Artificial intelligence can produce an answer that is clear, polished, detailed, and completely confident.

And still be wrong.

That is one of the most important things to understand about generative AI.

Good grammar is not evidence.

A professional tone is not evidence.

A detailed explanation is not evidence.

And confidence is certainly not evidence.

The ability of AI to produce convincing language can make incorrect information unusually easy to believe.

That is why learning to distinguish fluency from factual reliability is becoming an essential digital-literacy skill.

What Is an AI “Hallucination”?

You may have heard people describe incorrect AI-generated information as a hallucination.

The U.S. National Institute of Standards and Technology, or NIST, uses the more technical term confabulation.

NIST defines confabulation as generative AI producing and confidently presenting erroneous or false information. The organisation notes that these outputs are also commonly called hallucinations or fabrications.

A confabulation might involve an AI:

  • Inventing a historical event
  • Giving the wrong date
  • Attributing a quotation to the wrong person
  • Describing a study that does not exist
  • Inventing statistics
  • Creating a nonexistent court case
  • Citing a fake article or book
  • Giving an incorrect technical explanation
  • Combining several true facts into a false conclusion

The important part is that the answer may not sound uncertain.

It can look perfectly credible.

Why Can AI Produce False Information So Confidently?

To understand the problem, it helps to understand what a language model is doing.

A large language model has learned statistical patterns from enormous amounts of text.

When generating an answer, it predicts suitable continuations based on those learned patterns.

NIST specifically notes that large language models predict the next token or word based on statistical patterns in training data. That process can produce factually accurate responses—but it can also produce information that is inaccurate or internally inconsistent.

This creates an important distinction:

Producing plausible language is not the same operation as independently proving that every statement is true.

Imagine being extremely good at completing sentences.

You might know that this looks plausible:

“The study was published by Professor X at University Y in 2023.”

But producing a plausible university, professor, title, and year does not necessarily mean those details correspond to a real publication.

AI can sometimes construct the pattern of a correct answer without possessing reliable evidence for the details inside it.

The More Specific the Error, the More Convincing It Can Look

Some incorrect AI answers are obvious.

If an AI tells you that Paris is in Brazil, most people will recognise the problem immediately.

The more dangerous errors are often the ones surrounded by believable details.

Consider an answer containing:

  • An author's full name
  • A publication date
  • A journal title
  • A quotation
  • A statistic
  • A professional-looking citation

Those details create an impression of research.

But if the source does not exist, the additional detail has made the misinformation more convincing rather than more accurate.

NIST specifically warns that generative AI systems can produce confabulated citations or reasoning that appear to justify an answer and may cause people to trust it more than they should.

A Citation Is Only Useful If It Is Real

This deserves particular attention.

Suppose an AI says:

“According to a 2025 Harvard study, 73% of businesses…”

That sentence contains several things that make it sound authoritative:

  • A respected institution
  • A recent date
  • A precise statistic
  • The word “study”

None of those features proves that the study exists.

Before relying on the claim, ask:

Can I actually find the source?

And if you find something with a similar title:

Does the source really say what the AI claims it says?

Verification means checking the evidence—not merely checking whether the citation looks professionally formatted.

Real Sources Can Also Be Misrepresented

Fabricated sources are not the only problem.

An AI may identify a real article but incorrectly describe its findings.

It might:

  • Confuse correlation with causation
  • Exaggerate the conclusion
  • Misstate the sample size
  • Attribute one researcher's view to an entire institution
  • Mix findings from multiple studies
  • Omit an important limitation
  • Use an outdated version of information

So verifying a factual claim sometimes requires more than establishing that a webpage exists.

You may need to read the source itself.

Confidence Is a Style of Writing

Humans naturally interpret linguistic confidence as a clue to knowledge.

Compare:

“I think the answer might be 42.”

with:

“The answer is 42.”

The second sounds more authoritative.

But an AI system's confident wording does not necessarily reflect the kind of internal certainty a human expert might have after reviewing evidence.

It may simply be the form of response generated from the prompt and context.

That is why phrases such as:

“Definitely.”

“Research proves…”

or:

“The correct answer is…”

should not automatically increase your trust in the underlying claim.

Evaluate the evidence instead.

This Matters More When the Consequences Are Higher

Not every AI mistake carries the same risk.

If you ask an AI to suggest fictional names for characters and one is unusual, very little is at stake.

If you ask it for:

  • Medical information
  • Legal information
  • Financial information
  • Cybersecurity guidance
  • Academic citations
  • Business compliance requirements
  • Current laws or regulations
  • Safety instructions
  • Information used to make an important decision

the consequences of an incorrect answer can be much greater.

NIST highlights confabulation as particularly important in applications involving consequential decision-making because people may act on incorrect information precisely because the response appears confident.

The higher the consequence, the stronger the verification should be.

AI Is Still Extremely Useful

The existence of confabulations does not mean AI is useless for research or knowledge work.

It means AI should be used according to its strengths.

AI can be extremely useful for:

  • Generating questions
  • Brainstorming ideas
  • Explaining unfamiliar concepts
  • Summarising material you provide
  • Comparing arguments
  • Organising information
  • Drafting outlines
  • Identifying areas worth researching
  • Rewriting difficult information clearly
  • Accelerating an existing research process

The problem begins when useful assistance becomes unquestioned authority.

A better approach is:

Use AI to accelerate thinking. Use evidence to establish facts.

A Simple Verification Habit

When an AI-generated answer contains an important factual claim, ask four questions.

1. What exactly is being claimed?

Separate the factual statement from the surrounding explanation.

For example:

Claim: “The regulation came into force on January 1, 2026.”

That can be checked.

2. What is the source?

Look for the strongest available source.

For regulations, that may be an official government publication.

For scientific research, it may be the original paper.

For company announcements, it may be the company's official statement.

For statistics, it may be the organisation that collected the data.

3. Does the source actually support the claim?

Do not stop because the link exists.

Check whether it supports the specific statement being made.

4. Does the date matter?

Information can be correct and still be outdated.

This is particularly important for:

  • Prices
  • Laws
  • Political leaders
  • Software versions
  • Product specifications
  • Medical guidance
  • Company executives
  • Statistics
  • Offers and promotions

A source from three years ago may not answer a question about what is true today.

Be Especially Suspicious of Precise Details

Ironically, very specific information can sometimes deserve more verification, not less.

Pay additional attention when an AI gives you:

  • Exact percentages
  • Exact dates
  • Exact quotations
  • Study titles
  • Authors
  • URLs
  • Court cases
  • Regulation numbers
  • Product prices
  • Technical specifications

These details are highly useful when correct.

They are also highly misleading when invented.

Ask for Sources — Then Check Them

Asking an AI for sources can improve your research process.

But it does not remove the need for verification.

The correct workflow is not:

AI gives citation → therefore claim is true

It is:

AI gives citation → locate source → inspect source → confirm claim

That extra step is what turns generated information into evidence-backed information.

Cross-Checking Can Help

For important information, consider checking multiple independent sources.

If one AI-generated claim appears in:

  • An official government document
  • A reputable research institution
  • A respected news organisation

your confidence may reasonably increase.

But source quality matters more than simply counting how many webpages repeat the same statement.

Ten websites copying the same inaccurate information are not ten independent confirmations.

Don't Outsource Judgment

Perhaps the most important habit is psychological.

AI systems are becoming increasingly good at producing answers that feel finished.

That feeling can tempt users to stop thinking.

A polished paragraph may make a question feel settled.

A table may make numbers feel authoritative.

A citation may make a statement feel researched.

But presentation and reliability are separate qualities.

The user still needs to ask:

How do we know this is true?

Why It Matters

Generative AI makes producing clear, persuasive language easier than ever.

That is enormously useful.

It also means that incorrect information can arrive dressed in the language of expertise.

Understanding that distinction is becoming an essential part of digital literacy.

The practical rule is simple:

Use AI to accelerate research, analysis and thinking—but verify consequential factual claims against trustworthy sources.

Because an answer can be fluent.

It can be detailed.

It can be confident.

It can even contain citations.

And it can still be wrong.

Fluency is not evidence.