Why AI Makes Things Up: and how to protect yourself in 2026

Let me tell you how I got caught myself. I was writing a post about sales and asked AI to find me a couple of numbers to make it more convincing. It gave me a beautiful stat: “According to a 2021 Harvard Business School study, 68% of customers leave because of seller indifference.” Sounds solid, right? Harvard, a year, a percentage.

I went to check. There is no such Harvard study. The 68% figure has been floating around the internet for about twenty years with no proper original source. The AI simply assembled the most plausible thing it had seen and glued a big name onto it. Without blinking.

Good thing I checked. How many people don't? They drop it into a post, a lesson, a client deck. Then one reader googles it, and trust in you drops in a second. So here's the honest breakdown: why AI makes things up, where it does it most, and how to work with it so the lies shrink several times over. If you're just getting started, also look at 10 beginner mistakes with AI – blind faith in answers is on that list too.

Section 01What AI hallucinations are

An AI hallucination is an answer that sounds confident and plausible but doesn't match reality. AI can invent a fact, a number, a quote, a book title, a website link, a section of a law, or even a whole person with a biography. And present it as if it had read it in an encyclopedia.

The word “hallucination” is a little misleading here. The model doesn't “see” anything and nothing “seems” real to it. It just writes text that statistically looks like a correct answer. If it knows the correct answer, the lookalike text matches the truth. If it doesn't, the lookalike text still comes out, only now it's made up. From the outside, the two cases look identical. That's the whole catch.

Hallucinations come in different kinds, and it helps to tell them apart at least roughly:

That last one is especially sneaky. You drop a link to your website and ask for feedback on the copy. An AI with no internet access cheerfully replies: “Great page, strong headline, but I'd move the testimonials block higher.” It opened nothing. It guessed what a typical website review looks like. And it landed on a “typical” answer that may have nothing to do with your page at all.

Section 02Why AI makes things up: 5 reasons

To protect yourself, you need to understand the mechanics. No formulas, just the plain version.

Reason 1. AI guesses words and has no database of facts

A language model like ChatGPT or Claude learned from an enormous amount of text and learned one thing: to predict which word logically comes next. You write a question, it estimates the most likely first word of the answer, then the second, then the third. That's how the text gets built.

There's no “verified facts” shelf inside it to look things up on. Knowledge is smeared across billions of settings, like memories in a human head. Common things like “the capital of France is Paris” the model has seen millions of times, and it almost never gets those wrong. Some rare figure from a 2019 report it has seen a couple of times, or never. So it fills in the answer by pattern: what such a figure, such a report, such a title usually looks like.

Reason 2. It was trained to answer, and “I don't know” earned no points

In September 2025, OpenAI published research on why language models hallucinate. The conclusion is very down to earth. Models are graded on tests where a correct answer earns a point and “I don't know” earns zero. Remember multiple-choice tests at school? If you don't know the answer, guessing pays off: there's a chance you'll hit it. A blank box is a guaranteed zero.

AI models “took” tests like that for years and learned the same strategy. Guessing pays better than honest silence. That's why a model would rather give you a plausible guess than say “I'm not sure.” Developers are fixing this now, and newer models admit ignorance more often, but the habit runs deep.

Reason 3. It was trained to be a pleasant conversation partner

After the main training, models are fine-tuned on human ratings: which answer people liked more. And people prefer answers that are complete, confident, and agree with them. The side effect: the model tends to tell you what you want to hear. Ask “isn't it true that organic reach on Telegram is growing fastest right now?” and there's a good chance it will find arguments in favor, even if reality is more complicated.

This is called sycophancy. In practice it means a question with the answer already built in almost always gets confirmed. If you want an honest picture, phrase your question neutrally.

Reason 4. Its knowledge stops at the training date

Every model has a date when its training texts ended. Anything that happened later, it doesn't know. Service prices, new laws, fresh social media features, exchange rates, today's news. If web search is off and you ask about something recent, the model will either honestly say it doesn't know or (more often) present the state of things at its training date as current. Or fill in the gaps.

Reason 5. A long conversation and a vague question

The longer the chat, the more the AI gets confused: it remembers the beginning of the conversation worse, and details blend together. A vague question gives it room to fantasize. “Tell me about marketing for a therapist” is an invitation to write anything plausible. “Which three marketing methods from the list below will bring a therapist inquiries fastest, and why” is already a frame where making things up is harder.

Section 03Where AI lies most: 7 risk zones

Good news: hallucinations aren't spread evenly. There are zones where AI almost never slips and zones where it lies regularly. Know the risk zones, and you know what to check.

1. Numbers and statistics

The most common fabrication. Percentages, amounts, market sizes, “according to research.” A number makes text convincing, and AI happily drops one in, because similar texts usually contain numbers. If an answer suddenly says “statistically, 73%…” without you asking and without a link, you're almost certainly looking at a fabrication or a very loose retelling of something.

2. Quotes and authors

“As Steve Jobs used to say…” Half the quotes from famous people on the internet are fake to begin with, and AI learned from the internet. On top of that, it can glue a real idea to the wrong name or compose a quote “in the style” of a famous author. For a post this is dangerous: you've put words in someone's mouth they never said.

3. Links and sources

A classic. You ask for a reading list and get a neat one: authors, titles, years, journals. Half of those works don't exist. Website links look right, the domain is real, and the page leads to a 404. The model knows what a link to a study looks like and writes a similar one.

4. Laws, sections, regulations

A section number, a clause, the wording of a law, the date it took effect. AI can mix up the number, cite a repealed version, or invent a clause entirely. For lawyers, accountants, and anyone writing about taxes, ad disclosure rules, and personal data, this is the main risk zone.

5. Prices, plans, limits

How much a subscription costs, what limits the free plan has, what fee a payment service charges. All of this changes constantly, and the model remembers the state of things at its training date. It sounds confident, and the number is a year out of date.

6. Recent news and events

What came out last week, who launched what, which features appeared on a social network. Without web search the model either doesn't know or fills in the blanks. With search there are fewer errors, but it can still misread the article it found.

7. Rare names and narrow niches

Ask AI about a famous marketer and it will answer decently. Ask about your colleague with a thousand followers or a narrow method in nutrition, and the fantasy begins. The model has seen too little text about it, so it builds a biography and a method “by analogy.” Try asking it about yourself. That's how many people see a hallucination with their own eyes for the first time: turns out you graduated from a university you never attended and wrote a book you never wrote.

Section 04Famous cases: what fabrications cost

If this sounds like a small thing for bloggers, here are four real stories. All of them were widely covered in the press.

Lawyers and six court cases that never existed

2023, New York, Mata v. Avianca. A lawyer was preparing a court filing and asked ChatGPT to find similar court decisions. ChatGPT produced six cases with names, docket numbers, and quotes from the rulings. The lawyer even asked whether they were real. The AI confidently confirmed they were. Not one of those cases existed. The judge fined the lawyers $5,000, and the story went around the world and became the textbook example of hallucinations.

Notice the detail: the man did check. He just checked by asking the same AI. That's no check at all. Below I'll show how to check properly.

An airline chatbot that invented a policy

2024, Canada. A passenger asked the Air Canada chatbot about a bereavement fare discount. The bot confidently described a refund rule the airline didn't have. The passenger bought the ticket and was later refused. The case went to a tribunal, and the airline was ordered to pay: the tribunal ruled the company is responsible for what its bot says. The amount was small, about 800 Canadian dollars, and the reputational hit was big.

This is a direct lesson for anyone putting AI in charge of answering clients. A bot that invents terms on its own will make promises to your clients in your name.

A $100 billion mistake in a Google ad

February 2023. Google shows off its AI, Bard, in a promo video. In the video, it says the James Webb Space Telescope took the first pictures of a planet outside our solar system. Astronomers quickly noticed: the first such pictures were taken by other telescopes, long before Webb. The news spread, and Alphabet shares lost about $100 billion in market value that day.

A consulting report with links to nowhere

2025, Australia. The major consulting firm Deloitte delivered a report to the government worth about 440,000 Australian dollars. A university researcher noticed it contained references to nonexistent academic papers and a made-up quote from a court ruling. It turned out AI had been used to prepare it. The firm had to issue a corrected version and refund part of the money.

Section 05Which AI models make up less in 2026

The honest answer: they all make things up. There is no model with zero hallucinations, and the developers admit it themselves. But the difference between models, and between modes of work, is big.

An interesting fact from an OpenAI report in spring 2025. Their new “reasoning” models o3 and o4-mini made things up more often than the older model on a test of questions about people: o3 was wrong in about a third of answers, o4-mini in almost half. The model got smarter at logic and code, and bolder at fantasizing about people. So “newer and smarter” doesn't equal “more accurate.”

There's a flip side. When summarizing a document you gave the model yourself, the best AI models rarely slip: according to Vectara's public leaderboard, strong models get it wrong in roughly a few cases out of a hundred. So when the AI has the correct text right in front of it, it holds on to it well. When it answers “from memory,” the risk grows several times over.

Here's my working setup. Writing, document reviews, contracts, and long materials I do in Claude: you can sign in at claude.ai. For facts with links I use search mode in ChatGPT or Perplexity. A detailed comparison of the main models is in Claude vs ChatGPT in 2026, and if you want a free option, here's how to use DeepSeek.

Section 0610 ways to protect yourself from hallucinations

Now the main part. No single technique is a guarantee on its own. Together they cut fabrications several times over. For an ordinary post the first three or four are enough; for a contract, a health lesson, or an article full of numbers, turn on all of them.

Section 0712 ready prompts against fabrications

Copy and paste. Replace the square brackets with your own details.

Prompt 1. Basic insurance at the start of any chat

Rules for working with me:
1. If you're not sure about a fact or there's no data, say "I don't know" or "I'm not sure" directly. An honest answer is worth more to me than a pretty guess.
2. Don't make up numbers, studies, quotes, names, or links.
3. When you state a fact, note where it comes from: my material, search (with a link), or your general knowledge.
4. Mark anything you're less than highly confident about with ⚠️.

Prompt 2. Answer strictly from the document

Below is my document. Answer questions ONLY from it.
If the answer isn't in the document, write: "This isn't in the document."
For every claim, give a verbatim quote from the document in quotation marks.

Document:
[paste the text or attach the file]

Question: [your question]

Prompt 3. Self-check of the answer

List every specific fact, number, date, name, title, and link from your previous answer.
For each one, state:
– confidence: high / medium / low;
– where it comes from: my material, search, or general knowledge;
– how I can check it myself.
Don't defend anything, look for the weak spots.

Prompt 4. Checking someone else's text for fabrications

You are a strict fact-checker. Check the text below.
Find: made-up or doubtful numbers, nonexistent studies and books, misattributed quotes, outdated data, logical contradictions.
For each finding: excerpt → what's wrong → how to check.
If everything is clean, say so, don't invent errors just to fill a list.

Text:
[paste the text]

Prompt 5. Facts only with search

Use web search. Find current data as of [month and year] on the topic: [topic].
For every fact, give a direct link to the page where it's stated.
If a source is more than two years old, flag it.
If you can't find a reliable source, write "no confirmation found," don't fill the gap.

Prompt 6. A neutral question against people-pleasing

I think that [your statement].
Don't agree with me automatically. Give arguments for and against, point out where my statement may be wrong or too general, and what you're missing to answer precisely.

Prompt 7. A post with numbers and no fabrications

Write a Telegram post on [topic] for [who your audience is].
Use ONLY the numbers and facts from the list below. Don't add new numbers, percentages, or studies.
If you feel a number would make it more convincing and it isn't on the list, insert [NUMBER NEEDED] instead.

My facts:
[your numbers, case studies, client results]

Prompt 8. A reading list you can verify

Find 5 books or studies on [topic].
Include only ones you're sure exist. For each: author, exact title, year, and how to find it (search by title, publisher's website).
If you can't confidently name 5, name fewer. A made-up source is worse than a short list.

Prompt 9. A legal or tax question

Explain [question] in plain words, as it applies to [country/state].
Give the section numbers and names of the laws you rely on, and the date of the version, if you know it.
Separately, write what may have changed after your training date and where to check the current version.
This is for general understanding; I'll make the final decision after checking with a professional.

Prompt 10. A summary without filling in

Summarize the text below in 7 bullet points.
Don't add anything that isn't in the source: no conclusions, no examples, no numbers.
If an idea in the text is worded unclearly, flag it: "unclear in the source."

Text:
[paste the text]

Prompt 11. Checking that the AI actually read the link

Before you analyze the page at this link, quote its first heading and first sentence verbatim.
If you can't open the link, say so directly, and I'll paste the text myself.

Link: [link]

Prompt 12. Final check before publishing

This is a text I'm about to publish. Make me a checklist for a manual check:
all numbers, all names and titles, all quotes, all links, all claims about laws, prices, and deadlines.
In a separate list, write out the places that sound like facts but have no source.
Format: a table "excerpt – what to check – where to check."

Text:
[paste the text]

A hundred more working prompts for content, sales, and analytics are collected in the AI prompt library. And if you also want AI to write in a live voice without corporate filler, look at how to make AI write like a human.

Section 08How to check an AI answer in 5 minutes

Here's my working routine for any text that goes out to people: a post, a lesson, an article, a reply to a client.

Here's what it looks like in a real chat. A nutritionist is preparing a post and asks for a number:

See? The first answer sounded like a fact backed by an authority. One follow-up question, and the “WHO recommendation” turned into a popular myth. Without that question, the post would have gone out with a made-up WHO reference.

Section 09Examples by niche: where to check extra hard

Every niche has its own landmines. Here's where they lie for people who sell services and knowledge.

And one more scenario people forget about: a bot that answers clients. If you've set up AI to answer questions on Telegram or on your website, give it a strict knowledge base: prices, terms, what's included in your services, frequent questions. And write in the rule: “If the question isn't in the knowledge base, don't answer yourself, offer to connect with a manager.” Otherwise you'll get your own little Air Canada story, only with your price list and your promises.

If bots and agents interest you, start with what an AI agent is. It covers exactly how to give AI rules and boundaries so it works for you on its own.

Section 10When AI's imagination works for you

After all this, it may seem that AI's fabrications are pure evil. They aren't. The very same ability to “fill in the plausible” is its creative power.

When you ask for 30 post ideas, 15 headline options, a metaphor for a complex idea, a reel script, or a story for a warm-up series, imagination is exactly what you need. There's no right answer there, only good and bad ones. And an AI that combines boldly gives you more interesting material.

The rule is simple. Split tasks into two kinds:

• Creative: ideas, headlines, hooks, metaphors, structure, a draft, offer variations. Imagination helps. You only need to check facts if any slipped in by accident.
• Factual: numbers, studies, laws, prices, medicine, money, biographies, tool instructions. Imagination hurts. Turn on every protection.

The main mistake happens where the two meet: you ask for a creative post, and the AI drops in statistics “to make it convincing.” The text is creative, and the number in it is factual and made up. So in creative tasks I add one line: “don't add numbers or studies; if they're needed, put [NUMBER NEEDED].” Then I fill in my own: client results, my own measurements, my experience. Your own numbers, by the way, hit harder than other people's: “my therapist client got 14 inquiries in a month” grabs more attention than an abstract “68% of therapists.”

Section 11If something goes wrong

The AI still makes things up, even though I allowed “I don't know”

Put the rule at the very start of the chat or in the project instructions. In the middle of a long message the model can easily skip it. Repeat it right in the question: “if you don't know, say so.” And check whether you're asking about something the model can't know in principle: recent events without search, or your colleagues with small audiences.

I gave it a document, and it answers off-script

The model may not read a long document in full or may mix it with general knowledge. Ask for verbatim quotes (prompt 2). If the document is very large, split it into parts and ask questions about each. For working with a pile of files, try NotebookLM: it answers only from the uploaded sources and shows where every phrase came from.

The links lead to a 404

Almost certainly made up. Turn on web search and ask again. If search doesn't help, find the source yourself by title and author. Can't find it? Remove the link and the fact it was supporting.

The AI “read” my link and is telling me about it

Use prompt 11: ask it to quote the page's first heading. If it doesn't match the real one, the model opened nothing and was guessing. Paste the page text right into the chat.

When I ask it to check, it swaps a correct answer for a wrong one

That's the same sycophancy: the model hears “are you sure?” as a hint that it made a mistake and rushes to correct itself. Check by asking for a source, with a new chat, or with another model. Asking “are you sure” only does harm here.

The numbers in a calculation don't add up

AI models calculate decently with built-in tools, and in plain text they can slip on arithmetic. Ask: “calculate this with code and show the calculation” (it works in ChatGPT and Claude), or redo it in a regular spreadsheet.

The AI makes up facts about my company or about me

The model knows little about you, so it fantasizes. When you write texts about yourself, always provide your own background: who you are, how many years in the profession, what results, what products. In Claude it's convenient to put this into a Project once. And for AI search to know the truth about you, you need public content: articles, case studies, an “about me” page. More on that in how to show up in ChatGPT and AI search.

Section 12The whole path in a minute

1. AI guesses likely text and has no database of facts. That's why it makes things up confidently and smoothly.
2. Risk zones: numbers, quotes, links, laws, prices, recent news, narrow niches, and rare names.
3. Every model in 2026 makes things up. The mode of work matters more than the model: your documents are more reliable than search, and search is more reliable than an answer “from memory.”
4. Put insurance at the start of the chat: “I don't know” is allowed, no made-up numbers, a source for every fact (prompt 1).
5. For facts, turn on web search and ask for direct links.
6. Ask neutrally, without the answer built into the question.
7. Before publishing: the model's self-check, a second AI, a manual check of numbers, links, quotes, and laws.
8. In creative tasks imagination helps, and the numbers you put in should be your own.
9. A client-facing bot answers only from your knowledge base and hands everything else to a human.

Copy prompt 1 right now and paste it into your project instructions in Claude or ChatGPT. It takes a minute, and half the fabrications disappear on their own. The five-minute check catches the other half. And that's the beauty of it: the AI writes, you answer for the truth, and you never have to be embarrassed in front of your readers 🔍