Reading time: 12 minutes
ChatGPT makes so many mistakes because it predicts the most likely words, and those are not always correct. With little data on a small company, AI makes more mistakes, because it tends to guess instead of admitting doubt. The mistake can also come from a wrong page on the web.
In short
- According to OpenAI, models make things up because training and evaluations reward guessing more than admitting doubt.
- Without a cited source, wrong answers from ChatGPT tend to come from the model’s memory. With a source, however, they come from the web, which can also be wrong.
- In a Searchable report, brands flagged as false 9.2% of the facts AI engines stated about them. Pricing is where AI makes the most mistakes.
- OpenAI does not document a way to correct company data. So the fix happens at the source, and the timing depends on where the mistake came from.
Why does ChatGPT make so many mistakes and invent facts?
ChatGPT makes so many mistakes because it learned to predict the next word, not to tell fact from fiction. According to OpenAI, the tests that rate models also reward guessing. After all, saying “I don’t know” guarantees zero points, while a guess might be right. So, faced with a fact it doesn’t know, the model tends to take a chance.
In a September 2025 study, OpenAI explains the mechanism. In pretraining, the model learns from huge amounts of text that carry no true or false labels. Spelling follows patterns, so those errors fade with scale. However, arbitrary and rare facts, such as a birthday, follow no pattern, so they turn into hallucinations. See what AI hallucination is, with types and examples involving companies.
In OpenAI’s example, a widely used chatbot gave three different titles for one author’s PhD dissertation. None was correct. Likewise, when asked for his birthday, it gave three dates, all wrong.
Does ChatGPT make fewer mistakes when it admits it doesn’t know?
It does, at least in the test OpenAI uses as an example. On SimpleQA, a test of factual questions, gpt-5-thinking-mini skipped 52% of the questions and got 26% wrong. By contrast, the older o4-mini almost always answered and, as a result, got 75% wrong. Still, both got about the same share right: 22% and 24%. The authors sum it up this way:
“Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.”
Adam Tauman Kalai and coauthors, OpenAI study (2025)
For a company, then, the risk lies in facts that rarely appear in texts. Think of the partners of a small company or a clinic’s opening hours. OpenAI says newer models hallucinate less, but the problem persists in all large language models. Saying that ChatGPT lies, however, misses the point. As OpenAI explains it, ChatGPT guesses, and it can guess with confidence.
When ChatGPT makes mistakes, does the data come from training or the web?
It can come from either, and the answer itself gives a clue. Without search, ChatGPT answers from what it learned in training, which has a cutoff date. With search, however, it reads web pages and can cite sources. So, to tell which case it was, OpenAI advises checking whether the answer cites web sources.
In its Help Center, OpenAI explains that models learn from data up to a certain date. So they don’t know what came later, unless they use tools such as search. The same page warns that ChatGPT can sound very confident even in wrong answers.
In a 2025 post, Hashtag Treinamentos, a Brazilian training company, asked ChatGPT who its partners were. The answer swapped one partner’s name, made up another’s last name and invented a partner who doesn’t exist. Then, asked about its sources, ChatGPT said it had used knowledge acquired up to 2024.
Right after that, the team asked for the company’s CNPJ, Brazil’s business tax ID, and ChatGPT searched the web. With search, it listed five partners and cited a page from UOL, a Brazilian news portal. According to Hashtag, this time the answer was correct. In other words, the mistake came from the model’s memory, and search fixed it.
Can web search get it wrong too?
It can, because web pages can be wrong too. In an experiment by Ahrefs, which sells SEO tools, an analyst created a fictional brand. Then he spread three fake stories about it on other sites. As a result, five of eight AI products repeated the stories, even with an official FAQ on the brand’s site. ChatGPT, on the other hand, relied on the FAQ in most answers.
The guide to what GEO is explains the difference between the model’s memory and web search. The full path of an answer, from training to search, is in the guide on how ChatGPT works.
Where does AI make the most mistakes: price, address, partners or products?
AI makes the most mistakes on facts that change often. In the Searchable report, for example, pricing leads, with 13.5% of facts flagged as false. Products and services come in at 8.6%, history and founding at 6.9% and location at 4.8%. AI also gets partners’ names wrong, as the Hashtag case showed.
Searchable, which sells AI brand monitoring, gathered 32,556 facts that AI engines stated about 1,704 brands, through July 2026. Then an employee at each brand rated each fact as right or wrong. In total, brands flagged 9.2% as false. However, the report has no figure for Brazil.
| Type of fact | Facts flagged as false |
|---|---|
| Pricing | 13.5% |
| Funding and financials | 10.9% |
| Integrations and tech stack | 10.6% |
| Partnerships and clients | 9.9% |
| Products and services | 8.6% |
| History and founding | 6.9% |
| Location and coverage | 4.8% |
The report sorts false facts into three groups: once-true details that went out of date, confusion with another brand and pure fabrication. In addition, companies with up to 50 employees had 9.4% false facts. Large ones, with 250 or more, had 6.3%. Among the most-reviewed engines, ChatGPT had 8.1% false facts, while AI Overviews had 11.4% and Perplexity 11.5%.
How do you check whether ChatGPT makes mistakes about your company?
Start with facts that change often and that matter to buyers. For a clinic, for example, that means opening hours, insurance plans accepted and address. For a store or a software company, it means pricing, plans and integrations. Then ask ChatGPT about each item the way a customer would.
Also compare the answers with the brand’s site, profiles and directory listings. Whenever these channels disagree, AI can make mistakes about the company. The guide on why AI doesn’t cite your brand shows how to check each channel.
Can you ask OpenAI to fix the answer?
For wrong answers from ChatGPT about company data, OpenAI does not describe a correction channel in its documentation. The thumbs-down button sends feedback, and OpenAI may use the conversation to train future models. However, for wrong information about a person, such as a partner, the privacy policy allows a correction request, reviewed case by case.
According to OpenAI’s Help Center, anyone who marks wrong answers from ChatGPT with a thumbs-down gives the company feedback. On individual plans, OpenAI may then use the entire conversation to train its models. This applies even to people who turned off training on their chats. However, the page says nothing about correcting the answer or about timing.
OpenAI’s privacy policy is direct about the problem. ChatGPT generates answers by predicting the most likely words, but they are not always the most accurate. So anyone who finds wrong information about themselves can request a correction or removal through the privacy portal. OpenAI reviews each request based on applicable law and the technical capabilities of its models.
This route works for a real partner whose name AI swapped, but not for a wrong product price. Correcting ChatGPT inside the chat doesn’t help customers either. After all, according to OpenAI, ChatGPT’s memory personalizes answers for the account that uses it. For the step-by-step, see how to report a ChatGPT error about your company.
How do you trace ChatGPT mistakes back to the source?
Ask the question again and read the sources the answer cites. Whenever wrong answers from ChatGPT cite pages, open each one and look for the wrong fact. But if ChatGPT makes mistakes without citing any source, the fact likely comes from the model’s memory, built from old pages. In that case, search the web for whoever still publishes it.
- Ask again in a new chat, in a customer’s words, and request a web search. OpenAI lets you regenerate an answer with web search, when the option is available.
- Open each cited source. According to OpenAI, search results and citations can be incomplete, outdated or incorrect. So check whether the page supports the answer, and check its last update date.
- If there is no source, ask where the fact came from. In the Hashtag test, for example, the answer “conhecimento adquirido até 2024” (“knowledge acquired up to 2024”) showed the mistake came from training.
- Search for the wrong fact in quotes on Google and Bing. Then you’ll find the pages that still publish it, such as directories, old articles, marketplaces and company profiles.
- Record the date, engine, question, answer and source, so that you can compare after the fix.
For clinics and stores, check the Google Business Profile first, as the guide on GEO for local businesses shows. Then see the guide to the sources ChatGPT cites most to learn how to correct a fact at the source.
How long does it take for a fix to show up?
There is no official timeline, because it depends on where the mistake came from. If ChatGPT makes mistakes because of a search result, the answer may change after a new crawl of the fixed page. But if it came from training, the mistake may last until a model trained on newer data is in use.
Crawling is the part a company can speed up. According to the IndexNow FAQ, search engine crawlers can take days or even weeks to notice a change. With IndexNow, however, a site notifies search engines right away, including about price changes. Still, the protocol does not guarantee immediate indexing. According to Microsoft, Bing powers ChatGPT, as the guide on how to show up in ChatGPT explains.
In the model’s memory, the wait is longer. OpenAI’s developer documentation lists three flagship models, all from the GPT-6 line. Their knowledge cutoffs range from April 30 to May 18, 2026. Therefore, whatever changed after that is not in their memory. Meanwhile, without search, ChatGPT may repeat the old fact and give wrong answers.
To see whether the fix worked, repeat the same questions for a few weeks, always in new chats. Then count how many answers still contain the mistake. A single answer proves nothing, because answers change from run to run. The method is in the guide on how to measure what AI says about your brand.
What we found when analyzing 9 Portuguese-language pages on the topic
We ran 9 searches on why ChatGPT makes mistakes and invents company facts. Then we analyzed the 9 pages in Portuguese that came up and explain the problem. Four cover mistakes about company data, such as pricing, partners and opening hours. However, only one cites OpenAI’s 2025 explanation that evaluations reward guessing.
Only two tell training mistakes apart from mistakes caused by a web page. Three mention reporting the mistake to the platform, but none cites the correction request in OpenAI’s privacy policy. Likewise, only one mentions timing, and only one has figures on AI mistakes about brands.
Frequently asked questions
No. OpenAI describes wrong answers from ChatGPT as guessing, not lying. The model predicts the most likely words, which are not always correct. Also, because training and evaluations reward guessing, it tends to offer a plausible fact instead of admitting it doesn’t know.
Searchable offers one hypothesis: AI makes more mistakes because it has less information about them. So it fills the gaps with guesses. In the report, companies with up to 50 employees had 9.4% false facts, compared with 6.3% at large ones. Likewise, according to OpenAI, rare facts tend to turn into hallucinations.
Every AI makes mistakes, but the rate varies. In the Searchable report, ChatGPT had 8.1% false facts about brands. Google’s AI Overviews had 11.4%, and Perplexity 11.5%. However, Searchable sells AI brand monitoring, and the companies themselves reviewed the facts.
Asking for sources helps, but it doesn’t prevent mistakes. Sources show whether the answer came from the web or from the model’s memory, as in the Hashtag Treinamentos test. Still, OpenAI warns that citations can be incomplete, outdated or wrong. So open each source and check the fact.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Why language models hallucinate, a study on where hallucinations come from (2025).
- Searchable, The AI brand misinformation report, false facts about 1,704 brands (2026).
- OpenAI Help Center, Does ChatGPT tell the truth? (2026).
- OpenAI Help Center, Searching the web with ChatGPT (2026).
- OpenAI Help Center, How your data is used to improve model performance (2026).
- OpenAI, privacy policy, with the request to correct personal data (2026).
- OpenAI, model documentation, with knowledge cutoffs (2026).
- Hashtag Treinamentos, O ChatGPT erra! Aprenda a identificar e corrigir alucinações, in Portuguese (2025).
