In short
A draft pasted into an AI assistant is sent to the company that runs it. Depending on the plan, conversations can be used to train models, and human reviewers may read some of them. A draft can contain a source’s name or unverified information. Before using a tool, a newsroom checks where it stores data and what it does with it.
Pasting a draft into an AI assistant to proofread or shorten it is common practice. The text then goes to the company behind the assistant. How it is processed depends on the company and on the plan.
What the companies say
Information collected on 1 October 2026 from the companies’ official pages:
- ChatGPT: on individual plans (Free, Go, Plus, Pro), content is used by default to improve the models. A setting turns this off. Business, Enterprise and Edu plans are excluded by default.
- Gemini, consumer app: when activity is kept, conversations are used to improve Google’s services, including training. Human reviewers read some conversations. Reviewed conversations can be kept for up to three years.
- Perplexity: on the Free, Pro and Max plans, the “AI Data Retention” setting is on by default. It allows data to be used to improve the models. It can be turned off in the preferences, with no retroactive effect. Data from Enterprise plans is never used for training.
- Notion AI: content is used for training only with explicit consent.
Terms change. They often differ from one plan to another at the same company.
What a draft reveals
A draft often contains what a newsroom protects most: a source’s name, information still being checked, an angle before publication. In France, Article 2 of the Law of 29 July 1881 on the freedom of the press protects the confidentiality of journalists’ sources. The French Charter of Professional Ethics for Journalists asks journalists to keep professional secrecy and to protect their sources. Once an item is handed to a third-party service, it is outside the newsroom’s control.
Questions to ask about a tool
Before using an AI tool, a newsroom can check five points:
- Where is the data stored and processed?
- Is content used to train models, by default or with consent?
- How long is it kept?
- Can human reviewers read it?
- Can the data be exported and deleted?
At Verbatim
Content given to Verbatim is hosted and processed in Europe. It is never used to train third-party AI. A specialist cybersecurity firm oversees Verbatim’s security. A journalist’s style profile belongs to their account and only serves their agents. A newsroom that leaves Verbatim takes its data with it.
A quote, a document or a person can be marked OFF, so that a statement can be quoted without naming the source.
Questions
Is it risky to use a consumer AI assistant in a newsroom?
The risk depends on the plan and its settings. Hosting and training terms vary from one plan to another at the same company. For a sensitive document, check these terms before use.
What does hosting data in Europe mean?
The data is stored and processed in data centres located in Europe. This limits its transfer outside the European Union, which the GDPR regulates.