text: String) -> ModerationResult { let range = NSRange(text.startIndex..., in: text) if forbiddenPatterns.firstMatch(in: text, range: range) != nil { return .blocked(reason: .explicitContent) } return .passed } } We store the word list encrypted or load it from the server at startup—to avoid exposing the binary. The client-side filter is 10 times faster than the server: 10 ms vs. 100-300 ms. Our engineers are ready to audit your app—contact us for a consultation.
Backend-Proxy Architecture for API Key Protection
The only secure way is a backend-proxy. The app sends text to your server, the server calls the OpenAI Moderation API, and returns the result. Example request: POST https://api.openai.com/v1/moderations with Authorization: Bearer
Handling Edge Cases
OpenAI Moderation does not give a binary answer—it's probabilities. You need business logic for the "gray zone":
fun evaluateModerationResult(result: ModerationResult): ContentDecision { return when { result.flagged -> ContentDecision.BLOCK result.categoryScores["harassment"]!! > 0.7 -> ContentDecision.BLOCK result.categoryScores["harassment"]!! > 0.3 -> ContentDecision.REQUIRE_REVIEW result.categoryScores["sexual"]!! > 0.4 -> ContentDecision.REQUIRE_REVIEW else -> ContentDecision.ALLOW } } Content with REQUIRE_REVIEW goes into a manual moderation queue or is published with reduced visibility.
Example threshold configuration for categories
For the hate category, BLOCK threshold = 0.7, REQUIRE_REVIEW = 0.3. For sexual, REQUIRE_REVIEW = 0.4. Thresholds are selected based on the app's specifics.| Approach | Speed | Accuracy | Load |
|---|---|---|---|
| Client-side filter | 10 ms | 70% | Low |
| OpenAI Moderation API | 200 ms | 98% | Medium |
| Combined pipeline | 180 ms | 99.5% | Medium |
Multilingual Normalization
For Russian and translit, we apply normalization:
func normalizeText(_ text: String) -> String { var result = text.lowercased() let translitMap = ["a": "а", "e": "е", "o": "о", "p": "р", "c": "с"] for (latin, cyrillic) in translitMap { result = result.replacingOccurrences(of: latin, with: cyrillic) } result = result.replacingOccurrences(of: "(.)\\1{2,}", with: "$1", options: .regularExpression) return result } We check both normalized and original text—this gives +20% accuracy.
Process
- Analytics — study your app's specifics, UGC, platform requirements.
- Design — choose the stack (iOS/Android/cross-platform), draw the pipeline architecture.
- Implementation — write code: local filters, OpenAI integration, custom rules, normalization, rate limiting.
- Testing — load testing, A/B tests for thresholds, verification on real data.
- Deploy — configure monitoring, logging for appeals, CI/CD.
What's Included
| Stage | Result |
|---|---|
| Requirements analysis | Document with architecture and metrics |
| Design | Pipeline diagram, model selection |
| Implementation | Integration with OpenAI Moderation, local filters, normalization, rate limiting |
| Testing | Load test report, threshold tuning |
| Deploy | Documentation, team training, 1 month support |
Our Results and Guarantees
We are a team with 5+ years of experience in mobile development, certified Apple and Google developers. We have completed over 20 content moderation projects. We guarantee:
- Passing App Store and Google Play Review.
- Data confidentiality (NDA).
- Stable system operation under load.
Get a free engineer consultation—contact us.
Timelines and Cost
Basic integration (client-side filter + OpenAI Moderation) — from 2 to 3 days. Full system with pipeline, manual moderation, normalization, and analytics — from 2 to 3 weeks. Cost is calculated individually after an audit. The investment pays off by reducing blocking risks and saving on manual labor.