← Back to ThinkRoger.ai
KI & Privacy Guide

LLM Privacy & Tool Selection

Part 1: disable training on your data in the major cloud chatbots. Part 2: which tool for which job and effort level. Verified July 2026.

ChatGPT (OpenAI)

Applies to Free, Plus, Pro · account-wide across devices
Default: training ON
  1. Click your profile picture (bottom-left on web)
  2. Open Settings → Data Controls
  3. Toggle OFF Improve the model for everyone
Chat history stays intact. Temporary Chats are never used for training. Team/Enterprise/API are excluded by default.

Claude (Anthropic)

Applies to Free, Pro, Max · covers chats and coding sessions
Default: choice at signup, ON if accepted
  1. Click your initials/name (bottom-left)
  2. Open Settings → Privacy
  3. Toggle OFF Help improve Claude
Turning it off also reduces data retention for new chats. Conversations flagged by safety classifiers may still be reviewed for trust & safety.

Gemini (Google)

Setting is called "Keep Activity" - controls both history and training
Default: training ON
  1. Go to myactivity.google.com/product/gemini
  2. Toggle OFF Keep Activity
  3. Optionally delete past activity and set auto-delete to 3 months
  4. App: profile picture → Gemini Apps Activity → Turn off
Even when off, chats are kept 72 hours for service operation. Chats already reviewed by human raters are retained up to 3 years, disconnected from your account.

Microsoft Copilot

Consumer Copilot (copilot.com, Windows, macOS, mobile)
Default: training ON (consumer)
  1. Click your profile icon → your profile name
  2. Open Privacy (mobile: Account → Privacy)
  3. Toggle OFF Model training on text and Model training on voice
Opt-out does not stop use for product improvement, ads, safety and security. Microsoft 365 Copilot (work/school accounts) is never used for training.

Perplexity

Applies to Free, Pro, Max · logged-in sessions only
Default: training ON
  1. Click your account name (bottom-left) → All settings
  2. Open Preferences
  3. Scroll to Artificial Intelligence, toggle OFF AI data retention
Guest/logged-out use is not covered. For already-collected data, submit a request via Perplexity's data request form.

Grok (xAI / X)

Two places: X platform sharing and Grok itself
Default: training ON (posts + chats)
  1. On X: Settings → Privacy & safety → Grok
  2. Uncheck data sharing for training
  3. In the Grok app/grok.com: Settings → Data Controls, disable model improvement
By default xAI trains on your public posts AND Grok conversations. EU/EEA users can additionally object under GDPR Art. 21 via privacy@x.ai. Past data requires a deletion request via the xAI privacy portal.

Meta AI

Facebook, Instagram, WhatsApp · weakest controls of the group
Default: training ON, no simple toggle
  1. Open Meta Privacy Center and search Object to your information being used for AI at Meta (availability depends on region; strongest in EU/UK)
  2. Submit the objection form
  3. To delete AI chat data, type /reset-all-ais in a Meta AI chat
Public posts are used regardless in most regions. Since December 2025, AI chat interactions also feed ad targeting, with no opt-out for that use. If privacy matters, do not use Meta AI for anything sensitive.

Le Chat (Mistral)

Applies to Free, Pro, Student plans · EU-based provider
Default: training ON (consumer plans)
  1. Click your profile icon → Settings
  2. Open the privacy/data section
  3. Toggle OFF the model training option (labelled as data usage for model improvement)
Mistral scores well on privacy overall: limited data collection, EU jurisdiction (GDPR applies directly), clean opt-out. Team/Enterprise plans are excluded from training by default.

DeepSeek

China-based provider · data stored in the PRC
Default: training ON, weak controls
  1. Open Settings → Data in the app or web UI
  2. Toggle OFF Improve the model for everyone where available
  3. Delete chat history separately
Opt-out controls are significantly weaker than Western platforms and data is subject to PRC law. Treat everything you type as potentially retained. Not suitable for confidential or client data.

What an opt-out does not do

It is not retroactive. Data already baked into a trained model stays there. Some providers accept deletion requests for stored copies, none retrain models.

It is not zero retention. Providers still store chats for service operation, abuse detection and legal compliance, typically 30 days to 18 months.

It does not cover human review for safety. Flagged conversations can still be read by trust & safety teams at most providers.

The real rule: for genuinely confidential material, an opt-out toggle is not the control. Use an enterprise plan with contractual no-training terms, the API, or a local model.

Use caseEffortBest toolWhy
Quick factual question, definition, translationLowAny free chatbot; fast models (GPT-5 mini class, Gemini Flash, Claude Haiku)Speed beats depth. The cheapest model that answers correctly wins; don't burn premium quota here.
Current events, fact-checking, sourced researchLowPerplexity; ChatGPT Search; Gemini with groundingYou need citations and fresh data, not model brilliance. Perplexity's source-first layout makes verification fastest.
Drafting emails, posts, summariesLowChatGPT or Claude; Copilot if you live in Outlook/WordClaude tends to produce more natural long-form prose; Copilot wins when the draft must land inside Microsoft 365.
Serious writing: articles, reports, arguments with structureMediumClaude (Sonnet/Opus class); GPT-5 class as second opinionIterating on tone and logic pays off. Use one model to draft, another to critique - disagreement between models exposes weak spots.
Deep research with report outputMediumDeep Research modes (OpenAI, Gemini, Perplexity); Claude researchAgentic research runs 5-30 minutes and returns a cited report. Worth it when the answer drives a decision; overkill for lookups.
Working with long documents, PDFs, contractsMediumClaude or Gemini (largest usable context windows); NotebookLM for a fixed corpusNotebookLM answers only from your uploaded sources - the right constraint for document Q&A. Gemini handles very large files well.
Data analysis, spreadsheets, chartsMediumChatGPT (code interpreter); Claude analysis / Cowork; Gemini in SheetsInsist on tools that run actual code on your data. A model estimating numbers in its head is a liability.
Quick code snippets, debugging, regexLowAny frontier chatbot; inline IDE assistants (GitHub Copilot)Paste error, get fix. Inline completion is faster than chat for small edits.
Real software projects, multi-file codingHighAgentic coding tools: Claude Code, Cursor, Codex-class agentsAgents that read the repo, run tests and iterate outperform copy-paste chat by a wide margin. Setup cost pays back within a day.
Image generation and editingLowGPT image tools; Gemini (Nano Banana class); Midjourney for art directionChatbot-integrated generators win on convenience and text rendering; Midjourney still leads on aesthetic control.
Meetings, email triage, calendar workflowsMediumCopilot (M365) or Gemini (Workspace) - whichever suite you already useIntegration beats model quality here. The assistant inside your mail client sees context no external chatbot has.
Recurring automated workflowsHighAgent platforms (Claude Cowork/agents, custom GPTs), Make/n8n/Zapier with LLM stepsIf you do it weekly, stop prompting manually. One-time build, recurring payoff - the highest ROI category on this list.
Confidential or client dataHighEnterprise plans with no-training contracts; API; local models (Ollama, LM Studio)Consumer opt-out toggles are not a compliance control. Contractual terms or local execution are.

Decision shortcuts

Effort should follow stakes, not curiosity. A decision worth 10'000 CHF deserves a deep-research run and a second model as devil's advocate. A lunch recommendation does not.

Verifiability beats eloquence. When being wrong is costly, prefer the tool that shows sources or runs code over the one that writes most convincingly.

Ecosystem beats benchmarks. For daily work, the model inside your existing tools (Workspace, M365, your IDE) usually delivers more value than a marginally smarter model in a separate tab.

Two models, one hard question. For anything important, ask two different frontier models and compare. Divergence tells you where the uncertainty is.