finishReason: RECITATION on users' OWN legal documents — also returns content:{} with no role, crashing strict SDK parsers

Gemini API
1

Setup: Gemini Developer API (generativelanguage) via the Vercel AI SDK (@ai-sdk/google), streamText. Model: gemini-3-flash-preview. Use case: a legal assistant that summarizes/analyzes documents that our users upload themselves (contracts, leases, official police/court reports), these are the user’s own documents, not third-party copyrighted material.

Problem: We get finishReason: “RECITATION” recurrently (with a growing trend), mid-stream, on these self-uploaded legal documents. The prompts are large (we pass a lot of the document verbatim,▎ promptTokenCount ~255k in one case). When this happens, the response is cut off and the user sees a generic error.

A second, related issue: on RECITATION the API returns the candidate with content: {} (no role field):

{
“candidates”: [{
“content”: {},
“finishReason”: “RECITATION”,
“index”: 0,
“finishMessage”: “The generated content was filtered because it may contain material that resembles existing copyrighted works…”
}],
“usageMetadata”: { “promptTokenCount”: 255707, “candidatesTokenCount”: 347, … },
“modelVersion”: “gemini-3-flash-preview”
}

Because candidates[0].content.role is missing, strict client parsers (the AI SDK’s Zod schema) throw AI_TypeValidationError instead of surfacing a clean RECITATION result.

Questions:

  1. Why does RECITATION trigger on a user’s own uploaded documents (no third-party copyright involved)? Is this expected?
  2. Does passing large amounts of the document verbatim in the prompt increase the likelihood of RECITATION? Any best practices to reduce it (chunking, instructing the model to paraphrase, system-prompt▎ patterns)?
  3. Is there any way to disable or tune the recitation/copyright check on the Developer API, or is that only possible on Vertex AI?
  4. Is returning content: {} without a role the intended payload shape for RECITATION? It breaks strict schemas, would expect either a role or a clearer blocked-response structure.
  5. Is gemini-3-flash-preview more prone to RECITATION than the stable models (e.g. gemini-2.5-pro)?
  6. Does a high rate of RECITATION finish reasons have any impact on account standing, quotas, or rate limits? We want to make sure normal legal-document usage isn’t seen as abuse.

Any guidance appreciated

thanks!

Daniel