Generative AI first steps: What LLMs do well and badly
Pick an option and check it. The explanation shows after each answer. Take as long as you like; nothing is timed and nothing is sent.
15 questions
Question 1 · easy
Which list holds the four kinds of jobs a chat assistant is usually good at?
Drafting, summarising, explaining and translating all shape words, which is what a model learned from a huge amount of text. Exact dates, today's news, long sums and your private details are the other column: jobs it often gets wrong without a check.
Question 2 · easy
Meena teaches Class 6 in Puri. She wants help from a chat assistant with four tasks. Which one is it usually good at?
Explaining a common topic in plain words is shaping words, a strength. Yesterday's absent students are private data it never saw, this morning's result is a recent event, and an exact total of fees is a sum it predicts rather than calculates.
Question 3 · easy
Kiran runs a kirana shop in Rourkela. He asks a chat assistant for four things. Which one is the most likely to come back wrong?
Today's onion price is a recent event: the model learned from text up to some point in time, so it may be missing unless the app searches the web. Writing a closing message, cutting a pasted letter and putting a list into Hindi all shape words, so they are usually good, though he should still read them.
Question 4 · easy
The lesson says an assistant "does well" at drafting. What does does well mean here?
"Does well" means usually good, not always right, so you read every reply before you send or use it. "Always right" is the tempting mistake; and drafting is not limited to one-line replies or to English.
Question 5 · easy
Sameer opens a chat assistant and types: "How much money is left in my savings account?" He gives no other details. What is true?
An assistant does not know your bank balance, your marks or your office files unless you give them. Learning from a huge amount of text does not mean it saw Sameer's account, and it does not look anything up in his bank's records.
Question 6 · easy
Arjun asks a chat assistant for the score of a cricket match played in Cuttack yesterday. Why might the answer be missing or wrong?
A model learned from text up to some point in time, so news after that may be missing unless the app searches the web for you. Thinking it knows everything up to this minute is the common trap: reading a lot is not the same as reading today's news.
Question 7 · medium
Asha is a tailor in Bhubaneswar. She asks a chat assistant to add 18 customer payments and give the month's exact total. The reply shows a neat total in rupees. What should she do?
When an assistant writes a total, it is usually predicted piece by piece like any other text, unless the app hands the sum to a calculator, so a long or exact sum can be wrong. "A computer never makes mistakes" fails because a language model predicts; a sure look or the total written again in words proves nothing. Check with a calculator.
Question 8 · medium
Ravi, a clerk in Angul, pastes a six-page circular on new office timings and asks for five key points. Why is this task usually safe?
Summarising text you give it is a strength, because the facts come from your text, not from the model's memory. The circular need not be in what it learned, a short reply can still slip, and it does not check with anyone, so Ravi still reads the five points.
Question 9 · medium
Nisha wants a notice for her housing society translated from English into Hindi and into Odia. Which advice fits the lesson?
Translating works well between languages the model saw plenty of: English to Hindi usually comes out well, and Odia may need a closer read because it saw less Odia text. "Both will be perfect" forgets that does well still means read it, and an assistant can translate into Odia, just less reliably.
Question 10 · medium
Priya asks a chat assistant for the author and the year of a book on the history of Konark. The reply gives a name and a year, sounding very sure. What should she think?
The model writes from patterns and does not look anything up, so a wrong name or year can sound as smooth as a right one. A sure tone is not a check, and reading so much does not make exact facts its strength. Priya checks at a library or a trusted source.
Question 11 · medium
Meena asks: "Write a short message to parents. The school picnic is on Saturday and the bus leaves at 7 am." Which column does this task go in?
Ask whether the job is shaping words or supplying a fact. Here Meena supplies the facts and the assistant only shapes them into a message, so it is drafting, a strength (she still reads it). A day and a time are only a risk when the model has to supply them, and it does not know her school's plans on its own.
Question 12 · medium
Why is a chat assistant usually good at shaping words but often wrong about exact facts and long sums?
One reason explains both columns: the model learned the patterns of language and writes from them. That makes drafting, summarising, explaining and translating strong, but an exact fact, a recent event or a total is also predicted, not looked up or calculated. Task length or language is not the reason.
Question 13 · hard
Sanjay, a college student in Jajpur, lists six tasks for a chat assistant: - 1. Turn his points into a leave application - 2. The exact cut-off marks of an exam, announced this week - 3. Explain what a fixed deposit is, in simple words - 4. The total of 35 expenses, to the exact rupee - 5. Put a hostel notice from English into Hindi - 6. His college roll number
How many go in the gets wrong column?
Tasks 1, 3 and 5 are drafting, explaining and translating, so they are usually good. Task 2 is a recent exact fact, task 4 is a long sum, and task 6 is private data it never saw: 3 tasks. "2 tasks" leaves out the roll number, as if the assistant could see college records, and "4 tasks" moves the translation across, but translating is a strength.
Question 14 · hard
A shopkeeper in Balangir asks a chat assistant: "What is ₹120 + ₹80?" It says ₹200, which is right. He then asks for the total of 52 bill amounts and decides: "It got the first sum right, so this total is right too." What is the best reply to him?
Short sums often come out right, but the digits of a long total are predicted, not calculated, so they can be wrong while looking sure. One right short sum proves nothing about 52 amounts, and "always wrong" goes too far the other way. Some apps hand sums to a calculator, but you cannot always tell, so check the long total.
Question 15 · hard
Kiran pastes a long notice from the electricity office into a chat assistant. The notice does not say when the new rates start. He asks: "Summarise this in four points, and tell me if the new rates started last week." Which part should he check before he uses it?
Summarising a pasted notice is a strength, since the facts come from his text. The start date is not in that text, so the model would have to supply a recent fact from memory, which it often gets wrong: he checks it at the electricity office or its official page. Pasting the notice gives it only what is in the notice, not everything about the rates.
Your answers stay in this browser tab. Nothing is sent or saved.