Protecting sensitive information

AI does not do this task for you. It is the work of deciding what your organisation will and will not put into these tools, and writing that decision down before the deadline arrives.

Nothing visible happens when it goes wrong. No error message, no warning. The text simply now exists on a company’s servers, where by default it may be kept and used to train the next version of the model, and the person who pasted it will never know.

Never send this

Do not type, paste, upload or dictate into an online AI tool: the real name of a witness, survivor, informant or activist; a full testimony or interview transcript; case files or legal filings; an address, coordinates, or a description detailed enough to show where a specific person lives or hides; phone numbers, national ID or passport numbers; photographs showing an identifiable face where someone’s safety depends on that face staying unknown; or unpublished budgets and partner lists. Once it is sent you cannot take it back: deleting the conversation does not recall a copy that was already stored, or already used for training. ChatGPT is rated red here for exactly this reason. When the material is sensitive, use a tool that runs on your own machine, such as Whisper on your own computer, or do not use AI for that piece of work at all.

Write one page today

Three lists: what may never go into an AI tool, what may go in only with a named person’s permission, and what is fine. Name the person who decides the middle list, put the page where everyone can find it, and say it out loud in a meeting. Then open the settings of every AI account your team already uses and turn off the option that lets your conversations train the model: in ChatGPT it is under Settings, then Data Controls, then “Improve the model for everyone”, and it is switched on unless you switch it off.