AI Writing Tools That Don't Train On Your Work — The 2026 No-Training List

On April 24, 2026, GitHub quietly updated its Copilot terms so that the code you write — with Copilot's help, or just inside an editor where Copilot is enabled — is now used by default to train future models. You have to manually opt out. A month earlier, Atlassian made a similar shift for Jira and Confluence content on its lower tiers. Microsoft already does the same for consumer Copilot. ChatGPT, Claude, and Gemini have all defaulted to training on the consumer tiers for years.

The pattern is consistent. In 2024 and 2025, "does this AI train on my data?" was a question with a footnote answer. In 2026 it has become a question every writer needs to ask before they paste a single sentence of a manuscript into any tool — because the default has flipped, and the silent change is happening across every major platform you might be using.

This guide is the audit nobody has bothered to publish in one place. Which tools train by default. Which don't. How to opt out where opt-out exists. And — the part that matters for serious writers — which tools are actually incapable of training on your text because they never see it in the first place.

Tier 1: Tools That Train On Your Work By Default

These are the tools where, if you do nothing, your writing becomes training data for future models. All of them offer an opt-out — but the opt-out is buried in settings, sometimes per-device, and the providers reserve the right to change the policy at any time. The honest answer for any unpublished work is: assume it's being read until you've verified otherwise.

• ChatGPT (free, Plus, and Pro) — trains on chats by default. Opt out: Settings → Data Controls → toggle off "Improve the model for everyone." On free tier this disables chat history; on Plus/Pro it doesn't.
• Claude.ai (consumer, free and Pro) — trains by default since 2025. Opt out: Settings → Privacy → toggle off "Help improve Claude." Note: data already submitted before the opt-out can still be retained for up to 5 years per Anthropic's terms.
• Gemini (free, via google.com or the Gemini app) — trains by default. Opt out: Activity → Gemini Apps Activity → turn off. But: this also disables almost all features that make Gemini useful, including chat history.
• GitHub Copilot (Individual plan, post-April 2026) — trains on code by default since the April 24, 2026 policy change. Opt out: GitHub Settings → Copilot → toggle off "Allow GitHub to use my code snippets for product improvements."
• Microsoft Copilot (consumer, via Word, Outlook, Bing) — trains by default. Opt out is per-tenant and per-product; the path is genuinely confusing and Microsoft has reserved the right to change it.
• Notion AI (free and Plus tiers) — trains by default since 2024. Workspace owners must opt out at the workspace level; individual users cannot opt out independently.
• Grammarly (free tier with AI) — uses content to improve products by default. Opt out is per-account, in the privacy settings page.

Tier 2: Tools That Don't Train On Your Work By Default

These are the tools where the provider has made a contractual commitment not to train on your data — usually because they're sold to businesses with privacy as a key purchase criterion. The text still goes to the provider's servers (so a breach could expose it), but it doesn't feed the model.

• OpenAI API — no training on API data by default since March 2023. Documented in OpenAI's enterprise privacy commitments. This is the most reliable way to use GPT-4 / GPT-5 privately.
• Anthropic API — same policy. Data sent via the API is not used for training. The Anthropic Console for developers is also covered.
• ChatGPT Team and Enterprise — paid business tiers with contractual no-training guarantees. Significantly more expensive than Plus, but the only consumer-style interface OpenAI sells with this guarantee.
• Claude for Work / Enterprise — Anthropic's business tier. Same contract as the API in terms of no training.
• Gemini via Google Cloud Vertex AI — contractually guaranteed not to train on your data. Different product from consumer Gemini; sold to enterprise developers.
• GitHub Copilot Business and Enterprise — stricter than the Individual plan. Code is not used for training, and audit logging is available.
• Cursor (Pro tier with zero data retention enabled) — the AI-native code editor offers a true zero-data-retention mode in its Pro tier; when enabled, your code is processed but not stored or used for training.

Tier 3: Tools That Can't Train On Your Work (Because They Never See It)

The strongest privacy tier isn't "the provider promises not to train" — it's "the provider technically cannot train, because they never receive your plaintext in the first place." Two architectures get you here.

Local AI Models (Run on Your Machine)

Tools like Ollama, LM Studio, and GPT4All let you download and run open-source models (Llama 3, Mistral, DeepSeek, Qwen) entirely on your own computer. The model runs locally, your prompts never leave your machine, no provider ever sees your text. The trade-off: you need decent hardware (16 GB+ RAM, ideally an M-series Mac or a GPU), the responses are slower than cloud models, and smaller local models aren't as capable as GPT-4 or Claude.

For most prose drafting, however, a local 8B or 13B parameter model is more than enough — and the privacy guarantee is absolute.

Zero-Knowledge Encrypted Writing Apps

Tools like CipherWrite and Standard Notes encrypt your text on your device before it gets sent to the cloud for sync. The provider stores ciphertext only — to them, your manuscript is unreadable random data. AI training requires plaintext, so it's mathematically impossible for the provider to train on your work even if they wanted to. We covered the underlying mechanism in detail in our guide to zero-knowledge encryption for writers.

The Privacy Tier Comparison, At a Glance

The Recommended Stack for a Privacy-First Writer

If you want one assembled answer instead of a decision tree, this is the stack I'd recommend in 2026 for a novelist who cares about privacy without giving up AI assistance entirely:

1. Draft in a zero-knowledge editor. CipherWrite or Standard Notes. Your raw manuscript never sits in plaintext on a third-party server.
2. Use AI assistance through an API, not a chat interface. If you need brainstorming, prose feedback, or research help, use the OpenAI or Anthropic API (via a privacy-respecting wrapper) rather than ChatGPT.com or Claude.ai. The contractual no-training default makes a real difference.
3. For routine prose tasks, prefer local. Ollama with a recent 8B–13B model handles brainstorming, summarization, and basic prose suggestions privately. The quality is good enough for most non-final work.
4. Audit every tool quarterly. Privacy defaults change quietly. Bookmark the privacy settings page of every AI tool you use and revisit them every three months.

For the deeper context on why this matters specifically for unpublished work, see our explainers on whether OpenAI is reading your novel, the AI writing tools privacy audit, and how shadow AI in your existing tools may already be leaking your drafts.

The Bottom Line

The 2026 default for AI writing tools is training-on. The only question is whether you've opted out, whether you're paying for a tier with a contractual no-training guarantee, or whether you're running infrastructure that can't train on your work because it never sees it.

For published or shareable work, Tier 1 with the opt-out toggled is probably fine. For an unpublished novel, a personal journal, or any draft you wouldn't want a future model to be able to regenerate, move at least to Tier 2 — and ideally Tier 3. The cost is small. The peace of mind is permanent.