Introduction
If you have spent any time in author circles recently, you have probably heard the warning: do not put your manuscript into AI because it will steal your work. It is a concern that comes up constantly, and it is not entirely wrong — but it is not entirely right either. The truth sits somewhere in the middle, and understanding where that middle is matters if you want to use AI effectively without putting your creative work at risk.
This post breaks down exactly what happens to your writing when you use the major AI tools, what the real risks are, and what practical steps you can take to protect yourself while still benefiting from the technology.
The Fear vs the Reality
The fear is straightforward: you paste your manuscript into ChatGPT, Claude or Gemini, and it gets absorbed into the model, regurgitated to other users, or used to train the next version of the AI. Your unique plot twists, your carefully crafted prose, your unpublished ideas — all fed into a machine that anyone can access.
The reality is more complicated. Whether your data gets used for training depends on three things: which tool you are using, which plan you are on, and whether you have checked your settings. Most authors have not done that third one.
How Each Major AI Tool Handles Your Data
ChatGPT (OpenAI)
OpenAI’s approach differs significantly between consumer and business users. On the free and Plus plans, your conversations can be used for model training by default. However, you can switch this off. Go to Settings, then Data Controls, and toggle off “Improve model for everyone.” Once you do that, your future conversations will not be used for training.
For an extra layer of protection, you can enable Temporary Chat mode for sensitive sessions. This prevents conversations from being saved to your history and blocks them from being used for training or the Memory feature.
ChatGPT Enterprise, Team and API accounts have training disabled by default, backed by contractual data processing agreements. If you are using AI seriously for your author business, the paid tiers offer meaningfully better privacy.
Claude (Anthropic)
Anthropic recently updated their consumer terms in a way that every author should understand. Users on Free, Pro and Max plans are now asked to choose whether their data can be used to improve Claude’s models.
If you opt in, your conversations can be retained for up to five years for training and safety purposes — a significant jump from the previous 30-day policy. If you opt out, the 30-day retention period remains and your data will not be used for training.
One important detail: if you delete a conversation, it will not be used for model training regardless of your setting. So if you do use Claude for manuscript work and you are on a consumer plan, deleting the conversation afterwards adds a layer of protection.
Enterprise, API, Education and Government accounts are excluded from training entirely.
Google Gemini
Google operates a clear two-tier system. On the free consumer tier, your inputs may be reviewed by humans and used to improve Google’s AI models by default. Paid tiers — Workspace and Google Cloud — come with a commitment that your data will not be used to train the base Gemini models.
Google is widely considered the most difficult of the three to fully opt out of on the free tier, as doing so can restrict your account’s functionality.
What This Means for Authors in Practice
Let us be specific about the scenarios that matter most to working writers.
Pasting your full manuscript into a free-tier AI tool with default settings: This is the highest-risk scenario. Your text could potentially be used for training. Avoid this.
Using a paid plan with training opted out to get feedback on individual chapters: Much lower risk. Your text is processed to generate a response but is not retained for training. This is broadly comparable to sending your manuscript to a human beta reader — someone sees it, but it is not being republished.
Using the API to build your own workflows: The lowest risk of all three approaches. API usage across all major providers excludes your data from training by default.
Uploading your manuscript to a tool you have never checked the settings on: This is where most authors actually are, and it is an unnecessary risk. Five minutes in your settings can change everything.
Practical Steps to Protect Your Work
The Bigger Picture: Copyright and AI Training
Beyond your individual tool settings, there is a much larger legal battle playing out. In the UK, independent publishers have sent legal warning letters to major AI companies over the use of books for training without consent. The UK Government’s AI copyright report is shaping policy that could affect every author. In the US, multiple lawsuits are working through the courts.
These are systemic issues that individual settings cannot solve. But staying informed about them helps you make better decisions about which tools to trust and how to use them. We cover these developments in our weekly AI for Authors roundups.
The Bottom Line
AI is not automatically stealing your work. But it can use your work for training if you let it — and most tools default to exactly that on their free tiers. The good news is that every major platform now offers meaningful controls. You just have to use them.
Authors who approach AI with awareness rather than fear — or blind trust — will get the most from these tools while keeping their creative work where it belongs: in their own hands.