Foundation of AI › Unit 4
Unit 4 of Foundation of AI covers the consequences of the technology, and at eight hours it is the largest unit in the syllabus. It has five parts: ethical concerns in AI, covering privacy, bias and fairness, and job displacement; responsible AI, meaning transparency, accountability and human oversight; future trends including self-driving cars, AI in governance and AI in creative industries; and the career perspective, which is about opportunities in non-programming roles.
Examiners mark this unit differently from the others. A generic answer about how AI raises ethical questions scores about four out of ten. A named case, the principle it violated and one workable mitigation scores nine. Everything below is written with that in mind.
1. Privacy issues in AI
AI creates privacy problems that older technology did not, for a specific reason: it makes surveillance cheap. Recording a million faces was always possible. Searching those recordings for one person used to require a million hours of human attention. Now it takes seconds, and that change in cost changes what gets done.
CCTV and facial recognition
Indian cities operate among the densest CCTV networks in the world, and several police forces use facial recognition against those feeds. The technology is genuinely useful for finding missing children and identifying repeat offenders. The difficulty is that it works on everyone in the frame, not only on the person being sought, and nobody in that frame consented or was informed.
Facial recognition also fails unevenly. Accuracy is consistently lower for darker skin tones and for women, which means the cost of its errors is not distributed equally across the people it is used on. A false match in a policing context is not an inconvenience.
Online data
Every app you use produces a trail: what you searched, how long you paused on a post, where your phone was, who you messaged and when. Individually these are trivial. Combined, they support inferences you never disclosed, including health conditions, political leaning and financial distress. This is the real privacy issue with AI, and it is worth stating precisely in an answer: the concern is not that companies hold your data, it is that a model can infer facts about you that you never provided and would not have provided.
The Indian legal position. The Supreme Court held in 2017 that privacy is a fundamental right under the Constitution. The Digital Personal Data Protection Act, 2023 gives statutory form to parts of this, including consent requirements, purpose limitation and rights of correction and erasure. Citing both in an answer on AI and privacy is worth a mark or two, because most scripts cite neither.
2. Bias and fairness
Bias in AI does not mean the system holds an opinion. It means the system performs differently for different groups of people, usually because the data it learned from represented those groups unequally.
Below is a demonstration of exactly how that happens. Change the makeup of the training data and watch what it does to accuracy for each group, and to the headline accuracy figure the vendor would quote.
Group A and Group B here stand for any two populations that appear unequally in a dataset: skin tone, gender, language, urban and rural, or dialect in a speech system.
The point of this demonstration. The headline accuracy is measured on data with the same skew as the training set, so it stays high and hides the gap. Nobody wrote a discriminatory rule. Nobody intended this. The system is simply better at recognising the people it saw more of, and it will be deployed on the strength of a number that was never measured on the people it fails.
Where bias comes from
| Source | What happens | Example |
|---|---|---|
| Historical bias | The data faithfully records a world that was already unfair, and the model learns that pattern as the target | A hiring model trained on a decade of hiring in which few women were selected |
| Representation bias | Some groups appear far less often in the training data | Facial recognition performing worse on darker skin tones; speech systems failing on regional accents |
| Measurement bias | The thing being measured is a poor stand-in for the thing that matters | Using arrests as a proxy for crime, when arrest rates reflect where police patrol |
| Proxy variables | A permitted variable carries a prohibited one | Postcode standing in for caste, religion and income in Indian lending |
| Deployment bias | A system built for one purpose is used for another | A screening tool designed to rank candidates used to reject them automatically |
Bias in Indian contexts, which almost nobody writes about. Indian language models are trained on far more Hindi and English text than Marathi, Odia or Assamese, so quality varies by language and therefore by region. Surnames in Indian datasets carry caste information directly, so a model given names as a feature receives caste as a feature. Urban data is vastly more plentiful than rural. These are stronger examples than any imported one, and examiners notice.
What fairness requires
- Representative data. Check who is in the dataset before training, not after deployment.
- Disaggregated testing. Report accuracy separately for each group. A single headline number is the mechanism by which bias stays hidden.
- Proxy auditing. Test whether a permitted variable is carrying a prohibited one.
- Contestability. Anyone affected must be able to challenge a decision and get a human answer.
- Continuous monitoring. Fairness at launch does not survive a changing population.
3. Job displacement
The honest answer to “will AI take my job” is that AI takes tasks, not jobs, and a job changes when enough of its tasks go. Whether that ends the job depends on whether the remaining tasks are enough to justify the role.
Pick a role to see which of its tasks are exposed, which are not, and what that implies.
Choose a role above.
Three patterns run through all of those, and they make a good structured answer.
- Routine cognitive work is most exposed. Not physical work, which was the previous automation wave, but predictable desk work: data entry, first-draft writing, basic reconciliation, tier-one support.
- Work requiring physical dexterity in unpredictable settings is least exposed. Plumbing, nursing, electrical work, early-years teaching. The robot is much further away than the chatbot.
- Judgement, accountability and relationships resist automation. Not because AI cannot produce an answer, but because someone has to be answerable for it, and a model cannot be.
The historical pattern. Automation has consistently destroyed specific occupations while creating more jobs overall, but the people displaced and the people hired have rarely been the same people. That gap, not the total, is the policy problem. An answer that says AI will create more jobs than it destroys is incomplete unless it names this transition cost.
4. Responsible AI
Responsible AI is the set of practices meant to make AI systems safe, fair and accountable. The syllabus names three principles; there are four worth knowing.
The practical test for whether an organisation is doing this seriously is short: can it explain a refusal, does it publish accuracy by group, and is there a named human who can reverse a decision? If the answer to all three is no, its responsible AI policy is a document rather than a practice.
5. Future trends in AI
Self-driving cars. The technology works well in structured environments and remains difficult in unstructured ones. Indian road conditions, where lane discipline is loose and pedestrians, cattle, cycles and autos share the surface, are among the hardest driving environments in the world for an autonomous system. Expect assisted driving long before autonomous driving here.
AI in governance. Already in use for welfare targeting, tax anomaly detection, land record digitisation and public grievance routing. The gain is speed at scale. The risk is that an error in a government system removes someone’s entitlement rather than their convenience, and the person affected has the least power to contest it.
Creative industries. AI generates images, music, video and text competently enough to displace routine commercial creative work: stock imagery, product descriptions, basic layouts. It has also opened unresolved questions about whose work the models were trained on and who owns the output. The pattern so far is that the commodity end of creative work is being absorbed and the distinctive end is not.
6. AI and you: the career perspective
This section of the syllabus is aimed squarely at students who are not going to become programmers. The premise is correct: the largest employment effect of AI is not in AI jobs, it is in ordinary jobs that now assume AI fluency.
| Role | What AI fluency means in it | What to learn first |
|---|---|---|
| Digital marketing | Generating and testing campaign variants, audience segmentation, analysing performance data | Prompting for content at scale, plus enough analytics to judge whether it worked |
| Content creation | Drafting, repurposing across formats, translation, research assistance | Editing AI output critically. The skill is judgement, not generation. |
| Business analytics | Turning questions into queries, finding patterns, presenting findings | Spreadsheets, a BI tool, and how to tell correlation from cause |
| HR technology | Screening support, workforce analytics, internal knowledge assistants | Fairness auditing and the legal position on automated decisions about people |
| Creative industries | Concept development, iteration, production assistance | The parts of the craft AI does badly, which is where the value moved |
| Operations and customer experience | Support automation, demand forecasting, process redesign | Process mapping and knowing which decisions must stay human |
The one career sentence worth remembering. You are unlikely to be replaced by AI. You are quite likely to be replaced by someone in your own field who uses it well. That is the realistic version of this topic, and it is the version that belongs in your answer.
Model exam answers
Discuss the ethical concerns raised by artificial intelligence (10 marks)
Artificial intelligence raises four principal ethical concerns.
Privacy. AI makes surveillance cheap. Facial recognition applied to CCTV networks identifies everyone in the frame, not only the person sought, without their knowledge or consent, and accuracy is lower for darker skin tones and for women, so the cost of error falls unequally. Separately, models can infer undisclosed facts about a person, such as health or financial distress, from ordinary behavioural data. In India, privacy is a fundamental right under the 2017 Supreme Court judgment, and the Digital Personal Data Protection Act, 2023 gives statutory effect to consent and purpose limitation.
Bias and fairness. A model learns the patterns in its training data, including unfair ones. Representation bias means a system performs worse for groups that appear less often in the data; historical bias means a hiring model trained on past hiring reproduces past exclusion; proxy variables such as postcode reintroduce caste, religion and income even when those fields are excluded. A single headline accuracy figure conceals all of this, which is why accuracy must be reported separately for each group.
Job displacement. AI automates tasks rather than whole jobs, and routine cognitive work is most exposed. Historically automation has created more jobs than it destroyed, but the displaced and the newly hired are rarely the same people, and that transition cost falls on individuals.
Accountability and opacity. Neural systems cannot explain a particular decision, which is incompatible with a person’s right to reasons in lending, hiring or law. The accepted mitigations are transparency about when AI is used, a named accountable party, human oversight designed against automation bias, disaggregated testing before deployment, and a route by which an affected person can contest an outcome.
What is responsible AI? Explain its principles (7 marks)
Responsible AI is the set of design, development and governance practices intended to ensure AI systems are safe, fair, explainable and accountable to the people they affect.
Transparency requires that people know when AI is being used on them and on what basis it decided. Accountability requires a named person or organisation answerable for the system’s behaviour, since “the algorithm decided” is not an answer to a person refused a loan. Human oversight requires that consequential decisions be reviewed by a person who can overrule the system and is accountable for the outcome, designed to resist automation bias, which is the tendency of reviewers to defer to a machine. Fairness and safety require testing across groups before deployment, monitoring afterwards, and a route for affected people to contest a decision. The practical test of whether an organisation applies these is whether it can explain a refusal, publishes accuracy by group, and has a human who can reverse a decision.
Practice questions
Frequently asked questions
What are the main ethical issues in AI?
Privacy and surveillance, bias and unfairness, job displacement, and accountability for automated decisions. A fifth, increasingly examined, is the confident generation of false information by language models. In an answer, name each issue, give one real case, and state one mitigation. Generic discussion scores poorly on this topic.
Why is AI biased if it is only maths?
Because the maths optimises for patterns in the data it is given, and that data records a world that is already unequal. If a group appears rarely in the training data, the system will be worse at that group. If past decisions excluded a group, a model trained on those decisions learns exclusion as the target. No discriminatory rule is written anywhere, which is exactly what makes it difficult to find and to challenge.
Will AI replace human jobs completely?
No. AI automates tasks, and a job changes when enough of its tasks are automated. Routine cognitive work is most exposed; work needing physical dexterity in unpredictable settings, or judgement for which someone must be accountable, is least exposed. The realistic risk to an individual is not being replaced by AI but being replaced by someone in the same field who uses it well.
Which AI careers do not require coding?
Digital marketing, content strategy and editing, business analytics, HR technology and workforce analytics, AI policy and governance, product and operations roles, and customer experience design. Each requires fluency in what AI can and cannot do, judgement about its output, and domain knowledge, rather than the ability to build models.
What is the DPDP Act and why does it matter for AI?
The Digital Personal Data Protection Act, 2023 is India’s data protection law. It matters for AI because AI systems are built on personal data, and the Act sets requirements around consent, limiting use to the stated purpose, and rights of correction and erasure. It follows the 2017 Supreme Court judgment holding privacy to be a fundamental right.
Related reading on this site: AI literacy for professionals, role by role, and the 120-term AI glossary.
Written by Piyush Wairale, M.Tech, IIT Madras, founder of PiyushAI Edtech. Last reviewed 22 September 2026.
