Bias and Privacy: Fair Use and Protecting Your Data

When you use AI to produce a text, an analysis or a message, you are the one who signs it, sends it or publishes it. Two risks deserve special attention because they affect other people as well as you: bias, which can make outputs unfair, and the exposure of personal or confidential data. This lesson gives you concrete habits for both.

In the previous lesson you learned to verify facts. Here we deal with two less visible issues: what the model might assume without telling you, and what you might reveal without realizing it.

You remain responsible for the result

The tool assists; it does not carry the responsibility for you. If an AI-drafted email is rude, an AI-generated summary is wrong or an AI-assisted selection treats people unfairly, you are accountable, not the software. Keep that frame in mind as you read the rest of this lesson.

Understanding bias

You saw in chapter 1 that machine learning learns from examples. Language models learn from enormous quantities of human-written text, and that text carries human stereotypes, imbalances and blind spots. The result is bias: systematic tendencies in the output that favor some perspectives or groups over others.

Where bias comes from

• Training data: some languages, regions, cultures and viewpoints are far more represented than others; some groups are described through stereotypes.
• Tuning choices: decisions about what counts as a good or acceptable answer reflect the values and assumptions of the people who made them.
• Your prompt: loaded wording or an assumed viewpoint can steer the answer. "Explain why this policy is bad" invites a one-sided reply.
• The way it is used: even a neutral tool can produce unfair outcomes if used to rank or judge people.

What it can look like

• Stereotyped examples: always picturing a nurse as a woman and an engineer as a man, or associating certain names with certain jobs.
• Cultural defaults: assuming a particular country's holidays, laws, measurements or customs.
• Unequal quality: better results in widely used languages than in less represented ones.
• One-sided framing: presenting a contested issue as if only one view existed.
• Different tone: describing similar situations in more flattering or more negative language depending on the people involved.

How to reduce it in practice

1. Phrase prompts neutrally. Ask for "arguments for and against" rather than a conclusion you have already chosen.
2. Ask for several perspectives. "Present the main viewpoints, including those critical of each other."
3. Specify inclusive requirements. "Use varied examples and gender-neutral wording."
4. Test by swapping. Change a name, a gender, an origin or an age in your prompt and see if the answer changes in ways it should not.
5. Read critically. Ask yourself who might be missing from this answer, or described unfairly.
6. Keep humans in charge of decisions about people. Hiring, grading, evaluation, access to services and health decisions need human judgment, clear criteria and accountability. In the European framework, some such uses are regulated as high-risk, as you will see in the next lesson.

Before: Write a job ad for a salesperson.

After: Write a job ad for a sales role. Use inclusive, gender-neutral
language. Describe skills and results expected rather than personality
traits, avoid requirements that are not essential (such as an age range
or a native language), and keep it under 200 words.

Protecting confidentiality and personal data

What you type into a tool may be sent to external servers, possibly in another country. Depending on the terms, plan and settings, it may be stored, reviewed by people in some circumstances, or used to improve the service. You should assume that it leaves your control unless you have verified otherwise.

What counts as personal data

In the European framework, personal data is any information relating to an identified or identifiable person: a name, an email address, a phone number, a photo, an identification number, a precise location, and also information that becomes identifying when combined (a job title plus an employer plus a city). Some categories, such as health data, political opinions, religious beliefs or biometric data, are sensitive and benefit from stronger protection. We return to the GDPR in the next lesson.

What not to put in a prompt

• Personal data about other people (clients, students, patients, colleagues, family) unless you have a legitimate reason and the right to do so, and the tool is approved for it.
• Sensitive data of any kind, yours or others'.
• Passwords, access keys, security codes, banking information.
• Company secrets, unpublished results, confidential contracts, internal strategy.
• Material under confidentiality obligations (professional secrecy, non-disclosure agreements).

Anonymize before you paste

You can often get the help you need without the identifying details. Replace them with placeholders or generic descriptions:

Original:
"Marie Durand, 42, from Lyon, has been off sick since 3 March and
her manager Paul Girard wants to discuss her return."

Anonymized:
"An employee in their forties has been on sick leave for several
weeks, and the manager wants to prepare a conversation about the
return to work. Suggest a respectful structure for that meeting."

Notice that the second version still lets the model help, and it contains no name, no precise age, no place and no date. Remember that removing a name is not always enough: a unique combination of details can still identify someone. When in doubt, generalize more or do not use the tool.

Check settings and policies

• Look at the privacy and data settings: can you turn off history or the use of your conversations for improving the service?
• Check whether the plan you use has different data rules from a free one.
• Follow your school's or employer's policy on approved tools. Using an unapproved tool with work data can be a serious breach, even with good intentions.
• Be careful with connected features (access to email, calendar, drive, browser). They are powerful, and they widen what the tool can see.

Think about the people behind the data

Privacy is not only about rules: it is about respect. Before pasting a message from a friend, an email from a colleague or a student's work, ask yourself: would this person expect me to share it with an external service? If the answer is "probably not," anonymize it or leave it out.

Your turn: anonymize and de-bias

Exercise 1. Take a short message that contains real information (a work email, for example, or an invented one with names and dates) and rewrite it as an anonymized version that you could safely give to an AI tool, while still getting the help you need.

Exercise 2 (project). Choose a prompt from your notebook that concerns people (a job ad, a feedback message, a profile description). Run it twice, changing a name or gender in the input. Compare the outputs for differences in tone or content, and add a neutrality instruction to the prompt if needed. Add a "Data rule" line to every library entry saying what must never be pasted.

See a solution

Example of an anonymization: "Hi Sophie, following our meeting with Dr. Martin at the Saint-Louis clinic on 12 June about patient Mr. Lambert's file..." becomes "Hi [colleague], following our meeting with [specialist] at [organization] about a [case], draft a short, polite follow-up that asks for the missing documents." Names, institution, date and case details are gone, but the task is intact.

For the neutrality test, differences you might see include different adjectives (for example "assertive" versus "supportive") for the same behavior. Add to the prompt: "Describe behaviors and results, not personality. Use the same standards whatever the person's gender or origin." Data rule examples: "Never paste names, health information or client data; use placeholders such as [client]."