What is AI ethics?
Artificial intelligence (AI) is software that learns patterns from data and uses them to predict or decide. Ethics is the study of right and wrong. So AI ethics asks: is this AI fair? Is it safe? Does it respect people?
An AI does not have its own values. It copies the patterns in its data and follows the goal its makers gave it. That is why the people who design, train and use AI are responsible for what it does.
Five key principles
- Fairness: treat groups of people equally.
- Accountability: a named person or organisation answers for harm.
- Transparency: people can know that AI is used and why it decided.
- Privacy: personal data is protected.
- Safety: the AI works reliably and cannot easily be misused.
Bias in data and algorithms
Bias means a system leans unfairly towards or against a group. It can enter in three ways:
- Data bias: the training data has too few examples of some group, or it copies past unfair decisions.
- Design bias: the makers choose a goal or a feature that hurts a group (for example, using home postcode as a stand-in for ability).
- Use bias: a tool built for one place is used somewhere very different.
Example: a hiring AI trained on ten years of past hires, mostly men, may learn to rank women lower. Fixes: collect balanced data, test results for each group, remove unfair features and keep checking after launch.
Privacy and the ethics of AI data
AI needs huge amounts of data, and much of it is about people: photos, voices, locations, health and shopping history. Good practice:
- Consent: ask clearly before collecting.
- Data minimisation: collect only what is needed.
- Purpose limit: use data only for the stated purpose.
- Security: store it safely; delete it when no longer needed.
Many countries now have data protection laws, such as the EU's GDPR and India's DPDP Act 2023. Copyright is also an issue: AI trained on artists' or writers' work without permission raises questions of fairness and ownership.
Transparency, accountability and safety
Some AI models are a black box: even their makers cannot easily say why they gave an answer. Explainable AI shows the main reasons for a decision, so people can check it and appeal.
Accountability means a human is responsible. For high-stakes decisions (medical, legal, loans, exams) a person should review the AI's suggestion: this is called human in the loop.
Safety and misuse: AI can make mistakes with confidence ("hallucinations"), can be used to make deepfakes and false news, and can be tricked. Safe AI is tested, watched after release and has limits on harmful uses.
AI and social change
AI helps in medicine (reading scans), farming (crop disease alerts), translation and accessibility (voice for people who cannot see). It also brings problems:
- Jobs: some routine tasks are automated; new jobs appear; people need reskilling.
- Digital divide: those without devices, internet or skills fall further behind.
- Environment: training large models uses a lot of electricity and water.
- Over-trust: people may stop thinking for themselves.
Assessing social impact of an AI project: list who is affected, what could go wrong for each group, how to measure fairness, and how people can complain. Humans and AI work best together: AI handles speed and scale, humans bring judgement, care and values.
Try it: be the fairness tester
In the last 3D step, set the number of orange examples to 2, then 5, then 8. Write down the YES rate for each group. Which mix gives the smallest gap? At home: open any app on your phone and find its privacy settings. Which data does it collect? Would you agree to all of it?
Key formulas and definitions
- Bias: unfair data or design gives unfair results
- Fairness gap = YES rate of group A − YES rate of group B
- Consent: permission given clearly before data is collected
- Explainable AI: AI that shows the reasons for its decision
- Human in the loop: a person reviews AI decisions
- Deepfake: a fake video, image or voice made by AI
Worked examples
1. A face-unlock AI works for 99 of 100 light-skinned users but only 90 of 100 dark-skinned users. What is the problem and one fix?
Step 1: Error rates: 1% vs 10%, so the system is less accurate for one group. Step 2: This is bias, most likely from training data with too few darker faces. Step 3: Fix: add many more varied faces to the training data and test accuracy separately for each group before release.
2. A hiring AI says YES to 60 of 100 men and 30 of 100 women with similar marks. Find the fairness gap.
YES rate for men = 60%. YES rate for women = 30%. Gap = 60 − 30 = 30 percentage points. A large gap with similar marks shows the AI is unfair and must be fixed before use.
3. A school wants to use cameras with AI to mark attendance by face. List two ethical checks.
1. Privacy and consent: tell students and parents, ask permission, store face data securely and delete it later. 2. Fairness and accountability: test that it works for every student, and let a teacher correct mistakes instead of trusting the AI fully.
Common mistakes
- Thinking AI is neutral because it is a machine. It copies patterns, including unfair ones, from its data.
- Believing more data always fixes bias. More of the same unbalanced data keeps the bias.
- Blaming the AI itself. Responsibility lies with the people and companies who build and use it.
- Assuming a confident answer from AI is correct. AI can be confidently wrong, so check important facts.