AI · LESSON 23 / 25 · English
AI Ethics & Bias
Useful for some people is not automatically fair for everyone.
Content reviewed: 22 October 2026 · தமிழில் படிக்க ↗
Understand it in plain language
Responsible AI involves how data is obtained, who benefits, who may be harmed, and how people can question an outcome. Bias can enter through sampling, labels, objectives, interfaces, and deployment conditions. A fairness metric is one signal, not a universal definition of fairness. Examine the task and affected groups with appropriate domain expertise rather than assuming that one numerical fix settles the issue.
How the AI approach works
AI ethics asks who benefits, who bears the cost of errors, and how decisions can be questioned. It applies across the lifecycle, not just to the wording of generated answers. Problem selection, data collection, labels, optimization objectives, deployment settings, and user interfaces all embed choices about people.
Bias can arise from underrepresented data, inconsistent labels, historical decisions, proxy variables, or a product used outside its intended population. Removing a sensitive column does not necessarily remove information correlated with it. An aggregate score can hide systematic failures for a language, dialect, disability, or user group. Investigate with appropriate context rather than assuming that any rate difference proves one specific cause.
Fairness metrics answer different questions: equal selection rates, comparable error rates, or calibrated scores are not interchangeable goals. Some criteria can conflict under particular conditions. Choose criteria with affected stakeholders and domain expertise, document tradeoffs, and retain meaningful review and appeal. A number alone cannot settle whether a system is acceptable.
Data responsibility includes permission, purpose, minimization, retention, and access control. Explain relevant limitations without shifting all responsibility to users. A human review step must provide time, evidence, and authority to correct the outcome; a person who simply clicks approve is not meaningful oversight.
Break the idea into steps
- Document where data came from and whether you have permission to use it.
- Inspect performance across relevant conditions and groups without collecting unnecessary sensitive data.
- Provide clear limitations, human review, and a route to challenge consequential errors.
Worked example: connect the method to the task
A Tamil support bot misunderstands rural spoken phrases more often than formal Tamil. The impact is longer waits and repeated requests for those users. Investigate representative examples with consent, compare error rates by language style, improve coverage, and provide a working human-support path.
Do not “fix” this by collecting unnecessary identity details or forcing every user to write formal Tamil. The design should respond to the actual source of exclusion and track whether the improvement reaches affected users.
Apply the idea: check your reasoning
Think before revealing the answer
Is removing a gender field enough to establish fairness?
Reveal the answer and reasoning
No. Other features, labels, and deployment practices may preserve related disparities. Define the relevant harm and examine data, errors, context, and available remedies.
A common mistake: Removing a sensitive column does not necessarily remove its proxies or the effects of biased labels.
Before you move on
- Explain the main idea without repeating the technical label.
- Identify the approach, its assumptions, and the evidence needed to trust the result.
- Complete the challenge, then compare your reasoning with the answer.
Sources & a deeper next step
The refresh uses primary papers and official documentation. These optional links are in English; you can understand this lesson without reading them first. Product versions and availability can change.