Published July 6, 2026 · Policies last verified July 1–6, 2026 · Next scheduled review: October 2026
Key findings: we reviewed the published privacy policies, data-controls documentation, and help-center pages of 13 AI chat services. As of July 2026: 9 of 13 involve consumer conversations in model training by default or as the preselected choice (opting out is possible in most, not all); retention of a "deleted" conversation ranges from immediate local-only deletion to 30 days at most major providers; retention of conversations you never delete ranges from zero server-side storage to 5 years — or indefinite; and only 4 of 13 services are architected so that no readable conversation archive exists on the provider's servers at all.
A definition, because marketing blurs it: data retention in AI chat means how long the provider keeps your prompts, responses, and uploaded files on its servers — which is a separate question from whether it trains on them, and separate again from whether a human may review them. A service can honestly say "we don't train on your chats" while storing them, readable, for years. The table below keeps the three questions apart.
The Audit Table
Consumer tiers, default settings, as documented by each provider. Enterprise and API terms are usually stricter and noted separately below the table.
| Service | Where chats are stored | Retention (if you never delete) | After you delete | Used for training (consumer default) |
|---|---|---|---|---|
| ChatGPT (OpenAI) | OpenAI servers, tied to account | Indefinite | Scheduled for permanent deletion within 30 days; Temporary Chats kept up to 30 days | Yes, unless you turn off "Improve the model for everyone" |
| Claude (Anthropic) | Anthropic servers, tied to account | Up to 5 years if you allow training; ~30-day deletion cycle if you decline | Deleted per retention schedule (~30 days) | User choice at signup — allowing training is the preselected option |
| Gemini (Google) | Google servers (Gemini Apps Activity) | 18 months by default (user-adjustable 3–36); up to 72 hours even with Activity off | Removed from account; human-reviewed conversations kept up to 3 years, separately | Yes, with Activity on (opt-out via Activity off) |
| Grok (xAI) | xAI servers, tied to account | Indefinite | Deleted conversations removed within 30 days | Yes, opt-out available in settings |
| Meta AI | Meta servers, tied to your Meta accounts | Indefinite / unspecified | Deletion tools exist; interactions may persist in derived data | Yes; no opt-out for US users (EU/UK objection process exists) |
| Microsoft Copilot (consumer) | Microsoft servers, tied to account | Until deleted (conversation history) | Removed from history per Microsoft's deletion process | Yes in most regions, opt-out in settings (EEA excluded from training rollout) |
| Perplexity | Perplexity servers, tied to account | Until deleted | Removed per policy | Yes by default ("AI Data Usage" toggle, can be turned off) |
| Mistral Le Chat | Mistral servers (EU), tied to account | Until deleted | Deleted per GDPR erasure rights | Free tier: yes with opt-out; paid tier: no |
| DeepSeek | Servers in China | "As long as necessary" (unspecified) | Unclear; policy gives no firm timeline | Yes; no meaningful opt-out documented |
| Duck.ai (DuckDuckGo) | Your device only | Local until you clear it; upstream providers capped at 30 days, anonymized | Instant (local deletion) | No — barred by DuckDuckGo's provider agreements |
| Proton Lumo | Proton servers, zero-access encrypted | Ciphertext Proton cannot read, until you delete | Deleted; was never readable server-side | No |
| Venice AI | Your browser only | Zero server-side storage | Instant (local deletion) | No |
| Secret Chat AI | Your browser only (IndexedDB chats, OPFS files) | Zero server-side chat archive | Instant (local deletion); provider-side deletion requested per message and documented | No — providers process prompts under no-training/API terms |
Disclosure: Secret Chat AI is our product. Its row is held to the same standard as the rest — including the honest caveat that prompts are processed by the selected frontier-model provider at inference time, with deletion or no-storage handling requested and the outcome documented in a per-message privacy report.
What the Numbers Mean
1. "Deleted" almost always means "deleted in ~30 days"
OpenAI schedules deleted chats for permanent removal within 30 days (see OpenAI's chat and file retention policy); xAI documents the same 30-day window; Anthropic's deletion cycle is comparable. Exceptions survive everywhere for security review and legal holds — notably, a 2024–2025 US court preservation order forced OpenAI to retain even deleted consumer chats for a period, a useful reminder that provider policy is not the only force acting on your data.
2. The training default has quietly become "yes"
Nine of thirteen services involve consumer conversations in training by default or as the preselected choice. The direction of travel is one-way: Anthropic moved from no-training-by-default to a 5-year-retention training choice in September 2025; Microsoft began training on consumer Copilot conversations in most regions from early 2025; Meta offers US users no opt-out at all. The consumer default across the industry is now: your conversations are training data unless you act.
3. Human review has its own calendar
Google is the clearest documented case: conversations sampled for human review are kept up to 3 years, and that copy is not removed by deleting your activity. Most providers reserve similar safety-review rights with less specific timelines. If a human may read it and a separate copy may outlive your delete button, paste accordingly.
4. Only architecture removes the question
Four services — Duck.ai, Proton Lumo, Venice AI, and Secret Chat AI — are built so no readable conversation archive exists server-side (local-only storage, zero-access encryption, or both). Everything else is policy: real, but revocable, and revised more than once in the last twelve months. The practical rule for sensitive work: prefer tools where privacy is structural, and treat policy-based promises as valid only as of their verification date. Our head-to-head comparisons walk through the four architectural options individually.
Methodology
- Scope: 13 consumer AI chat services with meaningful market presence in July 2026.
- Sources: each provider's privacy policy, data-controls / help-center documentation, and official privacy announcements, as published between and verified on July 1–6, 2026. Representative sources: OpenAI's retention and Temporary Chat help pages, Google's Gemini Apps Privacy Hub, Anthropic's consumer-terms update (September 2025), DuckDuckGo's Duck.ai help pages, Proton's Lumo documentation, Microsoft's Copilot privacy FAQ.
- Defaults only: the table reflects consumer tiers at default settings. Enterprise plans (ChatGPT Enterprise, Claude for Work, Gemini for Workspace) and API access generally exclude training and shorten retention — the documented pattern is that API and enterprise terms are stricter than consumer terms at every major provider.
- Change policy: AI privacy policies changed frequently in 2024–2026; this page is re-verified quarterly and the verification date is updated at the top. If you spot a change we have not reflected, tell us via the contact page.
How to Cite This Audit
Cite as: "Who Stores Your AI Chats — and for How Long? The 2026 Data Retention Audit," Secret Chat AI, July 2026, secret-chat.ai/research/ai-chatbot-data-retention-audit-2026/. Journalists and researchers are welcome to reuse the table with attribution; if you need the underlying policy excerpts we collected, ask via the contact page.
Frequently Asked Questions
- Which AI chatbot stores your conversations the longest?
Among documented figures: Anthropic's Claude keeps conversations up to 5 years if you allow training, and Google keeps human-reviewed Gemini conversations up to 3 years. Meta AI and standard ChatGPT history are effectively indefinite until you delete. The shortest documented windows are the local-only tools (Duck.ai, Venice AI, Secret Chat AI), which keep nothing server-side.
- Does deleting a ChatGPT conversation really delete it?
Mostly, on a delay: OpenAI schedules deleted chats for permanent deletion within 30 days, with carve-outs for security and legal obligations — and litigation holds can extend that, as the 2024–2025 preservation order showed. Details in Does ChatGPT Store Your Conversations?
- Which AI chatbots do not train on your conversations?
By architecture or firm policy as of July 2026: Duck.ai, Proton Lumo, Venice AI, and Secret Chat AI. By configuration: ChatGPT with training disabled, Claude with training declined, Gemini with Activity off, Grok/Perplexity/Mistral with their opt-outs applied — and every major provider's enterprise or API tier.
- Is a "no training" promise the same as privacy?
No. A provider can decline to train on chats while storing them readable for years, sampling them for human review, and disclosing them under legal process. Retention, training, and human access are three separate questions — which is why this audit tracks them separately.