Why it is weaker on Indian questions

Your uncle asked you one simple thing. How much does the scheme pay, and what is the last date to apply. You typed it into an AI app and got four neat lines back, with an amount and a date sitting inside them.

At the block office you found out both were wrong. You were not careless. The tool is weaker on questions like yours, and researchers have measured how much weaker.

Why it knows America better than your block

The model learned from a very large pile of writing. Whatever the world wrote about often, it learned well. Whatever the world wrote about rarely, it only half learned.

Far more of that writing is about Western countries than about Indian districts. Now think about your own block. The exact papers your office asks for. The way your tehsil handles a correction. This year’s change in your board syllabus. Almost none of that is written down anywhere it could read.

So it has thin patterns to work from. It answers anyway, because answering is the habit it was trained into. You met that habit in why AI makes things up.

It also almost never admits the gap. One large study asked free AI assistants more than three thousand questions. Fewer than twenty got any reply that refused to answer. Saying nothing about doubt is the normal setting, not a sign of knowledge.

There is a simple way to sort your questions before you trust an answer. Ask yourself how many people would ever write this down. A famous history date has been written a million times. The pincode of your village, the name of your block officer, one scheme’s exact subsidy this year. Those may have been written once, or never. Rare and local means highest risk.

What the tests actually show

This has been measured with Indian material, not guessed at.

One test set was built from real Indian exam papers. It holds around 80,000 questions across 11 Indian languages. More than 40 models were run on it. The best one got about 74 answers right out of 100.

Its weakest subjects are the ones you need most. Law and governance. Arts and humanities. Science and maths came out better.

A second Indian test set, from 2025, covers farming, law, money and Ayurveda. The best model there averaged about 67 out of 100. Ayurveda was the weakest area, close to 63.

That second test also compared languages. The same models score roughly 8 to 10 marks higher on its English questions than on its Hindi ones.

These scores sit far below what the same models manage on the general English tests that companies like to show off. Read them as a picture from 2024 and 2025. The exact figures will move. The shape has not moved yet.

Ask in English, then ask for Hindi

A February 2026 study took real UPSC, JEE and NEET questions and asked them in both languages. Accuracy dropped clearly in Hindi.

So when the fact matters, type the question in English. Your English does not have to be good. Short and plain is enough. Then ask for the explanation in the language you read most easily.

Answer in English first.
Then explain the same answer in simple Hindi.
Solve step by step and show every step.
My question is:

Now the other half of the rule. Use one language per message.

Writing Hindi in English letters, with English words mixed in, makes the answer worse. A June 2026 study tested this at three levels of mixing. Accuracy fell as the mixing rose, by about 7 marks to about 17 marks out of 100, depending on the model. Even the strongest models lost ground.

So this is the weak way to ask:

mujhe is scheme ka last date aur eligibility batao simple words me

And this is the better way:

Explain the general eligibility rules of this scheme in simple English.
Do not guess any date or any amount.
If you are not sure, say you are not sure.

Mixed typing is fine for chatting and for English practice. For facts, pick one language and stay inside it. Using AI in Hindi shows how to type Devanagari on a cheap phone.

One more warning if you speak instead of typing. Voice input struggles badly with Indian names, village names and long numbers. A 2026 study of form filling by voice found accuracy fell steeply once real recorded speech replaced clean typed text. The accent, the proper names and the mixed-in English words all add up.

So end any spoken message with this line, and then check every letter.

Now repeat back to me exactly what you understood, item by item.

The questions it gets wrong with a straight face

Stop asking it these. It will answer each one in a calm voice, and it will often be wrong.

• The exact amount a scheme pays.
• Who is eligible and who is not.
• The last date to apply.
• The form number.
• The fee.
• The papers your local office will ask you for.
• A change in this year’s syllabus.
• This year’s exam pattern.
• Anything at all about your own application.

Notice what that list has in common. Every item is local, or exact, or tied to a date. And the wrong answer usually arrives with a very precise number inside it. Research on these mistakes found the bluffs are often confident and specific, not vague. A precise figure is not proof that it knows.

Where the real answer lives

Two habits protect you here. First, for a government scheme, check that the web address ends in gov.in. Second, type that address yourself. Do not tap a link the AI wrote for you.

Models invent addresses that look completely real. That is not bad luck. Building something that looks right is exactly how this mistake works. Later in the course, understanding schemes, forms and offices walks through a full office visit.

Remember that the answer box at the top of a normal search is AI too. You did not choose it, and it is not a source. One study published in April 2026 tested thousands of searches and found close to 9 answers in 100 were wrong. Google disputes that test. Either way, scroll past the box and open the real link when the answer matters.

Farming questions have their own version of this gap. A general AI has no live weather, no mandi rates and no sight of your soil card. So its crop advice comes out broad and safe, and it fits nobody’s field in particular. Farming questions it can and cannot answer covers what to do instead.

Three ways it is still worth using

The tool is not useless here. It is only in the wrong job when you ask it to remember your district.

First, give it the document instead of asking from memory. Copy the words of the notice, paste them in, and ask about that text. When a model is handed the source and told to stay inside it, wrong answers drop a long way. This is the biggest single gain in this lesson.

Below is a notice I copied from the scheme website.
Use only this text. Do not add anything from outside it.
Explain it in simple Hindi, in short sentences.
Then list what I have to do, in order.
Notice:

Second, use it to get ready for the visit. It is good at shaping questions even when it does not know your answers.

I want to apply for a widow pension in my block in Uttar Pradesh.
Write 10 short questions in simple Hindi that I should ask at the office.
Cover papers, fees and dates.
Do not tell me the answers.

Third, use it on official English. Paste an official page and ask for simple Hindi. Then check the work. Ask it to turn its own Hindi back into English, and see whether the meaning came through unchanged.

Giving it your own book, notes or photo shows how to hand it a document from a phone.

It is not your English, and here is what comes next

If it got your block wrong, you did not ask badly. The gap is in what it read, and no amount of better wording fills that gap. Expect the wrong answer, plan around it, and stop taking the blame for it.

Keep the split in your head. The general rule can come from the AI. Your own case comes from the office. Add one small habit for anything tied to time. Tell it today’s date and the year in your message, because it does not reliably know either.

When you cannot reach an official page today, at least ask the same question twice, in two fresh chats. If the two answers differ, neither one is knowledge. The two-chat test and three other checks turns that into a two minute habit.

The next lesson takes the most convincing mistake of all. Checking sources, links and quotes shows how a made-up reference is built, and how to catch one in under a minute.