AI Writing Assistants: Which One Respects Your Privacy?

You’re three chapters into your novel. You paste a chunk into ChatGPT to check the pacing. Ten minutes later, you’ve forgotten about it.

What you probably haven’t thought about: you just handed your unpublished writing to a company whose terms of service include an irrevocable, worldwide licence to use your content.

What “using your data” actually means

There are three distinct things a company might do with text you submit. Processing for inference: your text gets handled, you get a response, then it’s discarded. Storage without training: saved to your account history, but not fed into the AI models. And training: your words ground into the dataset that teaches the AI. Your prose, your style, your unpublished ideas, pulped into the model’s knowledge.

Most tools don’t commit to one level. They shuffle between all three depending on whether you’re paying, which settings you’ve toggled, and which feature you’re using.

ChatGPT

The free tier treats your conversations as training data by default. That character backstory you spent an evening developing: training data. OpenAI’s terms grant an irrevocable, worldwide licence to use your content. You can delete your account; that doesn’t scrub your words from future model versions.

Paid tiers offer an opt-out: conversations won’t be used for training unless you explicitly allow it. OpenAI still stores data for 30 days for abuse monitoring. And the delete button in ChatGPT removes conversations from your view—server-side deletion happens later, on their schedule.

Claude

Free tier conversations aren’t fed into Claude’s training, but they’re used for trust and safety purposes. Paid tiers get stronger protections: no training use at all, stored 90 days for safety monitoring, then deleted.

Claude’s privacy policy is unusually explicit. No selling your data, no third-party marketing, no user profiles. That clarity is rarer than it should be.

Grammarly

This is where it gets invasive.

Grammarly doesn’t just analyse text you deliberately paste in. Install the browser extension or desktop app and it watches every keystroke in every text field: your emails, your Slack messages, your manuscript in Google Docs. Everything.

They say they don’t use your content to train generative AI models. But they do use it to “improve their products and services,” which includes their suggestions engine. Technically distinct from AI training; a distinction that looks better on paper than it performs in practice.

Local tools: the clean alternative

Cloud services aren’t your only option.

Ollama is free software that runs AI models on your own machine. Download a model, run it entirely offline. Nothing leaves your computer: no account, no login, no company watching. Local models aren’t as capable as GPT-4 or Claude, but that gap is closing fast. For brainstorming, outlining, and catching plot holes, they’re good enough.

GPT4All is a friendlier version of the same idea: a desktop app that downloads and runs models locally, no terminal required.

What to ask before committing

Is your content used for training by default, not just “can you opt out”, but by default? How long is your data retained? What happens if the company gets acquired and the privacy policy changes? Can you actually delete your contribution to their training sets, not just your account?

And: do you even need this? The best privacy protection is not sending your work anywhere at all.

The practical reality

Most writers end up using a mix. ChatGPT for brainstorming because it’s fast. Local tools for manuscript work because you want control. Grammarly for emails, switched off for fiction. That’s reasonable.

The point is to make that trade deliberately, not by accident.

Your first draft isn’t training data unless you decide it is. The default settings want you to share everything. The defaults aren’t mandatory.