Week 11: Hallucinations and False Confidence

Welcome to Week 11 of the AI Literacy Course.

AI systems can produce answers that are fluent, detailed, and persuasive. They can also produce information that is fabricated, outdated, misquoted, unsupported, or applied in the wrong context.

The problem is not only that AI can be wrong.

The problem is that AI can be wrong in a style that sounds polished and certain.

The central rule for this lesson is:

Tone never replaces evidence.

Learning goals

By the end of this lesson, learners should be able to:

• Explain hallucination without treating AI as a conscious person

• Distinguish hallucinations from other AI output errors

• Recognise false confidence and misleading uncertainty

• Identify checkable claims

• Prioritise claims by importance and possible consequences

• Use a short verification chain outside the chatbot

• Correct errors transparently

• Match trust to evidence, task, and risk

1. Tone never replaces evidence

AI-generated text can appear trustworthy because it uses:

• clear grammar;

• professional vocabulary;

• precise numbers;

• confident statements;

• academic-looking citations;

• organised headings;

• detailed explanations.

These features may improve readability, but they do not prove that the information is correct.

Strong grammar is not evidence.

Length is not evidence.

Precision is not evidence.

Confidence is not evidence.

A claim becomes trustworthy when it is supported by appropriate evidence and used in the correct context.

2. What hallucination means

An AI hallucination is generated information that is fabricated or presented as factual without adequate grounding, often in a plausible and confident form.

Examples include:

• an invented report;

• a study that does not exist;

• a fabricated quotation;

• a false name or date;

• a link to a nonexistent page;

• a claim attributed to a document that does not contain it.

The word “hallucination” describes an output problem. It does not mean that the AI system sees, believes, remembers, or experiences something in the human sense.

Not every error is a hallucination

AI outputs can contain many kinds of problems.

An outdated rule may once have been correct. A calculation error may result from incorrect steps. A real source may be applied to the wrong country. These are serious errors, but it is useful to identify them precisely.

Instead of calling every problem a hallucination, ask:

What kind of error is this?

3. Why plausible errors occur

Language models generate responses by predicting likely language from learned patterns and available context.

They can reproduce the form of:

• a research citation;

• an official report;

• a legal explanation;

• a quotation;

• a confident technical answer;

without reliably establishing that every detail is real.

Some AI systems can also use:

• web search;

• databases;

• calculators;

• uploaded documents;

• retrieval tools.

These tools may improve grounding, but they do not guarantee accuracy. The AI may still:

• misunderstand the source;

• quote it incorrectly;

• combine information from different sources;

• omit an important limitation;

• apply information outside its proper context;

• add unsupported details.

Tool access may reduce some risks, but it does not remove the need to inspect important evidence.

4. Recognise different error types

AI outputs may contain several different kinds of error.

Error type	Meaning	Example
Fabrication	Information, a quotation, citation, or source is invented	A report that does not exist
Factual error	A checkable detail is wrong	Incorrect date, name, or number
Source mismatch	A real source does not support the claim	An article is cited for a conclusion it never makes
Ungrounded summary	AI attributes content to a supplied text that is absent	A summary includes a claim not found in the document
Outdated information	Information is no longer current	An old rule is presented as current
Context error	Information is applied outside its proper scope	Another country’s law is applied to Sweden
Reasoning or calculation error	The steps or conclusion do not follow correctly	An incorrect percentage calculation
Internal contradiction	Different parts of the response conflict	Two incompatible dates appear in the same answer

A single response may contain more than one error type.

For example, AI may name a real report but give it the wrong date and claim that it supports a conclusion outside its scope.

Partly correct answers are often harder to detect than completely invented ones because familiar details create trust.

5. False confidence and misleading uncertainty

False confidence occurs when the strength of the language exceeds the available evidence.

Warning phrases may include:

• “This is certainly true.”

• “The research clearly proves…”

• “According to the official report…”

• “It is well established that…”

• “The law requires…”

These expressions may be justified in some situations, but the wording itself proves nothing.

Cautious language is not proof either

AI may also say:

• “I may be mistaken…”

• “This is probably correct…”

• “Based on available information…”

• “There is an 80 per cent chance…”

Cautious wording and confidence percentages can also be unsupported.

The important question is not:

“Does the answer sound confident or cautious?”

The important question is:

“What evidence supports this claim?”

6. Mark and prioritise checkable claims

A checkable claim is a statement that can be compared with evidence.

Examples include:

• names;

• dates;

• numbers;

• quotations;

• named sources;

• causal claims;

• descriptions of laws or policies;

• medical, legal, financial, or safety advice;

• claims about what a document says.

Do not treat every sentence as equally important.

Prioritise a claim when it is:

• central to the answer;

• precise or suspicious;

• the basis for later conclusions;

• consequential if wrong;

• related to rights, health, safety, finance, law, assessment, or public information.

Use two questions:

1. How important is this claim to the answer?

2. What could happen if it is wrong?

For example, a minor wording error in a fictional brainstorm has little consequence. A false legal rule or incorrect medication instruction can cause serious harm.

The greater the possible harm, the stronger the evidence and accountable human review must be.

7. Use a short verification chain

When an important claim needs checking, use this sequence:

Step 1: Mark the claim

Copy or record the exact statement.

Do not rely on a vague memory of what AI said.

Step 2: Identify the possible error

Ask whether it may involve:

• fabrication;

• a factual error;

• a source mismatch;

• outdated information;

• a context problem;

• faulty reasoning;

• another error type.

At this stage, the classification is provisional. Verification may change it.

Step 3: Rank the risk

Consider the claim’s centrality and possible consequences.

Check high-priority claims first.

Step 4: Leave the chatbot

Asking the same AI:

“Are you sure?”

is not independent verification.

Asking another chatbot is not automatically independent verification either. Different systems may repeat the same unsupported claim.

Step 5: Locate an appropriate source

Depending on the claim, this may be:

• an official organisation;

• the original report;

• the supplied source text;

• an authoritative database;

• a reliable calculation method;

• a qualified person.

A search-result title or snippet is not enough. Open the relevant source.

Step 6: Check exact support

Ask:

• Does the source exist?

• Does it contain the claimed information?

• Does the relevant passage support the exact claim?

• Is the information current?

• Does it apply to the correct country, group, or situation?

A real citation can still be misused.

Step 7: Decide

Choose one action:

• Keep: The claim is adequately supported.

• Correct: The claim contains a repairable error.

• Reject: The claim is false, fabricated, or unsuitable.

• Escalate: A qualified or accountable person is needed.

In Lesson 12, we will develop source investigation further using the SIFT method.

8. Correct errors transparently

Finding an error may change more than one sentence.

Suppose AI recommends an action because “a 2025 government report” supports it. During verification, the learner discovers that the report does not exist.

Removing the citation is not enough. The learner must reconsider:

• the recommendation;

• the reasoning based on the report;

• the confidence of the conclusion;

• whether other evidence exists;

• whether the whole answer should be rejected.

Use a correction log:

Original claim	Error type	Risk	Evidence checked	Decision	Effect on the answer
				Keep, correct, reject, or escalate
				Keep, correct, reject, or escalate
				Keep, correct, reject, or escalate

A transparent correction explains:

• what was wrong;

• how it was checked;

• what changed;

• how the correction affected the wider answer;

• what remains uncertain.

Quietly replacing a false detail may hide how much earlier reasoning depended on it.

9. Build calibrated trust

AI literacy does not mean trusting everything. It also does not mean rejecting every AI output.

Calibrated trust means matching reliance to:

• available evidence;

• the task;

• possible consequences;

• source quality;

• local context;

• accountability.

AI use	Appropriate response
Brainstorming fictional names	Low verification need
Generating possible study questions	Review before use
Summarising an assigned source	Compare with the source
Explaining a current law or medical issue	Verify through authoritative sources
Supporting a high-consequence decision	Accountable human judgement required

Trust should increase when evidence is strong and applicable.

Trust should decrease when:

• evidence is missing;

• the source cannot be located;

• the claim is outside the source’s scope;

• information is outdated;

• the consequences of error are serious;

• no accountable person is involved.

The aim is proportionate reliance—not automatic trust or automatic rejection.

Classroom activity: Hallucination and confidence audit

Choose or receive an AI-generated response containing at least three checkable claims.

Use a neutral educational, historical, scientific, or civic topic. Avoid acting on unverified high-stakes medical, legal, financial, or safety information.

Step 1: Mark three claims

Choose claims that contain names, dates, numbers, quotations, sources, rules, causal statements, or other checkable information.

Step 2: Classify the possible errors

For each claim, identify a possible error type.

The classification may change after checking.

Step 3: Rank the claims

Rank them according to:

• importance to the answer;

• possible consequences if wrong.

Step 4: Verify the two highest-priority claims

Leave the chatbot and locate appropriate sources.

Check whether each source supports the exact claim and applies to the relevant context.

Step 5: Decide

For each checked claim, choose:

• keep;

• correct;

• reject;

• escalate.

Step 6: Record the impact

If a claim was wrong, explain how the correction affects the rest of the answer.

Step 7: Complete the correction log

Original claim	Possible error type	Risk	Evidence checked	Decision	Effect on the answer
Claim 1
Claim 2
Claim 3

Step 8: Reflect

Write approximately 300–500 words explaining:

• which language made the answer appear trustworthy;

• which claims you prioritised and why;

• what evidence you checked;

• which error types you found;

• what you kept, corrected, rejected, or escalated;

• how the corrections affected the answer;

• how much trust the revised response deserves.

Reflection questions

1. What is an AI hallucination?

2. Why should the word not be understood as a human experience?

3. Why is not every AI error a hallucination?

4. Which confidence signals can mislead users?

5. Why is cautious wording not evidence of accuracy?

6. Which claims deserve priority checking?

7. Why is asking AI “Are you sure?” not independent verification?

8. How can a real source fail to support a claim?

9. Why should a correction include its effect on the wider answer?

10. What does calibrated trust mean in one task from your own life?

Key vocabulary

Hallucination:
Generated information that is fabricated or presented as factual without adequate grounding, often in a plausible form.

False confidence:
Language that expresses more certainty than the evidence supports.

Checkable claim:
A statement that can be compared with evidence, such as a name, number, date, quotation, causal claim, or rule.

Fabrication:
Invented information, evidence, quotations, citations, or sources.

Source mismatch:
A situation where a real source does not support the claim attributed to it.

Ungrounded summary:
A summary that attributes information to a source even though the information is absent from that source.

Context error:
Information applied to the wrong country, population, time, purpose, or situation.

Independent verification:
Checking a claim outside the chatbot using appropriate evidence, tools, original material, or qualified people.

Correction log:
A record of a claim, possible error, evidence, correction, effect, and final decision.

Calibrated trust:
Reliance adjusted to the available evidence, task, risk, and possible consequences.

Summary

In Week 11, we learned that fluent AI language is not the same as reliable knowledge.

A hallucination is fabricated or inadequately grounded information presented as factual. But not every error is the same. AI outputs may also contain factual mistakes, source mismatches, outdated information, context errors, calculation problems, or contradictions.

The responsible process is:

Mark the claim → identify the possible error → prioritise the risk → check outside the chatbot → correct transparently → adjust trust

The goal is not to distrust every AI response.

The goal is calibrated trust: relying on AI only to the degree justified by evidence, context, task, and consequence.

Lesson 11 Interactive Quiz: Hallucinations and False Confidence

Choose one answer for each question. Then select Check my answers. You will receive feedback for every question. This practice quiz does not collect names or scores.

1. AI confidently cites a detailed government report that does not exist. What type of problem is this?

A. A context error B. Outdated information C. A fabrication or hallucination D. A formatting problem

2. AI accurately describes a rule from 2022, but the rule was replaced in 2025. How should the error be classified?

A. Outdated information B. A fabricated source C. An ungrounded summary D. An internal contradiction

3. AI cites a real scientific article, but the article does not make the conclusion attributed to it. What is this?

A. Independent verification B. A calculation error C. A spelling error in the citation D. A source mismatch

4. An AI system used web search and included links. What can the user conclude?

A. Every claim is verified because web search was used. B. Tool use may improve grounding, but important claims and sources still need checking. C. Only claims without links require checking. D. Search results remove the possibility of context errors.

5. AI says, “I may be mistaken, but there is an 85 per cent chance this is correct.” How should this wording be treated?

A. As reliable because the AI acknowledged uncertainty. B. As verified because a precise percentage was provided. C. As a style signal that still requires supporting evidence. D. As false because cautious wording always indicates weak knowledge.

6. Which claim should receive the highest verification priority?

A. “Swedish law currently requires this action before an employee can be dismissed.” B. “Blue is a calming colour for many people.” C. “This fictional character could be named Elian.” D. “Bullet points may make this list easier to scan.”

7. What is the strongest next step after identifying a high-priority factual claim?

A. Ask the same AI to answer again using a more careful tone. B. Ask a second chatbot and accept the claim if both systems agree. C. Trust the claim if several search snippets repeat it. D. Open an appropriate source and check whether it supports the exact claim.

8. AI accurately describes a workplace rule from another country and applies it directly to Sweden. What type of error is this?

A. An internal contradiction B. A context error C. A fabricated quotation D. A harmless language variation

9. A fabricated study was the main support for an AI recommendation. What should the correction include?

A. Remove the study title but keep the recommendation unchanged. B. Replace the citation with a general phrase such as “research suggests.” C. Correct the source problem and reconsider the recommendation, reasoning, and confidence. D. Hide the error so readers are not distracted by the correction process.

10. Which example best demonstrates calibrated trust?

A. Use AI ideas freely for fiction, review study questions, and require authoritative evidence and human judgement for high-stakes claims. B. Trust every response that includes citations and professional language. C. Reject every AI output because some responses contain errors. D. Use the same level of trust for brainstorming and legal decisions.

Reflection: Choose one confident AI claim. What type of error could it contain, how important would the error be, and what evidence would you use to check it?