AI hallucinations: why AI makes mistakes and how to check

In the lesson on writing AI prompts, you learned how to formulate requests. In the lesson on writing and editing text with AI, you learned how to create and revise content. Now you'll learn how to check whether the information in AI responses is accurate.

Consider this AI-generated statement: “The company was founded in 2018, its revenue is $15 million, and its profit increased because of advertising.” You need to verify the founding year, the revenue figure, and whether the claim about advertising is supported by evidence. Well-written text does not prove that any of these claims are correct.

What are AI hallucinations?

An AI hallucination is a fabricated, inaccurate, or unsupported claim that an AI model presents as fact.

For example, you ask for three research papers on a particular topic. The model provides author names, journal titles, and publication years. When you check them, you discover that one paper does not exist and another covers a completely different topic.

Hallucinations occur across different language models, including ChatGPT, Claude, Gemini, DeepSeek, and others. The types and frequency of errors depend on the specific model, task, available information, and generation conditions.

Why AI makes up facts

A language model generates text based on patterns learned during training and information provided in the current conversation. It produces a plausible continuation of your prompt. This generation process does not automatically verify every fact.

Models learn the structure of academic citations, legal explanations, and news reports. When accurate information is missing, a model may fill the gaps with details that fit the context.

Ways to reduce the risk:

• Provide the original document or verified data instead of relying on the model's assumptions.
• Ask a focused question with specific conditions.
• Allow the model to say “I don't know” when there is insufficient information.
• Independently verify important claims.

Errors can also result from outdated information, inaccurate training data, ambiguous questions, and lost details in long conversations. Sometimes the user includes a false assumption in the question, and the model agrees with it.

If you ask, “Why was this law passed in 2024?”, the model may start explaining the reasons even if the law was passed earlier or does not exist at all.

Why newer AI models still make mistakes

AI models continue to improve, but their reliability varies by task. A model may explain a common concept correctly while getting an obscure date, document number, or research paper title wrong.

A longer, more detailed answer is not necessarily more accurate. Additional details can make an answer appear credible even when some of them are fabricated.

There is no single hallucination rate that applies to every AI model. Results from individual benchmarks depend on the questions used and how answers are evaluated. You cannot meaningfully compare models using a single number without knowing how the test was conducted.

What kinds of mistakes does AI make?

Hallucinations often overlap with other types of errors. When checking an answer, it helps to identify exactly what might have gone wrong.

Error type	Example	How to verify
Fabricated fact	An award that does not exist	Official biography
Fabricated source	A nonexistent research paper	Journal archive or academic database
Misattributed quote	A quote attributed to the wrong person	Original text
Calculation error	15 × 200 = 3500	Calculator: 3000
Outdated information	An old subscription price	Current pricing
Reasoning error	Correlation mistaken for causation	Analysis of other factors
Incorrect conclusion	Revenue growth described as profit growth	Review expenses
Misrepresentation of a document	An exception is omitted	Comparison with the original
Sycophancy	The model agrees with a false premise	Verify the question itself

Another risk is mixing correct and incorrect claims. If an answer contains four verified facts, that does not automatically make the fifth claim reliable.

How to identify claims that need verification

Consider this example AI response. Some details are intentionally fabricated.

The arithmetic error in this example is clear: 12 × 5 = $60. The registration year and report cannot be classified as true or false without identifying the specific company and publication. The claim about advertising requires additional evidence.

A claim you cannot verify is not necessarily false.

After checking, assign one of three statuses:

• “Verified”: An independent source has been found that actually supports the claim.
• “Disproven”: Reliable evidence contradicts the claim.
• “Unverified”: There is not yet enough evidence to confirm the claim.

The last status is especially useful when dealing with obscure events, private reports, and unavailable documents. Do not replace missing information with a guess.

Can AI make mistakes in images and videos?

Yes. An AI image generator may add extra fingers, distort text, or create product details that do not exist. AI-generated videos may contain inconsistencies in faces, objects, and sequences of actions.

When enhancing an old photograph, a model may add details that cannot be recovered from the source image. The reconstructed detail may look realistic, but its historical accuracy remains unknown.

This lesson focuses primarily on verifying text responses. You'll learn how to evaluate images and videos in the corresponding sections of the course.

When can you trust an AI response?

The amount of verification needed depends on the consequences of an error.

Task type	Risk	What to do
Ideas, names, creative drafts	Low	Check whether the result meets your requirements
Explanations of common concepts	Moderate	Verify definitions
Dates, names, scientific information	Elevated	Find an original source
Summaries of contracts or documents	Elevated	Check terms and exceptions
Calculations and metrics	Elevated	Verify formulas and input numbers
News, prices, current regulations	High	Check an up-to-date publication
Medicine, law, personal finance	High	Consult specialized sources and qualified professionals

This is a qualitative scale for deciding how thoroughly to verify an answer. It does not indicate the probability that a particular model will make a mistake.

If AI suggests a name for a café, you can judge how it sounds and whether it fits your preferences. If AI tells you how much a mandatory payment will be, you need to verify the amount, applicable period, and basis for the calculation.

How to check an AI response: five steps

First, identify which details could affect your decision or the finished result. Then check each significant claim systematically.

Step 1. Identify claims that need verification

Consider this response: “The company was founded in 2018, operates in five countries, and generated $120 million in revenue.”

You need to verify the founding year, the countries where the company operates, and its revenue. For the revenue figure, you also need to establish the reporting period: month, quarter, or year.

If the answer is long, ask AI to prepare a list:

List all factual claims in your previous response. For each claim, explain what needs to be verified and which original source would provide the evidence. Do not describe claims as verified unless you have consulted external sources.

Review the original response yourself. The model may overlook an incorrectly worded claim or fail to identify an important exception.

Step 2. Find the original source

An original source is a document or other material that directly supports the claim you are checking.

For a product price, use the manufacturer's or seller's page with current pricing and terms. For a legal requirement, use the officially published law. For research findings, use the original scientific paper.

Follow these steps:

1. Find the exact name of the document, study, or organization.
2. Locate it using Google, another search engine, or the official website.
3. Find the section relevant to the claim.
4. Check the date, version, and scope of applicability.
5. Record the result: verified, disproven, or unverified.

For scientific research, useful resources include Google Scholar, journal websites, and academic libraries. For legislation, use the official legal portals of the relevant country.

Pay attention to where the information comes from. Be cautious with materials that have no author or publication date, unexplained figures, or references to unfamiliar publications.

Two websites may repeat the same mistake. Find the original publication instead of counting how many websites repeat the same claim.

Step 3. Double-check the numbers

Enter the original values and formula into a calculator or spreadsheet.

For example, 24 items at $7 each cost $168. If the AI model calculates $188, the multiplication is wrong.

Check units of measurement and reporting periods. You cannot directly compare monthly expenses with annual revenue. Dollars are different from thousands of dollars, and a change of 5 percentage points is not the same as a 5% increase.

Also verify that the correct metrics are being used. Revenue is not profit. The total value of orders does not necessarily equal the payments received.

If the model analyzed a file, check for missing rows, incorrect delimiters, or values expressed in different currencies.

Step 4. Check the reasoning

Correct numbers can still lead to an incorrect conclusion.

For example, an online store's revenue increases after an advertising change. The AI model concludes that the advertising change caused the increase.

But seasonal demand, product selection, or prices may also have changed. Additional evidence is needed to establish the cause.

Ask the model to consider alternative explanations:

Separate observations from conclusions. What evidence supports the proposed cause? What other factors could have affected the result? What additional information would help distinguish between these explanations?

Check whether the model is simply supporting the user's original position. A prompt such as “Prove that the new advertising increased sales” assumes the desired conclusion from the start.

A better approach is to ask whether the data supports a connection between the advertising change and sales.

Step 5. Compare the response with your question

Reread your original prompt and check whether every requirement has been met.

You asked for monthly expenses but received an annual calculation. You requested a price comparison, but the model compared features. You asked for a contract summary, but the AI added its own recommendations without distinguishing them from the document's contents.

These responses may contain accurate information while failing to complete the requested task.

After verification, correct the result. If you find one typo, a targeted edit is enough. If the input data or main conclusion is wrong, rebuild the answer using verified information.

What to do when AI keeps repeating a mistake

Sometimes a user points out an incorrect date, the model apologizes, and it provides a replacement. Then the new date turns out to be wrong too.

An apology does not mean the AI model has obtained new information. Asking again without an independent source may produce another guess.

Follow this process:

1. Stop asking the model to guess a new answer.
2. Find reliable evidence yourself.
3. Provide the model with the verified source or corrected data.
4. Ask it to revise the specific claim.
5. Compare the corrected response with the source.

If you cannot verify a claim, remove the unsupported detail or clearly state the uncertainty.

For example, instead of inventing a company's founding year, you could write: “The founding year needs to be verified against company registration records.”

If the model repeatedly misrepresents a document you provided, do not keep requesting corrections indefinitely. Extract the relevant passages manually or use another processing method.

How to get AI to check its own answers

Self-checking can help identify contradictions and create a plan for independent verification. But a model does not become an independent source simply by reviewing its own response.

Separate facts from assumptions

Review your previous response. Classify each claim as “Fact requiring verification,” “Assumption,” or “Insufficient information.” Explain what needs to be checked. Do not claim to have verified sources unless you actually opened them.

This helps you identify which parts are based on the original information and which rely on interpretation.

Do not ask the model to assign an exact accuracy percentage to its entire response. Without a dedicated evaluation, that percentage may create a false impression of measured reliability.

Look for weaknesses in the answer

Identify potential errors in your response. Check dates, numbers, names, quotes, exceptions, and cause-and-effect relationships. For each questionable claim, suggest an independent way to verify it.

The model may identify some problems. If it says “No errors found,” that only describes the result of its own review.

Recalculate step by step

List all input numbers and formulas. Show the calculation step by step. Check units, reporting periods, and the final total. Separately identify any metrics that cannot be calculated because information is missing.

Then repeat the key calculations using a calculator. A language model can make mistakes even when it provides a detailed explanation of its calculations.

Compare two AI models

Send the same question to two different AI models. Compare their factual claims, explanations, and acknowledged limitations.

The Neuromia AI chat offers Mia for free, while GPT, Claude, Gemini, and Grok are available for credits. You can switch models within the same conversation.

Keep in mind that previous messages may influence the next model's response. For a more independent comparison, use the same original context without including another model's completed answer whenever possible.

Matching answers from two models do not replace an external source. Both models may have learned the same common misconception.

Hands-on exercise: review an AI response

Check whether the model separates general recommendations from information that depends on the country, date, or current regulations.

If it lists specific requirements but cannot support them with an up-to-date source, mark those claims as unverified.

How to reduce AI hallucinations before they happen

In previous lessons, you learned how to structure a prompt. To improve factual accuracy, add a few more constraints.

Use a document as the basis for the answer

Instead of asking a general question about a contract, attach the file and ask the model to find a specific clause:

Use only the attached document. Find the contract termination terms and identify the relevant clauses with exact quotes. If the terms are not included, write “Not found in the document.” Do not add rules from other contracts.

This instruction reduces the need for the model to supplement the answer with general knowledge.

However, providing a document does not guarantee accuracy. The model may confuse clause numbers, misread a table, or overlook an exception.

Compare quotes with the original. If a PDF contains scanned pages, pay particular attention to whether the text was extracted correctly.

In Neuromia, you can attach PDF, TXT, and CSV files containing up to 60,000 characters of text per file, as well as images. Before analyzing a file, make sure the model has actually received the information it needs.

Break large tasks into smaller steps

When working with a long document, first ask the model to locate the relevant passages. Next, have it extract the facts. Then ask it to prepare a conclusion and identify any questionable details.

Check each intermediate result before using it in the next step. Otherwise, an initial error may carry over into the summary, calculations, and final recommendation.

Don't suggest the answer you want

A prompt such as “Explain why my approach is better” encourages the model to look for supporting arguments.

Instead, ask:

Compare two options using the specified criteria. List the advantages and disadvantages of each. Do not choose a winner if there is insufficient information.

The same principle applies to product evaluations, sales analysis, and business decisions.

Special cases: what needs closer verification

Links and scientific publications

Open every link included in the response. Check the publication's title, author, date, and content.

Even a real article can be cited incorrectly. An AI model may attribute a conclusion to a publication that does not actually contain it.

If you cannot find a source, do not include it in your bibliography. If necessary, replace the unsupported claim with a more cautious statement.

Quotes

Quotation marks do not guarantee that the wording is exact.

Find the original passage. Check whether a negation, exception, or condition has been omitted. If you cannot find the exact wording, do not present it as a direct quote.

Laws and contracts

AI can explain a legal term or help you prepare questions for a lawyer. However, it may confuse different versions of a law, countries, or conditions under which a rule applies.

Check the current official text. For decisions with legal consequences, consult a qualified professional.

Do not upload confidential contracts or other people's personal information without the necessary permission.

Health and finances

AI can explain general concepts, help organize expenses, and prepare questions for a professional.

Do not rely on an AI response as the sole basis for medical treatment, medication dosage, tax filings, or money transfers.

Verify medical recommendations with a qualified healthcare professional and financial obligations against official rules and documents.

Academic assignments

Compare quotes with the original books and articles. Verify bibliographic references and factual claims.

If AI summarizes educational material, check that it has not added events or conclusions that are absent from the original.

Follow your educational institution's rules on AI use.

Recent events and prices

A model may not know about the latest changes.

In the Neuromia chat, AI models do not perform internet searches. They respond using their existing knowledge and the materials you provide.

For current prices, news, regulations, and officeholders, check the information yourself using Google, another search engine, and official websites.

Document and spreadsheet analysis

A document may contain appendices, footnotes, and exceptions that the model overlooks.

A spreadsheet may contain blank rows, mixed units of measurement, or data covering different periods.

Before calculating, make sure the source data is complete. After calculating, verify the formulas and what the resulting metrics actually mean.

You'll use these skills in the next lesson on working with spreadsheets.

Hands-on exercise: find errors in an AI response

The following practice text intentionally contains three errors: an incorrect historical date, an unverified source, and a calculation error.

First, ask the model to identify what needs to be checked. Do not give it the correct answers in advance.

Then independently verify the date and source. Check the calculation using a calculator.

After verifying the information, write a corrected version of the text. Do not replace the unverified article with another publication unless you have personally checked that the replacement exists and supports the claim.

Hands-on exercise: compare two AI models

Send the same prompt to two models. Write down the dates, names, numbers, and qualifications included in their answers.

Compare the responses and find independent confirmation of at least one main claim.

If the answers differ, do not automatically choose the more detailed one. Verify both.

If they match, you still need an independent source.

Test your knowledge

Answer five questions about verifying sources, checking calculations, evaluating reasoning, and understanding the limitations of AI self-checking.

Checklist: how to verify an AI response

For creative writing, checking whether the result meets your requirements is usually enough. Calculations, legal documents, and medical questions require additional verification based on the potential consequences of an error.

Frequently asked questions

What are AI hallucinations in simple terms?

AI hallucinations are incorrect or fabricated claims that an AI model presents as facts. Examples include a nonexistent scientific paper, an incorrect date, or a made-up quote. The response may still sound convincing.

Why does AI lie even when you ask it to tell the truth?

An AI model generates a plausible response, but it may make mistakes when information is missing or it misinterprets the context. The word “lie” is misleading here: an incorrect answer does not imply intentional deception. Asking a model to tell the truth does not replace fact-checking.

Can AI hallucinations be eliminated completely?

There is no universal method that guarantees error-free responses. Providing verified source data, asking focused questions, using external sources, and checking claims independently can reduce the risk.

Which AI model makes the fewest mistakes?

It depends on the task, model version, and evaluation method. Results from one benchmark do not necessarily apply to every type of question. For a practical comparison, give different models the same tasks with known correct answers.

Do AI models search the internet?

Some AI applications support web search, while others respond without it. Models in the Neuromia chat do not search the internet. For current information, use external sources.

Can you trust AI for medical and legal advice?

AI can help explain general concepts and prepare questions for professionals. Decisions about medical treatment and legal obligations require checking official sources and consulting qualified professionals.

How do you verify AI-generated citations and links?

Open the link, locate the cited publication, and check whether it actually supports the claim. If the source does not exist or does not contain the relevant information, do not use it as evidence.

What to learn next

You've learned how to separate facts from assumptions, verify sources, recalculate figures, and identify errors in reasoning.

The next lesson is “How to use AI for spreadsheets and data analysis.” You'll work with source data, identify patterns, and verify calculations.

To prepare, find a small spreadsheet containing your expenses or sales. Check whether the units are consistent, whether any values are missing, and whether the reporting period for each metric is clear.

You'll use this data in the next hands-on exercise.