Foundation of AI › Unit 3
Unit 3 of Foundation of AI covers where artificial intelligence has actually landed in society, in seven hours. It has three parts: AI in different fields, which are education, healthcare, commerce and daily life; AI in decision making, covering hiring, loan approval and court decisions; and the advantages and limitations of AI. Every example on this page is Indian first, because an examiner has read the generic ones a hundred times.
The examination pattern for this unit is predictable. You will be asked to describe applications in one or two named sectors, and you will be asked separately for advantages and limitations. The marks are not in listing applications. They are in naming the system, saying which AI technique it uses, and stating what it gets wrong.
1. AI in different fields
Pick a sector to see the Indian examples, the technique underneath, and the honest limitation.
Choose a sector above.
Education
Two applications are named in the syllabus. Personalised learning apps track which questions a student gets wrong and adjust what is served next, so that two students using the same app see different sequences. Indian platforms do this at very large scale, and the government’s DIKSHA platform carries adaptive content for school subjects. Plagiarism checkers compare a submission against a corpus of published work and previously submitted assignments, looking for statistically improbable overlap; newer tools also attempt to detect AI-generated text, with accuracy poor enough that several universities have stopped relying on them.
Beyond those two, AI grades objective assessments instantly, flags students at risk of dropping out from attendance and submission patterns, and translates material into regional languages. The limitation worth naming in an exam: a personalised learning system optimises for the metric it is given, and if that metric is questions answered rather than concepts understood, it will efficiently make students good at the test and no better at the subject.
Healthcare
AI reads medical images, and this is where it performs best. Systems that screen retinal photographs for diabetic retinopathy are deployed in Indian screening programmes, where the number of patients needing screening vastly exceeds the number of ophthalmologists available. Similar systems assist in reading chest X-rays for tuberculosis, a problem of enormous scale in India. Medical chatbots handle triage, appointment booking and medication reminders, and under the Ayushman Bharat Digital Mission the underlying health records are being digitised in a way that makes such systems possible at national scale.
The limitation to name: a diagnostic model trained mostly on data from one population can perform noticeably worse on another, and the consequences of a false negative in medicine are not symmetrical with those of a false positive. This is why every serious deployment keeps a clinician in the loop rather than replacing one.
Commerce
Fraud detection in banking is the strongest Indian example in the entire syllabus. UPI processes billions of transactions a month, and every one is scored for fraud risk in well under a second, using models trained on historical transaction patterns. No human review is possible at that volume, which makes this a case where AI is not an improvement on a manual process but the only way the process can exist at all.
Stock market prediction is the example to handle carefully. Algorithmic trading systems do use machine learning, and they are fast and profitable at exploiting short-lived pricing patterns. What they do not do is predict long-term prices, because market prices already reflect available information and the market is not a stationary system. An answer that says AI can predict share prices is wrong; an answer that distinguishes short-horizon pattern exploitation from long-horizon prediction is a strong answer.
Customer support uses chatbots and voice bots for first-level queries across banking, telecom and e-commerce, increasingly in multiple Indian languages within a single conversation. The realistic measure of success is deflection rate, meaning how many queries never reach a human, and honest figures are far below what vendors claim.
Daily life
Smart home devices respond to voice commands, translation apps convert speech between Indian languages in near real time, and travel assistants plan routes and book tickets. DigiYatra uses facial recognition for boarding at several Indian airports, which is a good daily-life example precisely because it sits on the boundary of Unit 4: convenient, voluntary in principle, and a biometric database in practice.
2. AI and decision making
This is the part of the unit that matters most, because here AI stops recommending and starts deciding about people. The syllabus names three domains: hiring, loan approval and court decisions.
Below is a simplified credit-scoring model of the kind used in automated loan decisions. Move the inputs and watch the decision change. Then change only the applicant’s area and watch it again.
Now try this. Set an applicant who is comfortably approved, then change only the area field from metro to rural. Nothing about the person’s ability to repay has changed. The decision may have.
That is the central problem with automated decisions about people, and it is not a bug in the code. The model was never told about caste, religion or gender. It was given a postcode, which in India correlates with all three. A model trained on past approvals learns which postcodes were approved before, and so it reproduces the pattern of past lending, including its unfairness, while appearing neutral because no protected attribute appears anywhere in it.
| Domain | What AI does | The benefit | The danger |
|---|---|---|---|
| Hiring | Screens CVs, ranks candidates, scores video interviews | Handles volume no recruiter could read; can standardise the first filter | Learns from past hiring, so it reproduces who was hired before. Amazon abandoned a recruiting tool that had learned to downgrade women’s CVs. |
| Loan approval | Credit scoring from transaction history, repayment records and alternative data | Extends credit to people with no formal credit history, which matters greatly in India | Proxy discrimination through postcode and phone data; refusals that nobody can explain to the applicant |
| Court decisions | Risk assessment, case retrieval, translation of judgments | Speeds up research and translation in a system with an enormous case backlog | Risk scores have been shown to carry racial bias in the United States; opacity is incompatible with a right to reasons |
The rule to remember. AI should support a consequential decision about a person, not make it. The phrase examiners are looking for is human in the loop: a person who reviews the recommendation, can overrule it, and is accountable for the outcome.
3. Advantages and limitations of AI
Advantages
- Speed and scale. Scores every UPI transaction in under a second, which no human process could.
- Consistency. Does not tire, and applies the same standard at 4 a.m. as at 10 a.m.
- Pattern discovery. Finds relationships in large data that no analyst would think to look for.
- Availability. Works continuously, which matters for support, screening and monitoring.
- Reach. Extends scarce expertise, such as retinal screening, to places with no specialist.
- Cost per decision. Falls sharply once the system is built.
Limitations
- No common sense. Makes errors a child would not, because it has no model of the world.
- Data dependence. Biased or thin data produces a biased or unreliable system, reliably.
- No explanation. Neural systems cannot say why, which is unacceptable in law and lending.
- Brittleness. Performance falls sharply on inputs unlike the training data.
- No creativity or intent. Recombines patterns; it does not originate goals.
- Confident errors. A language model will state a false fact with no signal of doubt.
If you are asked for both in one question, do not write two lists and stop. The sentence that earns the last marks is the one connecting them: AI’s advantages and its limitations come from the same property. It learns patterns from data rather than reasoning about the world, which is exactly why it is fast, scalable and consistent, and exactly why it is biased, brittle and unable to explain itself.
Model exam answers
Explain any three applications of AI in society with examples (7 marks)
Healthcare. AI systems read medical images to screen for disease. In India, machine learning models screen retinal photographs for diabetic retinopathy and chest X-rays for tuberculosis, extending scarce specialist capacity to places with no ophthalmologist or radiologist. Medical chatbots handle triage and medication reminders. The limitation is that a model trained on one population may perform worse on another, so a clinician remains in the loop.
Banking and commerce. Fraud detection models score every UPI transaction for risk in under a second, using patterns learned from historical transactions. At a volume of billions of transactions a month, no manual review is possible, so AI is not an improvement on the manual process but the only means by which the process can exist.
Education. Personalised learning platforms track which questions a student answers incorrectly and adapt the next content accordingly, so two students see different sequences. Plagiarism detection systems compare submissions against large corpora for improbable overlap. The limitation is that such systems optimise the metric they are given, which may be engagement rather than understanding.
Discuss the role of AI in decision making and its dangers (10 marks)
AI now participates in three kinds of consequential decisions about people. In hiring, systems screen and rank CVs and score interviews, allowing organisations to process volumes no recruiter could read. In loan approval, credit-scoring models assess applicants from transaction and repayment data, which in India has extended formal credit to people with no conventional credit history. In judicial processes, AI assists with case retrieval, translation of judgments and risk assessment.
The dangers are consistent across all three. First, bias inheritance: a model trained on past decisions learns the pattern of those decisions, including their unfairness. Amazon abandoned an experimental recruiting tool after it learned to downgrade CVs containing markers of women’s colleges and activities. Second, proxy discrimination: even when protected attributes are excluded, correlated variables such as postcode reintroduce them, so the system discriminates while appearing neutral. Third, opacity: neural models cannot explain a particular refusal, which is incompatible with an applicant’s right to reasons. Fourth, automation bias: human reviewers tend to defer to a machine recommendation rather than genuinely reviewing it, so a nominal safeguard becomes a rubber stamp.
The accepted mitigation is to use AI to support rather than make consequential decisions, with a human in the loop who can overrule the system and is accountable for the outcome, regular auditing of outcomes across groups, and an explanation available for every adverse decision.
Practice questions
Frequently asked questions
What are the main applications of AI in India?
The largest deployments are fraud detection on UPI payments, disease screening from medical images such as retinal photographs and chest X-rays, personalised learning and translation in education, multilingual customer support, facial recognition at airports through DigiYatra, and route and demand prediction in transport and delivery. What these have in common is scale: each handles a volume that no manual process could.
Can AI predict the stock market?
Not in the way the question usually means. Machine learning systems do exploit short-lived pricing patterns profitably, which is what algorithmic trading is. They cannot reliably predict long-term prices, because prices already reflect available information and market behaviour changes in response to the very systems trying to predict it. In an exam, distinguish short-horizon pattern exploitation from long-horizon prediction; that distinction is the answer.
Why is AI in hiring considered risky?
Because a hiring model learns from past hiring decisions, and therefore reproduces whatever pattern those decisions contained. Amazon abandoned an experimental recruiting tool after it learned to penalise CVs associated with women. Removing gender from the data does not fix this, because other fields correlate with it. The mitigation is auditing outcomes across groups and keeping a human accountable for the final decision.
What does human in the loop mean?
That a person reviews the AI system’s recommendation before it takes effect, can overrule it, and is accountable for the outcome. It is the standard safeguard for consequential decisions such as loans, hiring, medical diagnosis and anything with legal effect. Its known weakness is automation bias: reviewers tend to defer to the machine, so the safeguard needs to be designed against that, not merely declared.
Written by Piyush Wairale, M.Tech, IIT Madras, founder of PiyushAI Edtech. Last reviewed 22 September 2026.
