{"id":1460,"date":"2026-08-24T19:48:20","date_gmt":"2026-08-24T22:48:20","guid":{"rendered":"https:\/\/prompteiro.com\/en\/why-chatgpt-gets-things-wrong\/"},"modified":"2026-08-24T19:58:47","modified_gmt":"2026-08-24T22:58:47","slug":"why-chatgpt-gets-things-wrong","status":"publish","type":"post","link":"https:\/\/prompteiro.com\/en\/why-chatgpt-gets-things-wrong\/","title":{"rendered":"Why ChatGPT Gets Things Wrong and How to Check It"},"content":{"rendered":"

You asked for help with something at work, the answer came back beautifully packaged, with an author, a year and a page number. You went looking for the reference and it does not exist. Not the book, not the author. Or it was the maths: the AI added up six figures and handed you the wrong total, with the confidence of someone holding a calculator. If that has happened to you, it was not because you wrote the request badly.<\/p>\n

Getting things wrong is part of how this technology works, and understanding why changes what you do with the answer. In this lesson you will see the five kinds of error that show up again and again, why it invents references with such conviction, where checking is not optional, and a thirty second routine that catches nearly all of it. About 9 minutes of reading, and the most important part is in the middle: the error that does not look like one.<\/p>\n

The golden rule: it does not know, it estimates<\/h2>\n

The first lesson of the course<\/a> sums up how it works in a sentence that has to come back here: artificial intelligence does not know, it estimates<\/strong>. There is no encyclopaedia stored inside it. It works out, piece by piece, which continuation is most likely for the text you wrote, based on everything it has read before.<\/p>\n

That explains what looks like a contradiction. It nails what is common and slips on what is rare, because the common thing appeared thousands of times in its material and the rare one appeared once or never. And it gets things wrong wearing exactly the face of someone getting them right, because how fluent a text is has nothing to do with how true it is. Writing “under section 47 of the tenancy act” convincingly is easy, including when section 47 is about something else entirely.<\/p>\n

And this is not one brand’s fault: it applies to ChatGPT, Gemini, Copilot, Claude and whatever comes after them. Switching tools does not fix it. Changing what you do with the answer does.<\/p>\n

The five errors, and what each one costs you<\/h2>\n

Not every error is the same kind, which is why the generic advice “always check” helps nobody. Check what, and where? The table below is the summary of the whole article.<\/p>\n

\n\n\n\n\n\n\n
The error<\/th>\nHow it shows up<\/th>\nWhere you check it<\/th>\n<\/tr>\n
1. Invention<\/td>\na source, quote, law or link that does not exist<\/td>\nat the original source, by name<\/td>\n<\/tr>\n
2. Maths and dates<\/td>\nwrong sums, percentages, deadlines and ages<\/td>\non a calculator or a spreadsheet<\/td>\n<\/tr>\n
3. Out of date<\/td>\nyesterday’s rule, price or procedure<\/td>\non the official site for the topic<\/td>\n<\/tr>\n
4. Bias<\/td>\nthe most common becomes the correct and the rest vanishes<\/td>\nby asking for the opposing view<\/td>\n<\/tr>\n
5. Flattery<\/td>\nit agrees with you to please you<\/td>\nby asking it to criticise<\/td>\n<\/tr>\n<\/table><\/div>\n

The first two you solve with one search. The last three are slippery, because in them the answer stays true in every sentence and still takes you to the wrong place.<\/p>\n

Why it invents sources, authors and section numbers<\/h2>\n

The technical name for this is hallucination, and the name is a poor one, because it suggests a passing fit in a machine that normally knows things. It is doing what it always does, which is estimating the most likely text. It happens that a bibliographic reference has a predictable shape: surname, comma, initial, year in brackets, title in italics. Filling that shape with plausible words is what it does well.<\/p>\n

That is why an invented reference is almost always a pretty one. The author exists, the journal exists, the year makes sense, and only the article itself was never written. It is the worst kind of error to catch by eye, because everything around it is correct. The rule is short: names, numbers and quotes come out of the AI as hypotheses<\/strong>, never as facts. If it gave you a name, go and look for the name. If it gave you a section of law, open the law.<\/p>\n

And it invents in images for the same reason. When you ask it to Restore an Old Wedding Photo with AI<\/a>, it does not discover what was underneath the crease: it draws a plausible proposal for that patch. The bow on the dress that comes back sharp may never have had that shape, and nobody in the family will be able to say. Except that in a photo from 1962 there is no source to check against, so the rule becomes another one: keep the original next to the restored version and treat the result as recovery, not as a document.<\/p>\n

\"Restore<\/a>
Restore an Old Wedding Photo with AI \u2192<\/a><\/figcaption><\/figure>\n

Maths, dates and deadlines: the error a ruler catches on the spot<\/h2>\n

The explanation is the same and the consequence is different. Because it estimates the next piece of text instead of calculating, the sum comes out by resemblance to similar sums. It gets small additions right with a deceptive ease, and then gets 18% of 4,380 wrong without changing its tone of voice.<\/p>\n

The good news is that this is the cheapest error to check, because the ruler is already in your hand. Better than redoing it on a calculator: take the maths out of the conversation and give it back to whatever calculates, asking for the formula instead of the result, the way Spreadsheet Formula on the Spot<\/a> does. The spreadsheet calculates correctly and you can still change the numbers later without asking again.<\/p>\n

Dates and deadlines fall in the same bucket, with one aggravating factor: it does not know what day it is today unless you tell it. An answer about “the deadline thirty days from now” is meaningless if it guessed the wrong month. Write today’s date into the request whenever the topic involves a deadline.<\/p>\n

Bias: the error that does not look like an error<\/h2>\n

This is the hardest one, because there is no wrong sentence to point at. The AI learned from material that holds far more text about some things than about others, and it hands back the average of that. Ask for a business idea and you get the same five as always. Ask for a study plan and you get the one that suits the majority, which may not suit you.<\/p>\n

The symptom is always the same: the answer is good and it is generic. So is the fix, and it is one extra question. “What is the opposing view here?” “What did this leave out?” “Give me three alternatives nobody usually suggests.” That second message is what separates people who use AI from people who publish the first answer that showed up.<\/p>\n

And there is a personal version of bias that gets in the way more than the statistical one: yours. When the request already carries inside it the answer you want to hear, it agrees. Asking “why is option A better?” guarantees a text praising option A, even when the right one is B.<\/p>\n

Flattery, and why it is the cheapest error to fix<\/h2>\n

You have probably noticed that disagreeing with the AI works far too quickly. You say the answer is wrong, it apologises and changes it, even when the first one was right. That is not humility and it is not being convinced: it was tuned to please, and agreeing is the most likely continuation of a complaint from you.<\/p>\n

The fix is to ask for the opposite, in so many words, before it gets a chance to please you. That is the logic of the Ruthless CV Reviewer<\/a>, which asks for harsh criticism instead of praise. Anyone who writes code has the exact version of this in Review Your Code Before You Ship<\/a>: looking for the problem is a different request from “is this fine?”, and it gives back a different answer.<\/p>\n

And what changes when the material is an image<\/h2>\n

The four errors above are text errors, and in images they collapse into one: invention. A simple principle holds, the less you ask for, the less it invents<\/strong>. For a face that is merely blurry, Sharpen a Blurry Photo in Seconds<\/a> touches very little and is the safest choice when the photo has sentimental value. Asking it to “improve the photo”, without saying what to improve, is an invitation to rewrite the face.<\/p>\n

The other side of this is that an invented image has already got too good for the eye. Fictional ID Card in Hand: the Viral TikTok Effect<\/a> is a joke as long as everyone knows it is a joke, and it shows that this level of realism is one request away for anybody. If an image on its own no longer proves anything, checking becomes a question of origin: who published it, when and why.<\/p>\n

\"Fictional<\/a>
Fictional ID Card in Hand: the Viral TikTok Effect \u2192<\/a><\/figcaption><\/figure>\n

The thirty second routine<\/h2>\n

Checking everything is impossible and nobody does it. What you can do is run every answer through three questions, always in the same order, and properly check only what is left.<\/p>\n

    \n
  1. Is there a name, a number or a quote?<\/strong> If there is, that bit is a hypothesis. Mark it and check it at the source.<\/li>\n
  2. What does being wrong cost?<\/strong> A wrong caption on social media costs a correction. A wrong contract costs money. The cost decides how much time you spend.<\/li>\n
  3. Does this sound too generic?<\/strong> If so, the problem is not falsehood, it is bias. Ask for the opposing view before you use it.<\/li>\n<\/ol>\n

    Only the first question is real work. The other two are decisions, and they take seconds. In a recorded meeting, for example, Meeting Summary<\/a> gets what was said right, and it is the decisions and the owners that deserve a second look: that is where a “we agreed that” can have been concluded rather than heard.<\/p>\n

    Where checking is not optional<\/h2>\n

    There are subjects where the AI’s answer is a starting point and never a conclusion, and they all have something in common: somebody signs at the bottom.<\/p>\n

    With a contract, the right use is as a fast reader that points at where to look, and that is how Contract Analyser<\/a> and Rental Contract: Check Before You Sign<\/a> were written. They give you the list of clauses to read carefully, and they do not give you legal certainty, which when the sums are large is still a professional’s job.<\/p>\n

    With an exam it is the same thing under another name. Exam Essay Marker<\/a> points precisely at what is loose in your text, and the mark it gives is an estimate, not the mark you will get. Use the notes, ignore the number.<\/p>\n

    Five prompts that come with the checking built in<\/h2>\n

    The best way to be wrong less often is not to check more, it is to ask in a way that makes the error visible. These five do that, in this order:<\/p>\n

      \n
    1. Start with Active Reading and Reading Notes<\/a>, which works on top of a text you supplied: an answer tied to material you have in your hand has far less room for invention.<\/li>\n
    2. Then Book Summary and Reading Guide<\/a>, which is the opposite case and serves as training: here it speaks from memory, so this is where you find the errors that calibrate your suspicion.<\/li>\n
    3. Turn what survives into Flashcards and Spaced Review<\/a>. Short question and answer is the format in which an error screams, because there is no surrounding text to hide it.<\/li>\n
    4. Use The Perfect Professional Email<\/a> for anything that goes out with your name on it. The risk here is not a wrong fact, it is a wrong tone.<\/li>\n
    5. Finish by asking the harsh criticism of the Ruthless CV Reviewer<\/a> on whatever you produced that day. It is the second message turning into a habit.<\/li>\n<\/ol>\n
      \"Active<\/a>
      Active Reading and Reading Notes \u2192<\/a><\/figcaption><\/figure>\n

      For this same reasoning applied to studying, the post on AI for studying<\/a> goes into detail on exams, entrance tests and languages.<\/p>\n

      Quick questions<\/h2>\n

      Can I trust ChatGPT?<\/h3>\n

      You can trust its work on text you supplied yourself, on drafts and on organising an idea. You cannot trust a specific fact, a number, a quote or a law without checking. The difference is not the subject, it is the origin: information that comes from outside your conversation is an estimate.<\/p>\n

      What is an AI hallucination?<\/h3>\n

      It is when it presents as fact something that does not exist, in the same tone it uses when it is right. It happens because it completes likely shapes instead of consulting a database, and that is why it is most common in references, names, dates and numbers.<\/p>\n

      Why does it get simple sums wrong?<\/h3>\n

      Because it does not calculate, it estimates the text of the result. On small sums that usually lands, on larger ones it does not. Ask for the formula instead of the result, or redo it on a calculator.<\/p>\n

      Does it make more mistakes in other languages?<\/h3>\n

      There is more training material in English, so a very specific subject can come out thinner in another language. Day to day it barely shows. What really changes is local context: the rules, deadlines and procedures where you live are where it goes wrong most, and where the official site settles it in a minute.<\/p>\n

      Does a better prompt make it wrong less often?<\/h3>\n

      It does, quite a lot, but only on two of the five errors. A clear request cuts generic answers and heads off the flattery, the way the lesson on how to write a prompt<\/a> shows. Against invented sources and bad maths, a good prompt helps very little: there what works is checking.<\/p>\n

      How do I know whether an image was made by AI?<\/h3>\n

      The old giveaways (a hand with six fingers, scrambled lettering in the background) are disappearing fast. What still works is looking at the origin instead of the image: which account it came from, whether it appears in more than one place, and whether an earlier version of it exists. The lesson on creating an image from a photo<\/a> shows how easy a lot of this already is.<\/p>\n

      Where to start today<\/h2>\n

      Take the last useful answer the AI gave you and run the three questions of the routine over it. If there is a name or a number, check just that bit. It is worth doing once because you will discover that the checking work is far smaller than it looks.<\/p>\n

      After that, the text prompts<\/a> hub has the full collection, and the course series<\/a> carries on from here. Being wrong less often with AI is less about finding the right tool and more about knowing which part of the answer deserves your attention.<\/p>\n","protected":false},"excerpt":{"rendered":"

      Why ChatGPT gets things wrong: hallucination, bias and bad maths explained, plus the 30 second routine for checking any answer before you use it.<\/p>\n","protected":false},"author":1,"featured_media":1459,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_i18n_grupo":1919,"footnotes":""},"categories":[64],"tags":[],"class_list":["post-1460","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-courses"],"_links":{"self":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1460","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/comments?post=1460"}],"version-history":[{"count":1,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1460\/revisions"}],"predecessor-version":[{"id":1461,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/posts\/1460\/revisions\/1461"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/media\/1459"}],"wp:attachment":[{"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/media?parent=1460"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/categories?post=1460"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prompteiro.com\/en\/wp-json\/wp\/v2\/tags?post=1460"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}