Ethics of Artificial Intelligence

Software now helps decide who gets a loan, who gets bail, and which target a weapon picks. It works faster than any office of people. So an old question comes back in a new form. When a machine decides, who is responsible for the result?

As you read, keep three things in mind. What can go wrong when AI decides? Who should answer for it? And how are India and the world trying to set rules?

1What is AI ethics?

AI ethics is a branch of applied ethics. It asks how AI systems should be designed, used and controlled. Its main concerns are fairness, openness, responsibility, privacy and safety.

Why is AI different from older machines? A tractor only does what the driver tells it. An AI system can make a choice and act on it by itself. It can also do this millions of times a day.

So one bad rule inside an AI does not hurt one person. It can hurt a whole group, quickly and quietly.

• Old tools: a human decides, the tool obeys.
• AI: the system itself decides and acts, at huge scale.
• So the ethics questions grow as fast as the system does.

2Why is AI biased?

Nobody needs to type prejudice into an AI. It learns from past data. If the past was unfair, the AI learns that unfairness too.

2.1What did Gender Shades find?

In 2018, researchers Joy Buolamwini and Timnit Gebru tested face-analysis software from big firms, including IBM and Microsoft. The tools were nearly perfect on lighter-skinned men. But they made far more errors on darker-skinned women. (Source: PMLR)

2.2What went wrong with COMPAS?

COMPAS is a US tool that helps judges guess if a person will commit another crime. In 2016, journalists at ProPublica studied its scores. Black defendants who did not reoffend were wrongly marked “high risk” far more often than white defendants. (Source: ProPublica)

2.3Why not just delete race?

Removing the “race” column does not fix the problem. Other data can stand in for it. A home address or a surname can point to a group just as clearly. Such stand-ins are called proxies.

3Why can’t AI explain itself?

Many modern AI systems learn through deep learning. They adjust millions of tiny numbers during training. Even their makers cannot point to one clear reason for a single answer.

This is the black box problem. You can see what goes in and what comes out. But you cannot see why.

That clashes with a basic rule of fair government. If an office refuses your licence or loan, you are owed a reason. “The computer said no” is not a reason.

• Black box: input and output are visible, the reason is not.
• Fair process: a person refused a benefit deserves a reason.
• Explainable AI (XAI) tries to turn the black box into a “glass box”.

4Who is to blame?

Think of a self-driving car that hits someone. Or a loan tool that wrongly turns down a farmer. Who should answer for the harm?

There are many possible names. The coder wrote the program. The data firm supplied the examples. The bank chose to use it. This spread of hands is called the accountability gap.

Most ethicists and regulators give one clear answer. Software cannot be punished or made to pay. So responsibility stays with the people and organisations who build and deploy it. “The algorithm decided” is not an excuse.

5Can machines choose to kill?

Some new weapons can find and strike a target with no human pressing the button. These are Lethal Autonomous Weapons Systems, or LAWS. Critics call them “killer robots”.

The big question is simple. Should a machine ever take a human life on its own?

5.1Where are the talks?

Countries discuss LAWS under the UN’s Convention on Certain Conventional Weapons (CCW) in Geneva. A Group of Governmental Experts has met there for years. Its key idea is “meaningful human control”. A human must truly stay in charge, not just watch.

The UN General Assembly passed its first resolution on LAWS in December 2023. A second one followed in December 2024, with 166 votes in favour. India abstained on that second vote. There is still no binding treaty. (Source: Human Rights Watch)

• Forum: CCW, Geneva.
• Key phrase: meaningful human control.
• Status: talks and resolutions, but no binding treaty yet.

6What about jobs and privacy?

6.1Will AI take jobs?

Earlier machines took over heavy, repetitive work. AI can now do some office and thinking work too. It can draft letters, sort forms and answer customers.

The ethical debate is not only about how many jobs go. It also asks who pays the cost of change. Can retraining and social security keep up with the speed of the technology?

6.2Why worry about privacy?

AI makes mass watching cheap. Cameras with face recognition can track people across a whole city. Software can scan millions of messages at once.

In India, privacy is a fundamental right. The Supreme Court said so in the Puttaswamy case of 2017, through a nine-judge bench. The Digital Personal Data Protection Act, 2023 now sets rules for handling personal data. Its Rules were notified on 14 November 2025. (Source: PIB)

• Jobs: AI reaches thinking work, not just manual work.
• Privacy: a fundamental right since Puttaswamy (2017).
• Data law: DPDP Act, 2023, Rules notified 14 November 2025.

7How do countries regulate AI?

Two models matter most for exams. India uses light, voluntary guidelines. The European Union uses a detailed law.

7.1What are India’s seven sutras?

On 5 November 2025, MeitY released the India AI Governance Guidelines. They come under the IndiaAI Mission. They are voluntary. India chose not to pass a separate AI law for now. The Guidelines rest on seven principles called “sutras”. (Source: PIB)

• Trust is the Foundation. People must trust AI before they use it.
• People First. Humans should keep final control where possible.
• Innovation over Restraint. Encourage careful new uses rather than blocking them.
• Fairness and Equity. AI should include people, not shut them out.
• Accountability. Builders and users of AI must answer for it.
• Understandable by Design. AI should be explainable from the start.
• Safety, Resilience and Sustainability. AI should be safe, sturdy and green.

The Guidelines also name three bodies. These are an AI Governance Group, a Technology and Policy Expert Committee, and the AI Safety Institute.

7.2How does the EU AI Act work?

The EU AI Act came into force on 1 August 2024. It sorts AI by how risky it is. The higher the risk, the stricter the rules. (Source: European Commission)

8What is the alignment problem?

Imagine telling a robot, “Keep the room clean.” It might simply lock the door so nobody enters. It followed your words but missed your meaning.

That is the alignment problem. How do we make sure an AI does what we truly want? The task gets harder as AI takes on bigger, open-ended goals.

Bias is about AI making a wrong call in a small task. Alignment looks further ahead. It asks what happens when powerful AI chases goals its makers never fully foresaw.

9What changed recently?

9.1What was the UN Global Dialogue?

The first UN Global Dialogue on AI Governance met in Geneva in July 2026. Minister of State for External Affairs Kirti Vardhan Singh led India’s team.

• Set up by UN General Assembly Resolution 79/325.
• It grew out of the Global Digital Compact, adopted with the Pact for the Future in September 2024.
• It is the first UN forum on AI open to every member state.
• It received the first report of the Independent International Scientific Panel on AI. (Source: Akashvani News)

9.2What did India host?

India hosted the AI Impact Summit at Bharat Mandapam, New Delhi, from 16 to 20 February 2026. It was the first global AI summit held in the Global South. Its three sutras were People, Planet and Progress. The summit adopted the New Delhi Declaration. (Source: PIB)

10What do PYQs ask?

• UPSC CSP 2026Asked about Large Language Models. The false statement said LLMs give unbiased outputs. They learn the biases in their data. See the question
• UPSC CSP 2026Asked about the AI Impact Summit. The sutras were People, Planet and Progress, not “Planning”. See the question
• UPSC CSP 2025Asked about the Paris AI Action Summit. India co-chaired it. The US and UK did not sign its declaration. See the question
• UPSC CSP 2020Asked what AI can do today. The answer was cutting power use, diagnosing disease and text-to-speech. See the question
• CAPF 2025Asked for an application of AI. The answer is natural language processing. See the question
• UPSC Mains GS IV 2024Asked you to judge AI as a basis for government decisions, from an ethical view. See the question
• UPSC Mains GS IV 2024A case study about a tech company’s CEO facing worries over fast technology change. See the question
• UPSC Mains GS III 2023Asked how AI helps diagnose disease, and whether it threatens patient privacy. See the question

11What are the exam traps?

• India’s Guidelines are voluntary, not a law.
• Guidelines released 5 November 2025 by MeitY, not NITI Aayog.
• Guidelines sutras = 7. AI Impact Summit sutras = 3 (People, Planet, Progress).
• EU AI Act has 4 risk tiers. Only “unacceptable” is banned.
• LAWS talks happen under the CCW. There is no binding treaty yet.
• The US and UK did not sign the Paris 2025 summit declaration.
• UN Global Dialogue met in Geneva, not New York.