澄清 AI 数据保留困惑:两种标准设计解析
数据留存(data retention)已成为 AI 基础设施中最容易被误解的术语之一。在一波 AI 安全相关头条报道之后,许多开发团队现在都在问一个简单的问题:我的 AI 服务商会不会保留我的数据?诚实的答案是,这取决于你以何种方式发送数据。把提示词发送到推理 API,它会被处理并返回,不会被留存。这种设计被称为零数据留存(ZDR)。上传数据集用于批处理任务,数据则会一直存储,直到你将其删除。两者都是标准设计,而且在两者中控制权都在客户手里。变化的只是由谁管理数据生命周期,以及在何处管理。
Fireworks AI 联合创始人兼 CEO Lin Qiao 提出了一种清晰的思考方式。在 2026 年 11 月 13 日发布于 X 的帖子中,她描述了两种标准 API 设计。每一种都决定了数据生命周期的控制权归属何处。同样的分野已经塑造数据处理系统长达数十年。本文将结合公开 API 文档中的示例,梳理这一框架。
AI 数据留存政策界定了平台会从你的请求中存储哪些数据,以及存储多久。OpenAI 的数据控制描述了两个层面。应用状态是某项功能必须保留的数据,例如上传的文件。滥用监控日志是为安全审查而留存的副本。平台可能在一个层面实现零留存,在另一个层面则不然。
AI 推理和训练 API 遵循两种生命周期设计。区别在于谁管理数据,以及在何处管理。
你创建一个文件或数据集,针对它运行任务,然后将其删除。平台会持续存储该数据并保持可用,直到你请求删除。除非你明确定义,否则平台无法知道你是否已经用完。因此删除是你的责任,需要通过平台的 API 来完成。
带上传数据集的批处理 API 就是标准例子:
数据一旦删除,就彻底消失了。据 Fireworks CEO 称,收到删除请求后,它不会保留任何数据。
OpenAI 的非批处理文件会一直保留,直到你主动删除。Fireworks AI 托管的训练数据集没有时间限制,由客户自行删除。Amazon Bedrock 和 Google 的批处理任务读写的都是你自己的云存储。
OpenAI 的批处理文件默认保留 30 天,Anthropic 的批处理数据为 29 天。应把到期失效视为例外情况,并逐项功能加以确认。
你发送数据,平台进行计算,结果再返回给你。不会留下任何应用状态。对话历史由你的应用自己持有,因此不需要平台做清理。这正是零数据保留(ZDR)所适用的模式。仅凭同步响应本身,并不能构成 ZDR。
有据可查的示例:
可以这样理解:把电子表格保存到云盘,与把表格里的数字交给计算器运算,两者并不相同。云盘会留存文件以备后用;计算器算出答案、把结果返回,什么都不保留。
在这两种设计中,客户都能完全掌控其数据的处理方式。区别在于数据生命周期的管理位置。在模式 1 中,体现为显式地创建和删除数据集。在模式 2 中,生命周期完全由你的应用负责。
答案是“否”。ZDR 的作用范围按端点(endpoint)划定,各平台的默认设置也各不相同。
OpenAI 默认最多保留滥用监测日志 30 天。Amazon Bedrock 的 none 模式不会将任何请求或响应数据写入持久化存储。
Fireworks AI 的文档说明,对开放模型默认提供 ZDR。其临时推理 API、Training SDK 和 Nexus 均带有 ZDR 保障。提示词与生成内容不会以静态形式留存。对于 Batch API,Fireworks 的文档指出数据在删除后不会被保留。
单个参数也能改变这一模式。在 Google 的 Interactions API 上,store 默认为 true。在 OpenAI,将 store 设为 false 并不等于 ZDR,后者需要审批。
文档核查时间为 2026 年 11 月 15 日。政策会变化,因此请对照所链接页面进行确认。OpenAI 和 AWS 指出,被标记为 CSAM 的内容即使在零保留模式下也可能被保留。
每一份留存副本都必须受到保护、被追踪并最终删除。从未写入磁盘的数据,日后也无法从磁盘中被取走。
监管方向也是如此,但有一点需要特别注意。根据 GDPR 第 5 条,临时处理有助于满足存储限制要求。数据最小化仍取决于你的应用发送了什么。
“只要工作流允许,临时处理都更可取。”——Fireworks AI 联合创始人兼 CEO Lin Qiao(发布于 X)
其理由与人為失误有关。临时处理不会留下被遗忘的数据集,也不会出现漏掉的删除调用。这还意味着需要保护和最终删除的留存副本更少。
存储型功能仍有其用武之地:基于数据集的微调需要静态数据(data at rest)。
延伸阅读:Lin Qiao Post、OpenAI 数据控制、Amazon Bedrock 数据保留、Google 零数据保留、Anthropic API 数据保留以及 Fireworks AI 零数据保留。
Amazon Bedrock
Fireworks AI
感谢 Fireworks 团队为本文提供的思想洞察与资源。本文为赞助内容。
Clearing Up the Confusion on AI Data Retention: Two Standard Designs Explained 一文最初发布于 MarkTechPost。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Clearing Up the Confusion on AI Data Retention: Two Standard Designs Explained)
Data retention has become one of the most misunderstood terms in AI infrastructure. After a wave of AI security headlines, many dev teams now ask a simple question. Does my AI provider keep my data? The honest answer depends on how you send it. Send a prompt to an inference API, and it is processed and returned without being kept. That design is called zero data retention (ZDR). Upload a dataset for a batch job, and it stays stored until you delete it. Both are standard designs, and in both the customer is in control. What changes is who manages the data lifecycle, and where.
Fireworks AI cofounder and CEO Lin Qiao set out a clear way to think about this. In a November 13, 2026 post on X, she described two standard API designs. Each one decides where control of the data lifecycle sits. The same split has shaped data processing systems for decades. In this article, we will go through that framework with examples from public API documentation.
TL;DR
- The framework: Two standard API designs: upload, use, delete, or send, compute, return.
- What it is: ZDR is a per-endpoint commitment not to persist covered prompts and outputs after an eligible request.
- Where it applies: Eligibility depends on the endpoint, request flags and account settings, not the vendor name.
- Stored features: Batch, file and dataset features keep data until it is deleted or expires.
- Bottom line: Prefer ephemeral endpoints. If you upload, deletion is your job, so never assume an expiry.
What is an AI data retention policy?
An AI data retention policy defines what a platform stores from your requests, and for how long. OpenAI’s data controls describe two layers. Application state is data a feature must keep, such as uploaded files. Abuse-monitoring logs are copies held for safety review. A platform can be at zero on one layer and not the other.
What are the two common API designs?
AI inference and training APIs follow two lifecycle designs. The difference is who manages the data, and where.
Pattern 1: Customer-managed retention (upload, use, delete)
You create a file or dataset, run jobs against it, then delete it. The platform keeps the data stored and available until you ask for deletion. It cannot know when you are finished unless you define it. So deletion is your job, done through the platform’s APIs.
Batch APIs with uploaded datasets are the standard example:
- OpenAI Batch: upload a JSONL input file, then create the batch.
- Amazon Bedrock batch inference: input JSONL files sit in an S3 location you choose.
- Gemini Batch API: jobs run from an input file you provide.
- Fireworks AI Batch: create a dataset, then submit jobs against it.
Once you delete the data, it is gone. According to Fireworks CEO, it keeps nothing after a deletion request.
OpenAI non-batch files persist until you delete them. Fireworks AI managed training datasets have no time limit, and the customer deletes them. Amazon Bedrock and Google batch jobs read and write your own cloud storage.
OpenAI batch files last 30 days by default, and Anthropic batch data 29 days. Treat expiry as the exception, and confirm it per feature.
Pattern 2: Ephemeral processing (send, compute, return)
You send data, the platform computes, and the result comes back. No application state stays behind. Your application owns the conversation history, so no platform cleanup step is needed. This is the pattern ZDR applies to. A synchronous response alone does not establish ZDR.
Documented examples:
- OpenAI Chat Completions: store defaults to false.
- Amazon Bedrock Converse: content is only used to generate the response.
- Google Gemini text generation: stateless when store is false.
- Fireworks AI Chat Completions: direct or through Nexus. The schema has no store field.
- Fireworks AI Training SDK, including Serverless Training: forward_backward sends tokenized batches per call. No dataset is uploaded, and training data is not stored at rest.
Think of saving a spreadsheet to a cloud drive versus passing its numbers to a calculator. The drive keeps the file for later. The calculator computes the answer, returns it and keeps nothing.
In both designs, customers fully control how their data is handled. What differs is where the lifecycle is managed. In pattern 1, it is explicit dataset creation and deletion. In pattern 2, it lives entirely in your application.
Does zero data retention cover every endpoint?
The answer is No. ZDR is scoped per endpoint, and defaults differ by platform.
OpenAI keeps abuse-monitoring logs for up to 30 days by default. The Amazon Bedrock none mode writes no request or response data to durable storage.
Fireworks AI documents ZDR by default for open models. Its ephemeral inference APIs, Training SDK and Nexus carry ZDR guarantees. Prompts and generations are not persisted at rest. For the Batch API, Fireworks document states that no data is retained after deletion.
A single parameter can also flip the pattern. On Google’s Interactions API, store defaults to true. At OpenAI, setting store to false is not ZDR, which needs approval.
| Platform | Synchronous inference: default | Zero retention option | Stored features (pattern 1) |
|---|---|---|---|
| OpenAI API | Abuse-monitoring logs are kept for up to 30 days, or longer where law requires. No application state on /v1/chat/completions, with exceptions. Chat Completions store defaults to false. | Modified Abuse Monitoring or ZDR, each by prior approval. Under ZDR, store is always treated as false. | Batch input files expire after 30 days by default. Batch output and error files expire 30 days after creation. Other files persist until deleted. /v1/files and /v1/batches are not ZDR eligible. Responses API: at least 30 days when store is true. |
| Amazon Bedrock | The Converse reference says content is only used to generate the response. The effective retention mode is the first non-inherit value across project, account and model default. New accounts and projects start in inherit. Models that require aws_review are retained for up to 30 days. | Set the mode to none. No request or response data is written to durable storage. | Batch inference reads and writes S3 locations you specify. Under none, requests to models that require retention are blocked. Model invocation logging to your own S3 or CloudWatch is off by default and customer-enabled. |
| Google Gemini Enterprise Agent Platform | In-memory and durable caching are on by default for supported Gemini models. Each clears within 24 hours of being written. Abuse-monitoring prompt logging applies to customers on the online Cloud terms. Master Agreement customers are exempt by default, except on designated frontier models. | Durable caching can be turned off while in-memory caching stays on. Leave request-response logging and Live API session resumption off. Set Interactions API store to false. | Batch inference uses Cloud Storage or BigQuery for input and output. Interactions API store defaults to true. On the separate Gemini API, stored interactions are kept for 55 days on the paid tier and 1 day on the free tier. |
| Anthropic Claude API | Standard policy: API inputs and outputs are deleted within 30 days. Content flagged for Usage Policy violations can be kept for up to 2 years. Covered Models: 30 days required. | ZDR by arrangement with sales, per organization. Not offered for Covered Models unless Anthropic expressly authorizes it. | Message Batches: retained 29 days. Files API: kept until deleted or expired. Neither is ZDR eligible. |
| Fireworks AI | ZDR by default for open models. Prompts and generations are not logged or stored without opt-in. They stay in volatile memory for the request. Prompt caches are held in memory only. Request metadata, such as token counts, is logged to run the service. Nexus uses ZDR on every account and plan. Turns it routes to your own Anthropic, OpenAI or Bedrock account follow that provider’s policy. | Already the default on Chat Completions, the Training SDK and Nexus. Enterprise admins can also enforce an account-wide ZDR policy. | Batch API: runs on uploaded datasets. Customers delete datasets through the console or the API. No data is retained after deletion. Managed training: uploaded datasets stay until the customer deletes them. |
Documentation checked November 15, 2026. Policies change, so confirm against the linked pages. OpenAI and AWS note that content flagged as CSAM can be retained even under zero retention.
Why is ephemeral processing the safer default?
Every retained copy must be protected, tracked and eventually deleted. Data that never reaches disk cannot be taken from disk later.
Regulation points the same way, with one important thing to note. Under GDPR Article 5, ephemeral processing supports storage limitation. Data minimisation still depends on what your application sends.
“Ephemeral processing is preferable whenever the workflow allows it.” Lin Qiao, cofounder and CEO, Fireworks AI (post on X)
The reasoning is about human error. Ephemeral processing leaves no forgotten datasets and no missed delete calls. It also means fewer retained copies to protect and eventually delete.
Stored features still have a place: dataset-based fine-tuning needs data at rest.
What should developers check now?
- Map every endpoint you call to pattern 1 or pattern 2.
- Default to ephemeral endpoints when you only need the result.
- For each upload, plan the delete call. Record any documented expiry. OpenAI supports expires_after.
- Audit flags that persist data, such as store.
- Keep secrets out of prompts and datasets. Rotate any that slip in.
- Define ZDR per processing path in contracts. Then verify the setting in your account.
Key Takeaways
- ZDR is an endpoint property, not a vendor label.
- Stored features keep data until it is deleted or expires.
- Prefer ephemeral when you only need the result.
- If you upload, deletion is your job. Never assume an expiry.
Further reading: Lin Qiao Post, OpenAI data controls, Amazon Bedrock data retention, Google zero data retention, Anthropic API data retention and Fireworks AI zero data retention.
Sources:
OpenAI
- Data controls in the OpenAI platform: https://developers.openai.com/api/docs/guides/your-data
- Chat Completions reference (create, store default): https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create
- Batch API guide: https://developers.openai.com/api/docs/guides/batch
- Batch API guide, upload your batch input file: https://developers.openai.com/api/docs/guides/batch#2-upload-your-batch-input-file
- Files API reference (create, expires_after): https://developers.openai.com/api/reference/typescript/resources/files/methods/create
Amazon Bedrock
- Data retention: https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html
- Converse API reference: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_Converse.html
- AWS Security Blog, Enforce zero data retention on Amazon Bedrock (July 7, 2026, updated August 13, 2026): https://aws.amazon.com/blogs/security/enforce-zero-data-retention-on-amazon-bedrock-with-bedrock-projects-and-service-control-policies/
- Create a batch inference job: https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference-create.html
- Format and upload batch inference data: https://docs.aws.amazon.com/bedrock/latest/userguide/batch-inference-data.html
- Model invocation logging: https://docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html
- Gemini Enterprise Agent Platform and zero data retention (updated November 12, 2026): https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention
- Abuse monitoring: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/abuse-monitoring
- Batch inference from Cloud Storage: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/capabilities/batch-prediction-from-cloud-storage
- Gemini API, Text generation: https://aistudio.google.com/docs/text-generation?codelanguage=rest
- Gemini API, Batch API input file: https://aistudio.google.com/docs/batch-api#input-file
- Gemini API, Interactions API overview (data storage and retention): https://ai.google.dev/gemini-api/docs/interactions-overview
Anthropic
- API and data retention: https://platform.claude.com/docs/en/manage-claude/api-and-data-retention
- How long do you store my organization’s data: https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
- Data retention practices for Covered Models: https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models
Fireworks AI
- Zero Data Retention (data handling): https://docs.fireworks.ai/guides/security_compliance/data_handling
- Zero Data Retention policy (Enterprise account setting): https://docs.fireworks.ai/accounts/zero-data-retention
- Secure Training: https://docs.fireworks.ai/guides/security_compliance/secure_training
- Chat Completions reference: https://docs.fireworks.ai/api-reference/post-chatcompletions
- Training SDK reference, forward_backward: https://docs.fireworks.ai/fine-tuning/training-api/reference/service-client#forward_backward-datums-loss_type-loss_fn_config=none
- Batch API: https://docs.fireworks.ai/guides/batch-inference#getting-started
- Delete Dataset reference: https://docs.fireworks.ai/api-reference/delete-dataset
- Serverless Training: https://docs.fireworks.ai/fine-tuning/training-api/serverless
- Prompt caching: https://docs.fireworks.ai/guides/prompt-caching
- Nexus, Data Security and Privacy: https://docs.fireworks.ai/nexus/data-security
- Nexus, APIs and SDKs: https://docs.fireworks.ai/nexus/apis-and-sdks
- Security bulletin, Update on October 2026 security incident (Training SDK: training data not stored at rest): https://fireworks.ai/security/bulletins/october-2026-security-incident-update
- Fireworks statement on data retention, “Data retention: the control and standard” (November 13, 2026): https://x.com/lqiao/article/2121434466082242896
Other
- Regulation (EU) 2016/679 (GDPR): https://eur-lex.europa.eu/eli/reg/2016/679/oj
Thanks to the Fireworks team for the thought leadership / resources for this article. This article is sponsored.
The post Clearing Up the Confusion on AI Data Retention: Two Standard Designs Explained appeared first on MarkTechPost.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(Clearing Up the Confusion on AI Data Retention: Two Standard Designs Explained),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。