Anthropic built its reputation as the privacy-conscious alternative to OpenAI. Constitutional AI, safety-first research, no training on customer data. For a while, that reputation was earned.
Then in August 2025, Anthropic quietly announced a policy change: consumer users would need to actively opt out if they didn't want their Claude conversations used to train future models. The deadline was September 28. If you missed it, your data was in.
If you're evaluating AI providers right now, this is the context you need.
What Does Claude Store by Default?
For consumer accounts, that are Claude Free, Pro, and Max, conversations are saved to your account until you delete them. Once deleted, they're removed from your chat history immediately but remain on Anthropic's back-end systems for up to 30 days before being permanently deleted.
That's the standard case. A few exceptions matter:
Policy violations. If a conversation is flagged for violating Anthropic's usage policy, the inputs and outputs are kept for 2 years. Trust and safety classification scores from that conversation are retained for 7 years.
Feedback you submit. Thumbs up/down ratings and bug reports are kept for 5 years.
Incognito mode. Conversations in Claude's incognito mode are never used for model training, regardless of your other settings.
The September 2025 Training Policy Change
Anthropic's previous stance was clean: consumer chats would not be used for training. That was the explicit promise. In August 2025, that changed. According to Anthropic's own announcement, they introduced an opt-in toggle, "You can help improve Claude", and gave users until September 28 to make their choice.
If you opted in, Anthropic could retain your conversations in de-identified form for up to 5 years and use them for model training. If you opted out, nothing changed. There's still a 30-day retention, no training use.
The reaction was immediate. Security researchers and privacy advocates flagged it as a "privacy pivot." The opt-in training setting extends data retention from 30 days to 5 years, a 60x increase in how long your conversations can sit in Anthropic's training pipeline.
How to Check and Change Your Settings
To see where your account stands: Claude.ai → Settings → Privacy → "Improve Claude for everyone."
If the toggle is on, your new conversations are eligible for training and retained for up to 5 years. Turning it off returns you to the 30-day retention standard.
Turning it off does not retroactively remove data already used for training. Like OpenAI, Anthropic doesn't unlearn from data once it's been incorporated.
How Is the Anthropic API Different for Privacy-Conscious Users?
The consumer product and the API have meaningfully different data policies, and the API is notably stronger.
The API's standard retention window is 30 days. Anthropic automatically deletes inputs and outputs from its backend within 30 days of receipt or generation, and none of it is used for model training without your express permission. (A shorter 7-day API window that Anthropic rolled out in September 2025 is no longer current; the commercial baseline has been back at 30 days for a while, per Anthropic's own Privacy Center.)
For enterprise API customers who qualify, Anthropic also offers a Zero Data Retention agreement. Under ZDR, inputs and outputs aren't stored at rest after the API response returns. One caveat that still applies: Anthropic can retain Trust and Safety classifier results even under ZDR, to enforce its usage policy.
There's a newer, sharper exception too. Since June 2026, Anthropic designates its most capable models, currently Claude Fable 5, Fable 5.1, Mythos 5, and Mythos 5.1, as "Covered Models". Any organization using one of these gets a mandatory 30-day retention window that overrides ZDR entirely, unless Anthropic expressly authorizes an exception. Anthropic's stated reasoning: some abuse patterns, like distributed jailbreak attempts or coordinated misuse campaigns, only become visible when requests are compared across a window of time rather than one at a time. Practically, if you're on a ZDR agreement and want to use a Covered Model, you need to turn on 30-day retention for that specific workspace; your other ZDR workspaces stay at zero retention. Anthropic's non-Covered models, including Claude Opus 5, still get full ZDR with no exceptions. Most meeting-summarization workloads, Anarlog's included, run fine on non-Covered models, so this mainly matters if your organization has specifically opted into the newest reasoning-class models for other work.
Claude for Work, Enterprise, Edu, and Gov
Commercial customers were explicitly excluded from the September 2025 consumer policy changes. If you're using Claude for Work, Claude Enterprise, Claude for Education, or Claude Gov, your data is not used for model training — that's covered under Anthropic's commercial terms and is not subject to opt-in or opt-out toggles.
Deleted conversations are purged within 30 days unless legally required otherwise.
What About HIPAA and GDPR?
HIPAA: Anthropic offers HIPAA-eligible services for qualifying healthcare customers, including a Business Associate Agreement. Web search itself is now HIPAA-eligible under the BAA; the current exception is web search's "dynamic filtering" sub-feature, which isn't covered. Covered Models (see above) add another wrinkle: they're only accessible under the BAA through the HIPAA-ready API or HIPAA-ready Claude Enterprise chat, not through Claude Code or Cowork. Standard consumer Claude is not HIPAA-compliant and should not be used with Protected Health Information.
GDPR: Anthropic supports GDPR compliance for commercial customers through a Data Processing Addendum. For EU-based consumer users, standard GDPR rights apply including access, deletion, portability, and can be exercised through Anthropic's Privacy Center. Consumer accounts don't automatically come with a DPA.
Claude's Reddit Lawsuit Worth Knowing About
In June 2025, Reddit filed a lawsuit against Anthropic alleging that Anthropic scraped more than 100,000 Reddit posts and comments without authorization to train Claude. Reddit presented evidence that Claude reproduced deleted Reddit posts with near-perfect accuracy.
The case has moved since then. Anthropic tried to move it to federal court, arguing Reddit's claims were really just copyright infringement in disguise. In March 2026, a federal judge disagreed and sent the case back to California state court, ruling that Reddit's breach-of-contract, unjust enrichment, and privacy-related claims raise "extra elements" beyond what copyright law covers: specifically, that Anthropic allegedly bypassed Reddit's technical safeguards and interfered with privacy commitments Reddit owes its own users. By August 2026, the presiding state judge was reportedly skeptical of Anthropic's remaining defenses on the privacy claims. The case is still active and unresolved as of this writing.
This isn't about your personal data directly. But it's relevant context for how Anthropic has approached training data acquisition, and it matters when evaluating whether a company's stated privacy values match their actual behavior.
How Claude.ai Compares to Using Claude Through Anarlog
| Claude.ai (Consumer) | Anthropic API via Anarlog | |
| Where notes are stored | Anthropic's servers | Your device (local SQLite) |
| Retention after deletion | Up to 30 days on backend | 30 days by default, 0 under ZDR* |
| Used for training | Yes — unless opted out (Sept 2025) | Never |
| If opted into training | Up to 5 years, de-identified | Not applicable |
| HIPAA BAA available | No (consumer) | Depends on the AI provider and contract |
| Can switch AI providers | No | Yes — OpenAI, Mistral, local models |
*Covered Models (Claude Fable/Mythos-class) are the exception: they carry a mandatory 30-day retention regardless of ZDR. Anarlog doesn't require those models for meeting summaries.
Anarlog is an open-source AI notepad for meetings that lets you bring your own Anthropic API key. When you connect it, your meeting data goes through the API, so it follows the 30-day (or zero, under ZDR) API retention, never used for training, rather than through the consumer Claude.ai product where training opt-ins and longer retention windows apply.
And if Anthropic's policy trajectory gives you pause, you're not locked in. Anarlog supports OpenAI, Mistral, Google Gemini, and local models via Ollama. Your notes stay on your device regardless of which AI processes them. Switching providers doesn't mean starting over.
That's what actual control looks like. It's not just a privacy toggle that defaults to whatever Anthropic decides next quarter.
Download Anarlog and use the AI provider your security team actually trusts.