Skip to main content
Provider Comparison

Your AI prompts are not private by default

Every prompt you send to OpenAI, Anthropic, or Google passes through their infrastructure — and each provider handles your data differently. Understand what they log, what they train on, and how to lock it down before your data ever leaves your machine.

Quick Answer

OpenAI, Anthropic, and Google all say they do not train on your API data — but each handles data retention, abuse monitoring, and ZDR differently. None of them default to zero data retention. Your prompts and responses are logged for abuse monitoring (30 days by default for OpenAI, not retained by default for Anthropic, limited time for Google's paid tier). The safest approach: protect your data before it leaves your machine — so it never reaches the provider in the first place.

OpenAI Data Handling
API tier — consumer products have separate terms
Training on API DataNot used for training (since March 1, 2023)

Data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in. This applies to all API endpoints. Consumer products (ChatGPT free/plus) have separate terms.

Default RetentionUp to 30 days

Abuse monitoring logs may contain prompts, responses, and metadata. Retained for up to 30 days by default — longer if legally required or to protect services from harm. Audio endpoints and the moderations endpoint have zero-day retention.

Zero Data RetentionAvailable

Zero Data Retention (ZDR) is available by contacting the OpenAI sales team. When enabled, eligible endpoints do not retain any customer content. ZDR-eligible endpoints include chat/completions, responses, embeddings, audio, moderations, completions, and realtime. Not eligible: conversations, assistants, vector stores, files, fine-tuning, batches, and video.

HIPAA / BAAAvailable

Business Associate Agreement (BAA) available for eligible endpoints. Requires ZDR or Modified Abuse Monitoring (MAM). HIPAA eligible endpoints include chat/completions and embeddings (with ZDR).

What Gets Stored
Stored
  • Prompts and responses (30 days, abuse monitoring)
  • Application state for conversations and assistants (until deleted)
  • Files and vector stores (until deleted, 30-day grace after deletion)
  • Fine-tuning job data (until deleted)
Not Stored
  • Audio transcription/speech content (0 days)
  • Moderation endpoint content (0 days)
  • ZDR-enabled endpoints (no storage, eligible endpoints only)
  • Training data from API usage (never, since March 2023)
Audit Rights

Enterprise agreement required for audit rights. Standard API terms do not include independent audit provisions.

Region Controls

Data processed in US data centers. Enterprise customers can request regional processing through sales. No self-service region selection for standard API users.

Key Caveats
  • Image/file inputs are scanned for CSAM — if flagged, retained for manual review regardless of ZDR/MAM settings
  • Prompt caching uses GPU-local storage up to 24 hours
  • Conversations, assistants, and vector stores are NOT eligible for ZDR
  • ZDR must be requested through sales — not self-service

Side-by-Side Comparison

DimensionOpenAIAnthropicGoogle
Training on API DataNo (since Mar 2023)No (never)Paid: No / Free: Yes
Default RetentionUp to 30 daysNot retained (standard)Paid: limited / Free: may retain
Zero Data RetentionAvailable (sales)Available (sales)Not available
HIPAA / BAAAvailableAvailableVertex AI only
Audit RightsEnterprise onlyEnterprise onlyEnterprise Cloud
Region ControlsUS default / enterpriseUS default / AWS regions40+ regions (Cloud)
CSAM ScanningYes (all images/files)Not disclosedPer safety policy
Consumer vs. APISeparate termsSeparate termsSeparate terms

Where Your Data Goes

Every prompt flows from your machine to the provider's cloud. Without Shield, sensitive data reaches the provider before any redaction happens. With Shield, PII and secrets are stripped on your machine — the provider never sees them. ZDR helps, but it only affects what happens after the provider receives your data.

Your MachinePrompt + DataSHIELD (optional)Redacts PII/SecretsBefore data leavesPROVIDER CLOUDOpenAI / AnthropicGoogle⚠ DEFAULT PATHLogs retained 30 days✓ ZDR PATHNo data at restTRAINING?API: No (all 3 providers)Your NetworkExternal Internet

The critical insight: provider data policies only matter for data they receive. Shield prevents data from reaching them at all.

Common Questions

No — all three major providers have policies against using API data for training without consent. OpenAI stopped using API data for training in March 2023. Anthropic has never used API data for training without express permission. Google does not use Paid Service data for training. The critical distinction: these policies apply to API usage — consumer products like ChatGPT, Claude.ai, and Gemini (free) have different terms. Always use the API tier (not consumer products) for business data, and verify your organization's data usage settings.
Zero Data Retention (ZDR) means the provider does not store your prompts and responses at rest after the API response is returned. Both OpenAI and Anthropic offer ZDR — but you must contact their sales teams to enable it. OpenAI's ZDR covers chat/completions, embeddings, audio, and moderations (not conversations, assistants, or files). Anthropic's ZDR covers the Messages API and Claude Code (not Console, Managed Agents, or consumer products). Google does not offer a standalone ZDR toggle — enterprise customers use Cloud data residency controls instead.
Not completely — and this is the most important nuance. Even on paid API plans, providers retain data for abuse monitoring. OpenAI retains data for up to 30 days. Anthropic retains Covered Model data for 30 days. Google logs paid API data for a limited time for abuse detection. The data is not used for training, but it is briefly stored. If you need zero retention, you must explicitly request ZDR from OpenAI or Anthropic. This is not the default — you have to opt in.
All three providers support HIPAA, but with important differences. OpenAI requires ZDR or MAM plus a signed BAA for HIPAA workloads — and only chat/completions and embeddings are eligible. Anthropic requires a signed BAA for HIPAA readiness, covering the Claude API. Google's HIPAA support is through Vertex AI (not the standalone Gemini API). The common thread: you need a BAA, you must use the API tier (not consumer products), and you should verify which specific endpoints are covered.
Start with the provider's published data usage documentation. OpenAI's data controls page lists every endpoint's retention behavior. Anthropic's API and data retention docs cover ZDR and HIPAA eligibility. Google's Gemini API terms distinguish paid vs. unpaid data usage. Beyond documentation: request a DPA (Data Processing Addendum), ask for SOC 2 or ISO 27001 audit reports, and — for enterprise agreements — negotiate audit rights. If you're sending sensitive data, Shield adds a layer that redacts it before any provider ever sees it.
Consumer products (ChatGPT, Claude.ai, Gemini app) and API products have completely separate data policies. Consumer product data may be used for training, reviewed by humans, or retained differently. API data is generally not used for training. This is why organizations should never let employees paste business data into consumer AI tools — it bypasses the API's data protections entirely. Shield helps enforce this boundary by sitting at the API layer and catching sensitive data before it reaches any provider.

Related Articles

LLM Data Exposure: The Hidden Supply Chain
Interactive map of the AI supply chain. See what leaks at every layer.
AI Data Residency & Sovereignty
Where your data physically lives, which laws apply, and how to keep it in your jurisdiction.
On-Prem vs. Cloud AI Security: The Trade-offs
Compare sovereignty, compliance, latency, cost, and control across deployment models.

Don't bet on provider policies alone

Every provider says they protect your data — but they all log it, retain it, and scan it. Shield stops your passwords, customer data, and company secrets from ever leaving your computer in the first place. Provider policies become irrelevant when the data never reaches them.

Talk to Our TeamHow Shield Works

Last updated: July 19, 2026