Using AI to Generate App Content Dynamically

Static content doesn't adapt: the same tips for every user, the same practice questions every day. Generated content can adapt to a learner's level, a user's goals or a product's attributes. It can also be wrong, inappropriate, or inconsistent with your app's design, unless you engineer guardrails.

A safe generation pipeline

Structured output instead of free text

@Serializable
data class QuizQuestion(val question: String, val options: List<String>, val correctIndex: Int, val explanation: String)

@Serializable
data class Quiz(val topic: String, val questions: List<QuizQuestion>)

class QuizGenerator(private val model: GenerativeModel) {

    suspend fun generate(topic: String, level: String): Quiz? {
        val prompt = """
            Create 5 multiple-choice questions about "$topic" for a $level Android learner.
            Each question: 4 options, exactly one correct (correctIndex 0-3), a one-sentence explanation.
            Use only well-established facts; avoid version-specific numbers.
        """.trimIndent()
        repeat(2) {                                         // one retry on invalid output
            val raw = model.generateContent(prompt).text ?: return@repeat
            val quiz = runCatching { Json { ignoreUnknownKeys = true }.decodeFromString<Quiz>(raw) }.getOrNull()
            if (quiz != null && quiz.isValid()) return quiz
        }
        return null                                         // caller shows curated fallback quiz
    }
}

/** Business-rule validation the schema can't express. */
fun Quiz.isValid(): Boolean = questions.size == 5 && questions.all { q ->
    q.options.size == 4 && q.correctIndex in 0..3 &&
    q.options.distinct().size == 4 && q.question.length in 10..300
}

Configure the model with responseMimeType = "application/json" and a responseSchema (see structured output example) so it's constrained to your shape, then still validate rules such as distinct options.

Generate on the client or the server?

Moderation and policy

• Use model safety settings and an additional moderation step for user-visible content.
• Google Play's policies require apps with AI-generated content to prevent prohibited content and to provide ways for users to report offensive content. Review the current policy.
• Avoid generating content about real people, or anything that could be mistaken for authoritative advice.

Presenting generated content

• Label it ("Generated by AI"), and let users regenerate, report or hide it.
• Keep a curated static fallback so the feature never shows an empty state.
• Cache results so the same content doesn't change unexpectedly between visits.

Use cases

Common mistakes

• Rendering raw model text with HTML or Markdown that can break the layout or inject links.
• Trusting numbers, prices or facts generated by the model.
• Generating per screen view instead of caching, which drives cost and inconsistency.
• No reporting path for bad content.

Frequently asked questions

Conclusion

Generated content is valuable when it's structured, validated, moderated, cached and labelled, with a static fallback. Generate shared content on the server, personalise carefully on the client, and never let raw model output reach users unchecked.

Next step: make your app's content findable with Build an AI-Powered Search Feature.