Learn AI /Responsible AI, Human Rights and Safety /Privacy
Privacy
- Starter
- Responsible AI, Human Rights and Safety
- Human rights practitioners
- Policy officers
Also called: data protection, personal data.
Back to the milestone you came from
In five seconds
What it is
Privacy means controlling how information about identifiable people is collected, used, shared and kept, so they are not exposed, tracked or harmed unfairly.
In plain words
Privacy means using information about people only when needed for a clear purpose, without exposing them through careless or unfair data practices.
An everyday example
At a food distribution point, a team needs to estimate tomorrow's meal quantities. One worker proposes writing every visitor's full name, exact home address and family details. Another worker instead counts household sizes by area and keeps only the numbers needed for planning. Both methods can help estimate food, but the second exposes fewer people to unnecessary risk if a paper list is lost or seen by the wrong person.
Real privacy work is more complex than choosing between two paper lists. In practice, risk depends not only on how much data is collected, but also on who can access it, whether data sets can be linked, how long records are kept, what legal basis applies and whether people can be re-identified from details that seem indirect.
Remember thisPrivacy requires control throughout the data lifecycle. Collect the minimum, use it only for the stated purpose, limit access and sharing, and delete it when it is no longer needed.
This term is contested
The broad importance of privacy is widely recognised, but the exact legal definitions, boundaries and enforcement rules vary across jurisdictions and sectors.
Understand it
Understand it
Why it matters
Privacy affects safety, dignity, trust and rights. If too much personal data is collected or used carelessly, people can be exposed, tracked, profiled or treated unfairly. In AI systems, privacy problems can spread quickly because data is easy to copy, combine and analyse at scale. For a non-specialist, the key decision question is simple: do we really need this data, for this purpose, in this form, with these people able to access it? If the answer is unclear, the risk is often higher than it first appears.
How privacy risk grows across a data workflow
After reading this diagram, a learner should be able to explain that privacy depends on the whole data lifecycle, not just on keeping data secure.
01
1. Collect
Data enters the system from forms, devices, messages or records. Main question: is each data item necessary for the stated purpose?
Draw a box labelled 'Collect' with small icons for a form, camera and document feeding in. Add a note below: 'Only what is needed?'.
02
2. Store and combine
Data may be saved, linked with other records or copied to other systems. Main question: who can access it, and can people be identified more easily after linking?
Draw a second box labelled 'Store and combine' with a database icon and two arrows merging. Add a note below: 'Who can access? Can records be linked?'.
03
3. Analyse or model
Software analyses patterns or generates outputs. Main question: does the analysis use personal data in a way people would reasonably expect, and are sensitive uses reviewed by humans?
Draw a third box labelled 'Analyse or model' with a simple chart and gear icon. Add a note below: 'Expected use? Human review for sensitive cases?'.
04
4. Share or act
Results are shown, shared or used to support action. Main question: could the output reveal personal information or unfairly affect a person?
Draw a fourth box labelled 'Share or act' with an arrow to a report and a person icon. Add a note below: 'Could this expose or harm someone?'.
A left-to-right flow diagram with four boxes shows personal data moving through a typical AI workflow. Under each box, a short note names the key privacy question at that stage. Read from collection to outcome to see where unnecessary or risky handling can be reduced.
On hover or tap, each box expands to show one good practice: minimise, limit access, review purpose, and check exposure before sharing.
Describe this without the diagram
The diagram explains privacy as a four-stage flow. First, data is collected, and the question is whether every item is truly needed. Second, data is stored and combined, and the question is who can access it and whether linking records makes identification easier. Third, data is analysed or used in a model, and the question is whether this use matches the stated purpose and whether sensitive uses get human review. Fourth, results are shared or acted on, and the question is whether the output could reveal personal information or unfairly affect someone. The main message is that privacy risk can appear at each stage, not only during storage.
With reduced motion: A static four-box horizontal flow replaces animation. Each box includes its privacy question directly underneath, with no expanding panels.
Privacy is about information linked to a person and whether and how it should be handled. That can include names, phone numbers and ID numbers, but also location, photos, voice, online behaviour or a combination of details that points to one person.
In an AI system, information goes in first. It may be collected from forms, messages, images, sensors or past records. Then the system or the people operating it store it, combine it, analyse it or share it. What comes out might be a prediction, a ranking, a summary or a decision support output. Privacy asks: was this information really needed, was the use clear, was access limited, and could the person be identified or affected?
Privacy is not the same as data security. Data security means protecting information from theft, loss or unauthorised access. Privacy is broader. A system can be secure but still invade privacy if it collects too much, uses data for a new purpose without a proper basis, or keeps it longer than needed.
Privacy is also not the same as confidentiality or consent, though they are related. Confidentiality limits access so only authorised people can see, use or share information for a legitimate purpose. Consent is one possible permission basis in some situations, but not the only one, and it may not be valid if people cannot freely refuse. Good privacy practice usually means collecting the minimum needed, limiting access, explaining the purpose, checking risks and applying extra safeguards to sensitive uses.
How “Counting meals without naming everyone” maps onto the real thing
| In the story | In the AI system |
|---|---|
| writing full names and exact addresses | collecting more personal data than the task requires |
| counting household sizes by area | using less detailed or less identifiable data for the same purpose |
| estimating tomorrow's meals | producing an operational output from input data |
| a lost or exposed paper list | privacy harm when personal data becomes visible or misused |
A worked example
A community-centre coordinator must complete the following week's staffing workplan before close of business and wants to forecast visitor numbers without reusing the sign-in register.
Starting from: Weekly totals by centre, day of the week, public-holiday indicators and a weather forecast. Names, contact details, exact visit times and addresses are excluded.
- The coordinator defines the purpose narrowly: estimating overall demand for staffing and supplies.
- Only centre-level totals and the limited contextual data needed for forecasting are copied into the working file.
- The forecasting tool produces expected ranges by centre and day. The coordinator checks that sparse totals are combined where necessary so an individual cannot be singled out.
- Only the aggregate forecast is entered in the workplan. Access to the working file is restricted, and the file is deleted when no longer needed under the applicable retention requirements.
Result: The coordinator obtains a useful forecast without circulating the sign-in register or keeping an unnecessary copy of personal data.
In a United Nations workplace
At a small field presence with unreliable connectivity, a records officer is finalising the attendance record and note for the file after a community town hall. A participant asks whether names, phone numbers and open-text questions will be entered into an AI summariser. The front office wants an answer before close of business.
Removing names would not settle the issue. Contact details, question text, location data and device metadata may identify participants directly or in combination, while analysis may infer information they never provided explicitly. The records officer needs to confirm the purpose, minimise the fields used, restrict access and account for temporary offline copies, retention and deletion.
Before you try thisNames, contact details, open-text responses and associated metadata must not be pasted into an AI service that has not been approved for that information. Removing obvious identifiers is not enough, since combinations of details may re-identify participants or reveal sensitive information. Use only approved tools and data, with documented access, retention and deletion controls.
Where you may hear it
- A vendor demo during a procurement meeting
- Clearance of terms of reference for an AI tool
- An internal audit interview on system access logs
- An inter-agency working group reviewing a data-sharing arrangement
- A front-office review before a new chatbot goes live
Avoid confusion
Avoid confusion
Often confused with
Data security
Data security means protecting digital or paper information from unauthorised access, change, loss or theft throughout its use, storage and sharing.
Confidentiality
Confidentiality means limiting access to information so only authorised people can see, use or share it for a legitimate purpose.
Consent
A person freely agrees to a specific use of them or their information after getting clear, relevant information and a real chance to say yes or no.
Side-by-side
Common misconceptions
Common misconception: If names are removed, privacy is no longer a concern.
Correction: Removing names helps, but it may not be enough. People can sometimes be identified from location, age, job title, rare traits or by linking one data set with another. Privacy review should consider re-identification risk, not just obvious identifiers.
Common misconception: Privacy only matters for secret or embarrassing information.
Correction: Even ordinary data can become sensitive in context. A travel pattern, attendance record or search history may reveal health status, political activity, religion, relationships or location over time. Harm often comes from combination and inference, not from one dramatic data point.
Common misconception: If a system is secure, privacy is solved.
Correction: Security protects data from unauthorised access, but privacy also asks whether the data should have been collected, combined, analysed or kept in the first place. A well-protected system can still violate privacy if it uses personal data in ways that are unnecessary or unexpected.
Quick check
Which action best supports privacy in an AI project forecasting service demand?
The second option is best because it limits collection and disclosure to what the forecast needs. Deleting excessive data later does not justify collecting it. Removing names only from the output leaves unnecessary identifiable processing. Restricted access helps, but does not justify keeping all source records indefinitely.
Go deeper
Go deeper
The simple version
Privacy means using information about people only when needed for a clear purpose, without exposing them through careless or unfair data practices.
The technical version
Privacy concerns the lawful and appropriate handling of personal data across its lifecycle: collection, storage, use, combination, sharing, retention and deletion. In AI contexts, privacy risk can arise from direct identifiers such as names, from indirect identifiers such as location and device data, and from inference, where analysis reveals something about a person that was not explicitly provided. Key mechanisms for reducing risk include data minimisation, purpose limitation, access controls, retention limits, de-identification with awareness of re-identification risk, and additional controls for sensitive processing. The term is often confused with security, confidentiality and consent, but they address different parts of the problem: security protects against unauthorised access, confidentiality limits access and use to authorised people for a legitimate purpose, and consent is only one possible basis for processing in some contexts.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
The UNESCO recommendation gives an official high-level framework for how privacy fits into AI ethics and governance, which suits policy and organisational readers.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
The OHCHR B-Tech Project provides human rights-oriented material on technology risks and responsibilities, which suits readers focused on rights impacts and accountability.
Intermediatelink checked 2026-07-31
EncyclopaediaPrivacy (opens in a new tab)Wikipedia
Wikipedia gives a broad introductory overview of privacy concepts and background, which suits general readers starting from scratch.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Privacy (opens in a new tab)YouTube search
The YouTube search results page can help readers find different video explanations of privacy, which suits people who prefer to compare multiple introductory formats.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Privacy (opens in a new tab)Reddit search
The Reddit search results page shows community discussions and questions about privacy, which suits readers who want to see common confusions and public debate.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Privacy requires control throughout the data lifecycle. Collect the minimum, use it only for the stated purpose, limit access and sharing, and delete it when it is no longer needed.