What Hallucination means
Suppose you ask ChatGPT in which year a particular school in your upazila was founded. The answer comes at once, with a specific year, without any hesitation. You check and find the year is wrong.
AI did not lie, because to lie you have to know the truth. It invented. This inventing is what is called Hallucination.
If you translate the word into Bangla as "maya" or "motibhrom" (illusion, delusion) it gives the wrong idea, because no confusion is happening here. AI is doing exactly what it is meant to do, only the result does not match reality. Once you understand the reason, catching the mistakes becomes easy too.
The root cause: fitting words, not truth
Inside AI there is no system for checking what is true and what is false.
What a Model does is look at your text and work out which word fits best next. Then it puts that word in and does the same sum again, and word after word an answer stands up.
Now notice this. The sentence "founded in 1972" sounds just as fitting as "founded in 1968". In terms of language there is no difference between the two. The difference is in reality, and the Model has no direct contact with reality.
From this comes the most important point. For AI, a correct answer and an invented answer are made in the same process. So there is no way to judge truth from the tone, the confidence or the neatness of an answer. This is the opposite of human habit, because when a person hesitates we can hear it in their voice.
How information sits inside a Model
Going a little deeper makes this clearer.
Inside a Model there is no table or list that says "such and such school, founded 1972". Inside there are only billions and billions of numbers, which were set bit by bit during Training. The information is not stored separately there, it is spread out.
Here is a comparison. Try to recall a book you read long ago. You remember the main story, you remember the characters' natures, but not what was in which chapter or which sentence was on which page. If you force yourself to say it, you will guess.
The Model's situation is close to this, with one difference. You at least feel that you are guessing. The Model does not feel it, because for it "I remember" and "I am making it up" are the same calculation.
For this reason one rule holds almost always. It is reliable on ideas and explanations, not on specific numbers, dates and names. It will explain "how interest works" well, but it will guess "what was the interest rate last year".
Why it is weak at numbers and counting
There is another interesting reason, and knowing it explains many odd mistakes.
AI does not see whole words. It sees Tokens, small pieces of text. A word may be one piece, or two or three. It does not even notice letters.
So on questions like "how many words are in this text" or "how many times does the letter 'r' appear in this word", it makes funny mistakes. It is not counting. It is guessing what the result of counting should look like.
It is the same with big sums. To do 437 times 28 it does not work step by step. It puts in a number that looks fitting as a result. On small numbers it often comes out right, on big ones it does not.
The fix is simple. Do the arithmetic in a calculator or a spreadsheet, and give AI the job of explaining the result. In this split, both sides do what they are good at.
Other reasons mistakes increase
The date Training ended. Every Model is built with information up to a certain time, called the Knowledge cutoff. It does not know events after that, but it does not stop because it does not know. It guesses and says something.
Little information about Bangla and Bangladesh. On the internet there is far less Bangla writing than English, and even less writing on local matters in Bangladesh. Where information is thin, the share of guessing is larger. So on specific questions inside the country the rate of mistakes is noticeably higher.
Vague questions. The foggier the question, the more gaps AI has to fill, and every filling of a gap is a guess.
A weak habit of saying "I do not know". During Training, people rated tidy answers as good, and nobody liked hearing "I do not know". So the Model learned that it is safe to say something. In newer Models this has been fixed a little, but the lean remains.
Calculation and counting. AI is weak at multiplying and dividing big numbers, or at questions like "how many words are in this text", because it does not work out the sum, it guesses what the result of the sum looks like.
Assuming a mistake inside the question. If you ask "what is in section 7 of such and such law", and that law has no section 7, it will often still make up a section. It usually does not challenge what the question takes for granted.
When AI agrees with you
This problem gets less discussion, but it is the most dangerous in daily use.
You got a correct answer. Then you said, "no no, I know this is wrong". Often AI will apologise at once and change the answer, and the new answer will be wrong. It has learned to please you, not to argue.
So when you verify, never hint at which answer you want. Do not say "this is wrong". Say, "where did you get this information from, and how sure are you?" The way you ask changes the quality of the answer.
Try one more test. Ask about something that does not exist at all, such as asking about the seventh chapter of an invented book. If AI neatly describes that chapter, you will know how carefully to take its answers on that subject from now on.
One mistake, from start to finish
Suppose a shopkeeper in Mymensingh wants to know about renewing a trade licence. They wrote:
What do I need to renew a trade licence and how much is the fee?
The answer came at once, very tidy. A list of papers, a description of the steps, and a specific amount of money. No hesitation anywhere.
They took that amount, took the money and went to the office. There they found the figure did not match, because the fee depends on the type and class of the business, and there are differences from place to place too. They had not brought one paper, because it was not on AI's list. Total loss: a whole morning and the fare.
It is worth noticing where the mistake happened. The description of the steps AI gave was roughly right, because that was general knowledge. The mistake was exactly where there was a specific number or a specific paper's name, that is, where local and current information is needed.
The second time they changed the question:
Explain the general process of renewing a trade licence. Do not write any specific fee or amount. Instead, give me a list of questions I should ask when I phone the office.
What came this time was useful, because this time they were given the very job it does well, which is putting things in order. And the specific number came from where it should come from, which is the office.
Another mistake, one that is hard to catch
Suppose a coaching centre teacher in Rajshahi is making a note for class. They wanted to add a reference in support of a point, so they wrote "give me a source for this information".
A source came, looking perfect. A book's name, an author's name, a year of publication, even a page number. They put it in the note.
The problem was caught two weeks later, when a student went to look for the book and could not find it. The author's name was real, the book's name looked familiar, but that author has no book with that name.
This mistake is more dangerous than the first example, because here it takes time for the mistake to be caught, and by then it has spread. A trade licence mistake costs a morning. An invented reference costs trust.
So the rule is plain. Asking for a source and verifying a source are two separate jobs, and AI cannot do the second one for you.
What to add to the question to cut mistakes
Preventing a mistake is cheaper than catching it, and most of that work happens in the question. There are a few lines that, when added, narrow the path for invented facts to get in.
Write the prohibitions. Say not only what you want, but what you do not want.
Before: Write a note on this subject. After: Write a note on this subject. Do not write any specific number, year or statistic. Do not give any invented source. Where you are not sure of a fact, write "needs verifying".
Tie it to a source. When a document or text is in front of it, it has no chance to speak from memory.
Before: Tell me about this law. After: I have pasted the document below. Use only the information inside this text, nothing from outside. What is not here, say "not in the document".
Ask for the places of uncertainty. When the work is done, ask, "list which lines of this text I need to verify". The list will not be perfect, but it gives a start on where to put your hand.
Doing these three habits together does not bring mistakes to zero, but what remains is much easier to spot.
How risky each kind of question is
| Kind of question | Risk of a mistake | What to do |
| Tidying text, summaries, translation | Low | Read it and check it yourself |
| Understanding general ideas, such as how interest works | Low | Check the idea against a second source |
| Famous and old facts | Medium | Verify if in doubt |
| Specific numbers, dates, statistics | High | Do not use without seeing the original source |
| Sections of law, government circulars, fees | Very high | Go to the government website or a professional |
| Medicine, dose, diagnosis | Very high | Do not decide without a doctor |
| Religious quotations and references | Very high | Verify from the original text or a scholar |
| Information about a specific person or organisation | Very high | Verify it yourself, especially before spreading it |
| Today's prices, news, schedules | Very high | Do not trust it unless search is on |
| Mathematical calculation | High | Check it in a calculator |
Notice the pattern in the list. Risk goes up exactly when the answer has a specific number, name or date. On explanations and ideas, AI is far more reliable.
Invented sources and links
The most misleading mistakes come with sources.
If you tell AI "give a source", it will give one, looking exactly like the real thing. A book's name, an author's name, a page number, even a link. The trouble is that it makes these sources the same way. The author's name may be real, the book may be real, but the book does not say that thing. Sometimes you open the link and find the page does not exist.
So the rule is simple. Being given a source does not mean it is verified. If there is a link, open it. If there is a book reference, match the book. An unverified source is worse than no source, because seeing a source makes us feel safe for no reason.
With religious quotations this risk needs to be remembered separately. AI can get the number of a verse or a hadith wrong, and can even invent a quotation that does not exist. Never share these without verifying.
What to do when in doubt, in eight steps
Once you have an answer in hand, following this format catches most mistakes.
- Read the answer and mark the numbers, dates and names. The risk is stored in these three things. The rest, such as explanation or structure, is usually safe.
- Ask yourself: should this information be written about a lot on the internet? If the subject is local, new or small in scale, assume the risk is higher.
- Ask exactly the same question again in a new chat. If the two answers differ in a number or a name, you know it is a guess. There is a reason not to do it in the old chat: the earlier answer is in front of it there, so it will cling to its own words.
- Ask from the opposite side. Take the information you got and ask, "in which document is this information, and what exactly is written there?"
- Ask where it guessed. "Write separately which parts of this answer you are sure of and which you guessed." This sometimes brings out the weak spots, but do not rely on it, because inside AI there is no real gauge for measuring certainty, so even that level is invented text.
- If there is a link, open it. A link being there does not mean a search was done, links can be invented too. Check both that the page really exists and that it says that thing.
- Test it with a subject you know. Ask a question from your own profession whose answer you know for certain. How it does on that tells you how far to trust it on nearby subjects.
- For decisions about money or health, do not stop at AI's answer alone. You need a second source, at least one: a person or a government document.
How much time to give to verification
You do not need to run all eight steps on every answer. That will waste time instead of saving it. Set the level of verification by the cost of a mistake.
| Kind of work | What is lost if wrong | Level of verification |
| Draft of a Facebook post | Nothing, just write it again | Read it once |
| Personal notes or ideas | Time | Glance over it |
| Email sent to a client | Reputation | Check every promise and date |
| Notes or questions given to students | Learning and trust | Match every fact with the book |
| Announcement of shop prices or offers | Money | Match it exactly against your own ledger |
| Government rules or fees | Money and time | Verify at the office or the official site |
| Health or medicine | Body | Leave AI out, go to a doctor |
| Legal documents or contracts | Long-term liability | Do not use without a professional's eyes |
Mistakes users make
Part of the mistake is AI's, and part is our habit.
| Our mistake | What happens | Fix |
| Believing a tidy answer | A confident mistake spreads | Look at the numbers, not the tone |
| Feeling safe on seeing a source | An invented source passes as true | Open the link, match the book |
| Verifying again and again in the same chat | It clings to its own earlier words | Ask again in a new chat |
| Pressing with "this is wrong" | It apologises and moves to a wrong answer | Ask "where did you get it from" |
| Giving it calculation work | Wrong sums come out quietly | Calculation in a spreadsheet, explanation in AI |
| Thinking a newer Model means no mistakes | The habit of verifying loosens | With a better Model, verifying matters more |
| Sharing a screenshot | The mistake spreads in your name | See at least one source before spreading |
When not to ask AI at all
The most effective way to cut mistakes is not catching them but not asking it certain questions. This list is short, but it is the most valuable.
Where the answer is a specific current number. Today's dollar rate, today's market price, government fees, exam dates, train times. There are official sources for these, and that is faster and safer.
Where the cost of a mistake is paid with the body. Medicine doses, a child's fever, problems in pregnancy. You can use it to understand ideas, not to decide.
Where someone has to take responsibility and sign. A legal reply, tax filing, an audit opinion, a land deed. You can take a draft, but no Model will take the blame for a mistake.
Where the quotation itself is the point. Religious verses and hadith, sections of law, an exact quotation of someone's words. Close is not enough here, it has to be exact, and exactness is its weakest spot.
Where the matter is about one particular person. Someone's education, career or an allegation. If invented facts about this spread, the harm falls on that person, and the blame stays on your shoulders.
Where you already have the right source yourself. Your own ledger, your own contract, your organisation's documents. Looking there is faster, and it is the correct answer.
Do newer Models make fewer mistakes
They do, but not zero, and it will not be zero in the near future.
Newer Models say "I do not know" more, warn more, and with search attached, mistakes fall by a good deal. But falling is not the same as not happening.
There is also an apparent opposite. The better a Model gets, the fewer the mistakes but the more believable they are. Earlier a wrong answer jarred the eye, now it does not. So when you use a good Model, do not cut back the habit of verifying, hold on to it.
People who say "the new version no longer makes mistakes" are either not verifying or selling something.
Questions and answers
Does AI lie on purpose? No. It has no intention or purpose. It just puts in fitting words, and a fitting thing to say is not always true.
So is it wrong to use AI at all? It is fine, you just have to pick the task. Where you will catch a mistake at once, such as tidying text or finding ideas, use it without hesitation. Where a mistake will cost money, health or reputation, human verification is a must.
What if the answer says "I am sure"? That sentence is also words made the same way, not a measured certainty. Inside AI there is no reliable gauge for "how sure am I", so do not take a declaration of confidence as proof.
Do image-making AIs make mistakes too? Yes, of a different kind. Text in the picture comes out jumbled, the number of fingers on hands goes wrong, and things that do not exist in reality appear. Mistakes are especially frequent when putting Bangla text in a picture, so match every letter before you publish.
Do mistakes fall if I ask using my own business documents? They fall by a good deal, because then it speaks from the text in front of it, not from memory. Even so, it can take information from the wrong place or mix up two numbers, so match important figures against the document. And remember that before giving confidential documents to an outside service, you need to think about permission and security.
I pointed out a mistake yesterday, so why is it making the same one today? Because the Model does not change when you correct it. That correction worked only inside that chat, because the text was in front of it. In a new chat it does not know you or your correction. Keep the conditions you need again and again in a note and paste them at the start each time.
If search is on, will it stop making mistakes? Mistakes fall, but do not end. The pages it reads may be wrong too, and it can also get mixed up when summarising what it read. So opening a link and looking cannot be dropped.
If I ask two different AIs the same question, is it verified? Partly. If the two answers differ, you know the matter is uncertain, and that is useful to know. But if both give the same answer, that does not make it true, because both learned from the same kind of internet text, so both may have the same mistake. For specific facts you will finally need a human source or a document.
In short
- Hallucination means AI inventing information, and it happens because it does not verify truth, it only puts in fitting words.
- Correct and invented answers are made in the same process, so you cannot judge truth from a confident tone.
- Risk is highest where the answer has a specific number, date, name or source, and mistakes are even more likely on local matters in Bangladesh because the information is thin.
- Sources and links AI gives can be invented too, so a source being there does not mean it is verified, and you must open the link.
- When you verify, do not hint at which answer you want, and do it in a new chat, because in the same chat it clings to its own earlier words.
- Newer Models make fewer mistakes but sound more believable, so decisions about money, law and health always need a second source.