AI principles · Chapter 7 of 8
Privacy, bias and owning the result
What you paste into an AI tool may be kept, so share only what the task needs; bias comes from data and design and falls on some people more than others; say when AI helped, and own what you send.
Goals
- Hide the personal details in a customer's message, including one the detector misses, and copy the result.To do
- Build a line that says what AI did and how you checked it.To do
- Pass the check.To do
Step 1 of 4: The read
What you paste may be kept
When you paste text into an online AI chat assistant, it goes to the company that runs the service. Depending on the service, your plan and your settings, it may be stored, and it may be used to train later models. Before you paste anything private, look in the service's settings and terms for what it does with your chats.
Share only what the task needs
Personal data is anything that points to a person: a name, an email address, a phone number, a home address, an order or ID number. Most tasks do not need it. A summary of a complaint works just as well with “[Name 1]” in place of the customer's name.
Sharing only what a task needs is called data minimisation. Taking each personal detail out before you paste, or swapping it for a placeholder, is called redacting. A tool can find many details for you, but it misses some, so read the text yourself too.
Bias comes from data and from design
Bias here means that a system's outputs lean one way, for or against some groups of people. It can come from skewed data, as chapter 1 showed. A system trained on past hiring decisions can learn to favour the kind of people hired before. It can also come from design, the choices its builders make about what to include, test and leave out. Either way, the harm falls on some groups more than others, often those who were rare in the data or missing from the tests. When an output is about people, ask who it might treat worse.
Say when AI helped, and own the result
A disclosure line says what AI did and how you checked it, such as “AI translated this letter. A colleague who reads both languages checked it.” Some schools, employers and publishers ask for one. It does not hand over the responsibility: what you send or publish is yours, whoever wrote the first draft.
In the visual
A detector looks for personal details in a made-up customer message. It is not an AI but a few fixed rules that run in this page, such as a pattern for email addresses and a short list of first names. Nothing you type leaves the page. You hide what it missed, see why, and copy the result. Then you build a disclosure line.
Step 2 of 4: Try it: Redact, then disclose
Redact, then disclose
Run the detector on a customer's message, hide what it missed and copy the text you would paste. Then build a line that says how AI helped.
Mode:RealReal: the detector and the disclosure line are worked out in this page by fixed rules, not by an AI, and nothing you type leaves it. The customer, the neighbour and their details are made up for this chapter, and the detector's list of names is kept short on purpose to show what a narrow list misses. The page knows where each detail sits, so it can tell you when one is still showing.
Step 3 of 4: Check
Check
Next: Question 2 of 4
Step 4 of 4: What you can do now
What you can do now
- Take names, emails, phone numbers, addresses and ID numbers out of a text before you paste it, and check what a tool missed.
- Explain how bias can come from the data a system learned from and from how it was built, and who it can harm.
- Write a line that says what AI did and how you checked it.
- Treat what you send or publish as yours, whoever wrote the first draft.