Before putting personal, client, or confidential information into an AI service, identify the exact product, account type, deployment route, and features you will use. Then check separately what the provider may use the data for, how long each kind may remain, what deletion does, and which tools or third parties also receive it. “Not used to train models” does not mean “not stored,” and a delete button does not necessarily erase every copy immediately.
What happens to data after you send it to an AI service?
There is no single answer for a provider’s entire brand. A consumer chat app, business interface, API endpoint, cloud-hosted deployment, and optional integration can have different terms, retention behavior, and controls. Even within one service, prompts, files, feedback, safety records, and application state may be treated differently.
As an Amazon Associate I earn from qualifying purchases.
Use the provider’s current documentation and your organization’s applicable contract to establish the rules for the exact configuration. The examples below cover selected OpenAI API, Anthropic API and commercial/consumer offerings, and Google Gemini Apps and Workspace materials; they do not establish the rules for every provider or settle jurisdiction-specific legal obligations.
What exactly are you evaluating?
Write down the scope before comparing privacy claims. Include the product and tier, account type, interface or API, deployment route, organization, models and endpoints, enabled tools, and any marketplace or third-party integration. This matters when data passes through another platform: Anthropic says that Amazon Bedrock and Google Cloud Agent Platform use the cloud provider as data processor, while its first-party API documentation describes Anthropic as processor. Anthropic’s API retention documentation explains the distinction.
#1 Best Overall
For each candidate, use a scope record such as:
- Product and account: consumer, business, enterprise, API, or a named cloud deployment.
- Route: interface, endpoint, project, model, and region, where specified.
- Features: file uploads, voice or video, memory, connectors, agents, feedback, or other tools that may change what is stored or shared.
- Owner: the organization responsible for configuring controls and reviewing terms.
What data will the service and its connected tools handle?
Inventory both what people submit and what the service creates around those submissions. A prompt is only one possible data category. Google’s Gemini Apps Privacy Hub, for example, lists prompts and uploads, generated content, app, browser and device information, connected-app information, location, and subscription information. Google’s Privacy Hub describes the categories for Gemini Apps.
- Prompts, outputs, and conversation history
- Files, images, audio, video, and other uploads
- Feedback, account identifiers, and usage metadata
- Connected-app content, tool calls, and generated summaries
- Application or feature state, such as stored files or other data needed to operate a feature
Trace each connection beyond the primary provider. OpenAI notes that remote MCP servers are third parties and that their retention policies apply to data sent to them. OpenAI’s API data-controls documentation describes endpoint and third-party considerations. Map cloud platforms, connected applications, MCP servers, and other subprocessors, then check the terms that govern each recipient.
What may the provider do with each category?
Ask about purposes separately rather than relying on a broad phrase such as “your data is protected.” For each category, find whether it is used to provide the requested service, maintain or secure it, prevent abuse, conduct human review, improve products, train models, personalize outputs, or meet legal obligations. Check which setting or contract controls each purpose, whether a change applies only to future data, and what happens to feedback or interactions flagged for policy enforcement.
Training and storage are separate questions. A provider may say data is not used for model training while retaining it for service operation, safety, abuse prevention, feedback, or legal reasons. Anthropic’s API documentation states, in the context of API retention, “Retained data is never used for model training without your express permission”; that statement should not be generalized to every Anthropic product. Read the API documentation in its stated scope.
Consumer policy settings may differ from commercial and API terms. Anthropic’s August 28, 2025 announcement says users of its Free, Pro, and Max consumer plans can choose whether data is used to improve Claude. It says the change does not apply to commercial services or API use. For consumer users who allow model-training use, the announcement describes a five-year retention period for new or resumed chats and coding sessions; users who do not make that choice remain on the stated 30-day period. These are consumer-policy statements, not a rule for Anthropic business or API accounts. Anthropic’s announcement gives the scope and date.
How long can different data remain?
Build a retention table by data type, not by provider alone. Record the ordinary period, when its clock starts, whether the data appears in user-visible history, backend deletion timing, who can delete it, and exceptions. The examples below are not directly interchangeable: they refer to different products, data categories, and retention mechanisms.
| Offering and data category | Published period or rule | Scope and qualification |
|---|---|---|
| Anthropic commercial API inputs and outputs | Within 30 days | Anthropic’s commercial FAQ says automatic backend deletion is within 30 days of receipt or generation, subject to longer user-controlled feature retention, agreement, Usage Policy enforcement, or law. Commercial retention FAQ. |
| Anthropic commercial product: deleted chats | Visible history immediately; backend systems within 30 days | The FAQ says deleted chats disappear from visible history immediately and are deleted from backend storage systems within 30 days. It also says flagged Usage Policy violations may result in inputs and outputs being retained for up to 2 years and trust-and-safety classification scores for up to 7 years. Commercial retention FAQ. |
| Anthropic consumer Free, Pro, and Max: chats and coding sessions | Five years if the user allows model-training use; otherwise the stated 30-day period continues | Anthropic’s announcement dated August 28, 2025 describes this consumer-policy distinction for new or resumed sessions. It does not describe commercial or API retention. Consumer terms announcement. |
| Google Gemini Apps: temporary chats and chats with Keep Activity off | 72 hours | The Privacy Hub, last updated September 24, 2026, says these chats are retained with the account for 72 hours for response and protection purposes. It separately says reviewed chats and related data may be retained for up to 3 years after activity deletion. Gemini Apps Privacy Hub. |
| Google Gemini Apps Activity auto-delete setting | 18-month default; options include 3 months, 36 months, or indefinite | The same Hub describes an activity setting, not a retention promise for every data category. Gemini Apps Privacy Hub. |
| OpenAI API abuse-monitoring data | 30 days for several endpoints | OpenAI’s endpoint table shows different retention and application-state behavior by endpoint; other endpoints have different settings or no abuse-monitoring retention. This is not one universal API retention rule, and Zero Data Retention eligibility also varies by endpoint. API data-controls documentation. |
Do not compare these figures as if they measure the same thing. A history auto-delete setting, a backend deletion window, an abuse-monitoring period, and retention after a policy flag are different kinds of commitments.
What does deletion or “zero retention” actually cover?
Find out what a user’s delete action changes immediately and what happens later in backend systems. Ask specifically about backups and recovery systems, files, feedback, flagged content, safety classification, vector stores, retained summaries, legal holds, and data already sent to tools. A control’s name is not enough: check its covered categories, exclusions, timing, and exceptions.
“Zero Data Retention” is not necessarily a blanket setting across every product feature. OpenAI says approved organizations can set retention controls at organization and project levels, but ineligible endpoints or capabilities may retain application state even when Zero Data Retention is enabled. Anthropic says ZDR is enabled per organization and covers only eligible API features; it does not cover all consumer and commercial interfaces or third-party integrations. OpenAI’s endpoint guidance and Anthropic’s API retention documentation describe their respective limits.
Rank #4
Deletion exceptions matter too. Anthropic identifies policy-enforcement and legal exceptions in its commercial retention FAQ. Google says human-reviewed conversations are not deleted when Gemini activity is deleted. OpenAI’s endpoint documentation describes application-state retention and endpoint-specific differences. Read the relevant terms for the product and feature you actually plan to use rather than assuming that deleting visible history removes every associated record.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Are the controls available for your account and deployment?
Confirm who can enable a control—the end user, administrator, or provider under a contract—and whether it requires approval. Check if it is organization-wide or project-level, and whether the specific endpoint, model, tool, and mode qualify. Record any settings that need an administrator or contractual change before users submit sensitive information.
Also distinguish consumer services from business protections. Google’s materials say qualifying Workspace use is not human reviewed or used for generative AI model training outside the customer’s domain without permission. Its FAQ warns that users without qualifying Workspace licenses are subject to different terms when using the Gemini app. Google’s Workspace FAQ explains that distinction. Do not infer Workspace protections from a personal Gemini account, or the reverse.
Best Value
How should you compare candidate services?
Use one row per exact product and configuration, then compare the same data categories and purposes. A compact vendor-review table can include:
| Comparison axis | What to record |
|---|---|
| Scope | Product, tier, interface or endpoint, deployment route, models, tools, organization, and evidence date |
| Data categories | Prompts, outputs, uploads, feedback, identifiers, metadata, feature state, and data sent to integrations |
| Use | Service operation, safety, human review, improvement, training, personalization, and legal purposes |
| Retention and deletion | Period by category, when the clock starts, user-visible behavior, backend timing, deletion controls, and exceptions |
| Controls and exclusions | Who can configure controls, approval requirements, eligible endpoints or features, and any state not covered |
| Other recipients and commitments | Subprocessors, cloud platform, connected tools, applicable contract, and any regional commitment established in the terms |
When a source does not specify an item, write “not stated in the reviewed source” and follow up with the provider rather than filling the gap with an assumption. A shorter retention period in one category does not by itself make a service more private if another service has a different scope, stronger controls, or different exceptions.
What should you do when the terms do not answer your use case?
- Ask the provider to confirm the applicable terms for your precise account, deployment, endpoints, and enabled features.
- Request clarification on any unstated category, exception, deletion timeline, subprocessor, or control limitation that affects your data.
- Have security or privacy staff review configuration, access, integrations, and the data flow before adoption.
- Obtain qualified legal review when your industry, jurisdiction, contract, or data type creates specific obligations; provider documentation alone does not determine compliance.
The cited materials are selected provider documentation, not a complete comparison of every AI service. They do not establish data-residency terms, contractual protections, legal compliance, or the current terms of Microsoft Copilot and other providers. Verify the primary terms that govern your actual service and configuration.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
