COMP 3270 / Module 4 / Unit 10 / Lesson 1

Lesson 10.1 · Week 7 · Reading: Chapter 7

Claims, Hallucinations, and Fabricated Content

Identify and separate the claims in AI-generated content. Distinguish hallucination from fabrication and factual error. Recognise fabricated references, quotations, and statistics. Explain why fluent, well-formatted text does not establish reliability.

Estimated study time: 3–4 hours Prerequisite: Units 1–9 Module 4 · Week 7

Table of Contents

1. Lesson Overview

This lesson is the first of four in Unit 10, Hallucination, Error, and Uncertainty. It establishes the foundational skill of claim identification—the ability to separate an AI-generated response into individual statements that can be examined independently. You will learn to distinguish claims from opinions, instructions, questions, and recommendations; to decompose compound claims; and to define hallucination with precision. You will also learn to distinguish fabrication from factual error and to recognise the warning signs of fabricated references, quotations, events, and statistics. Finally, you will examine why fluent, well-formatted language can create an illusion of credibility, and why that illusion must never be mistaken for evidence.

By the end of this lesson, you will be able to dissect any AI-generated passage into its constituent claims, flag potential hallucinations and fabrications, and identify the claims that most urgently require verification. This work lays the foundation for Lessons 10.2, 10.3, and 10.4, and for the detailed verification methods you will learn in Unit 11.

⚖️ Verification reminder: An AI system is not an independent authority for verifying its own output. Always apply human critical judgment. All examples in this lesson are synthetic—do not rely on any specific claim, statistic, or source reference without independent verification.

2. Learning Outcomes

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

  1. Define claim, hallucination, fabrication, and factual error with precision.
  2. Distinguish factual claims from opinions, recommendations, questions, instructions, interpretations, and value judgments.
  3. Decompose a compound AI-generated response into independently assessable atomic claims.
  4. Distinguish hallucination from fabrication, factual error, and unsupported claim.
  5. Identify fabricated sources, citations, quotations, events, statistics, credentials, and technical details in synthetic AI outputs.
  6. Recognise warning signs in citation-like references and quotation-like passages.
  7. Explain why fluent language, detailed formatting, and citation-like structures do not establish reliability or truth.
  8. Apply the first phase of the TRACE framework: identifying claims and error signals.

3. Before You Begin

4. Retrieval Practice

These questions draw on Units 2, 3, 6, 8, and 9. Answer them briefly, then check the feedback.

Question 1: In your own words, what is the “AI Paradox” as introduced in Unit 1, and why does it matter for evaluating AI-generated content?

Show feedback

Sample answer: The AI Paradox is the observation that generative AI systems can produce fluent, coherent, and seemingly authoritative content while simultaneously lacking genuine understanding, fact-checking capacity, or reliable access to truth. They are remarkably good at sounding correct, but they have no built-in mechanism for ensuring that what they produce corresponds to reality. This matters because it means we cannot rely on fluency or polish as a proxy for accuracy; we must independently verify every consequential claim.

Key point: Fluency ≠ reliability.

Question 2: From Unit 3, explain the distinction between fluency and factual reliability in the context of generative AI. Why is a fluent, well-structured response not necessarily reliable?

Show feedback

Sample answer: Fluency refers to the linguistic smoothness, coherence, and stylistic polish of the output—how naturally it reads. Factual reliability refers to whether the claims in the output are supported by evidence and correspond to reality. A generative AI system is trained to produce fluent text that matches patterns in its training data; it is not trained to verify the truth of those patterns. Consequently, a fluent response may contain fabricated details, unsupported claims, or subtle errors presented with high confidence. Fluency is a feature of the language model, not a certificate of accuracy.

Key point: Evaluate claims, not prose.

Question 3: From Units 6 and 9, what is meant by “human authority” in the context of AI-assisted work, and what does it mean to “escalate” a claim or output?

Show feedback

Sample answer: Human authority means that a qualified, accountable person—not the AI system—must make final decisions about the validity, appropriateness, and use of AI-generated content, especially when the content has consequences for learning, professional practice, or public life. To “escalate” means to refer a claim, output, or uncertainty to a person with greater expertise, authority, or responsibility when the issue falls outside one’s own competence, when evidence is conflicting or insufficient, or when the potential consequences of error are high.

Key point: AI assists; humans decide and are accountable.

Question 4: Briefly recall the purpose of the PCAFE-V and SCOPE frameworks from earlier units. How do they relate to the work of claim identification?

Show feedback

Sample answer: PCAFE-V (Purpose, Context, Audience, Form, Evidence, and Verification) is a framework for evaluating the suitability and risk of using generative AI in a given task. SCOPE (Source, Claim, Omission, Purpose, Evidence) helps structure the initial review of an AI-generated output. Both frameworks require the user to identify what is being claimed, for whom, and with what evidence. Claim identification is the natural first step in both frameworks: you cannot evaluate evidence, omissions, or context until you have identified the specific claims the output contains.

Key point: Claim identification is foundational to all other evaluation frameworks.

5. Opening Scenario: A Synthetic Academic Passage

Read the passage below carefully. It is a synthetic, AI-generated extract from a simulated student essay on the history of educational technology. The passage contains multiple kinds of defects, including a fabricated source, an invented quotation, an unsupported statistic, and an ambiguous recommendation. Do not assume any part of it is true. After reading, consider the questions that follow.

📌 Passage — “The Role of Digital Platforms in 21st-Century Learning” (synthetic extract)

According to a 2022 study by Martinez and Kowalski published in the Journal of Educational Technology, “the integration of adaptive learning platforms has been shown to improve student outcomes in STEM disciplines by an average of 34% across 15 controlled trials” (Martinez & Kowalski, 2022, p. 112). This finding is consistent with the broader shift toward data-driven instruction, which has been endorsed by the International Society for Technology in Education (ISTE) as a “cornerstone of modern pedagogy” (ISTE, 2021). Furthermore, a meta-analysis conducted by the Learning Analytics Research Group at Stanford University concluded that students who used AI-powered tutoring systems scored 28% higher on standardised assessments than those who received traditional instruction (Chen, 2020).

Given these compelling results, educators should consider integrating adaptive platforms into all core curriculum areas as soon as possible. However, it is equally important to ensure that teachers receive adequate training in interpreting the analytics generated by these systems. The evidence suggests that the benefits of adaptive technology are most pronounced when implementation is accompanied by professional development and ongoing technical support.

Note: The names, publication details, and statistics in this passage are synthetic. Do not rely on them as factual.

Initial Questions

  1. What claims does this passage make? List as many as you can.
  2. Which statements appear to be factual claims? Which are interpretive, predictive, or recommendations?
  3. What evidence would you need to verify the claimed 34% improvement?
  4. What warning signs do you notice in the references and quotations?
Show guided analysis (initial)

Initial observations:

  • The passage makes several factual-sounding claims: a 34% improvement, a 28% higher score, the existence of a 2022 study by Martinez and Kowalski, a 2021 ISTE statement, and a 2020 meta-analysis by Chen.
  • It also contains recommendations (“educators should consider integrating…”), interpretive claims (“consistent with the broader shift…”), and a conditional claim about training.
  • The references look plausible (journal name, author names, dates, page numbers) but may be fabricated. The quotation marks around the 34% finding suggest a direct quote, but we cannot verify it without checking the source.
  • The statistic “34% across 15 controlled trials” is highly specific and may be false precision or entirely fabricated.
  • The passage does not cite a primary source for the ISTE endorsement; it only mentions the organisation and a year.

Key takeaway: A passage can be fluent, well-structured, and packed with plausible details—and still contain serious errors or fabrications. The first step is to separate every claim and then evaluate each one independently.

6. Core Lesson Content

6.1 What Is a Claim?

Definition — Claim

A claim is a statement that presents something as true, probable, likely, appropriate, or supported. A claim is an assertion that can be examined, questioned, verified, or refuted. Not every sentence contains a claim; questions, commands, and exclamations generally do not, though they may imply claims.

In the context of generative AI, a claim is any proposition that the system has generated that invites acceptance, belief, or action. Claims may be:

  • Factual: “The population of Canada in 2021 was approximately 38 million.”
  • Quantitative: “The intervention improved test scores by 12%.”
  • Causal: “The policy caused a decline in unemployment.”
  • Comparative: “Model A outperforms Model B on benchmark X.”
  • Interpretive: “The results suggest a shift in consumer preferences.”
  • Predictive: “The system will reduce error rates by 20% next year.”
  • Evaluative: “This approach is more effective than the alternative.”
  • Normative (value-based): “Educational institutions should adopt adaptive learning technologies.”
  • Legal or regulatory: “The policy requires annual reporting.”
  • Professional: “In clinical practice, this guideline is considered standard.”
  • Historical: “The treaty was signed in 1945.”
  • Technical: “The system uses a transformer-based architecture.”

Importantly, a single sentence may contain multiple claims. For example:

“The 2022 study by Martinez and Kowalski demonstrated that adaptive learning platforms improve STEM outcomes by 34%, and this finding is consistent with previous research on personalised instruction.”

This sentence contains at least four distinct claims:

  1. A 2022 study by Martinez and Kowalski exists.
  2. That study demonstrated an improvement.
  3. The improvement was 34%.
  4. The finding is consistent with previous research.

Why this matters: If you treat the whole sentence as a single claim, you may miss that one part is fabricated while another is accurate. Decomposition is the first step in responsible error analysis.

🧩 Quick check

Consider the statement: “The committee recommended, based on the 2021 survey, that the policy should be revised.”

How many claims can you identify? List them.

Show possible decomposition
  • A committee exists (or existed).
  • The committee made a recommendation.
  • The recommendation was based on a survey.
  • The survey was conducted in 2021.
  • The content of the recommendation is that the policy should be revised.

Note: The first four claims are factual-sounding; the fifth is a normative claim (a recommendation). Each requires different evidence.

6.2 Atomic and Compound Claims

Definitions — Atomic and Compound Claims

Atomic claim: A claim containing one principal proposition that can be examined independently.

Compound claim: A statement containing several propositions that may require different evidence, each of which can be separated into atomic claims.

For example:

  • Atomic: “The report was published in 2023.” (One proposition.)
  • Atomic: “The report was authored by Dr. Singh.” (One proposition.)
  • Compound: “Dr. Singh’s 2023 report recommends a 15% budget increase.” (Contains: the report exists, it was published in 2023, Dr. Singh is the author, it recommends a 15% increase.)

Practice: When you encounter a compound claim, separate it into its atomic parts. Then evaluate each atomic claim independently. This prevents a single fabricated detail from contaminating your assessment of the entire passage.

Error-analysis heuristic

Before you ask “Is this statement true?”, ask: “What exactly is being claimed?” Write down each atomic claim. Only then begin to verify.

6.3 Hallucination Defined

Definition — Hallucination (in generative AI)

[INSERT DEFINITION FROM CHAPTER 7]

At a general level, treat hallucination as generated content that is presented plausibly but lacks adequate grounding or alignment with reality, supplied evidence, or authoritative sources.

Important: Do not use “hallucination” as a label for every undesirable output. It is a specific category of error, distinct from fabrication, factual error, omission, and uncertainty. Do not anthropomorphise the system—hallucination is a pattern in the output, not a mental state.

Key characteristics of hallucination:

  • The content is plausible—it looks and sounds like legitimate information.
  • It is ungrounded—it does not correspond to verifiable reality.
  • It is presented with confidence, often without hedging or qualification.
  • It may be internally coherent, making it more deceptive.

Example (synthetic): An AI system generates a response stating that “the 2020 Global Education Report by UNESCO found that 72% of students preferred digital textbooks over print.” If no such report exists, or if the report says something different, this is a hallucination.

Misconception: Hallucination is the same as “lying.” It is not. An AI system does not have intent, belief, or knowledge of its own falsehood. Hallucination is a systemic artefact of how generative models produce text from patterns in training data.

Why it matters: Hallucinations can appear in any domain—education, medicine, law, finance, science, journalism—and they can be extremely difficult to detect if you rely only on the internal coherence of the text.

6.4 Fabrication vs. Factual Error

Definitions — Fabrication and Factual Error

Fabrication: Generating details that are represented as real but are invented or unsupported. Fabrication often involves specific, concrete items that can be checked: sources, quotations, events, statistics, credentials, laws, reports, and technical specifications.

Factual error: A claim that conflicts with reliable evidence—a wrong date, an incorrect name, a false numerical value, an inaccurate description. A factual error need not be intentionally deceptive; it is simply incorrect.

Key distinction: A fabrication is an invention presented as real; a factual error is a misstatement of reality. Both are problems, but they require slightly different responses. A fabrication may be wholly invented (e.g., a nonexistent source), while a factual error may be a misremembered or incorrectly generated detail (e.g., a date that is off by one year).

Critical warning: Generative AI systems are particularly prone to fabricating sources, citations, and quotations because they have been trained on patterns of scholarly writing without any connection to the actual sources. They know what a citation looks like, but they do not know whether the cited work exists or whether the quoted text is accurate.

Examples of fabrication (synthetic):

  • A reference to “Smith, J. (2021). Foundations of AI Ethics. Cambridge University Press” that does not exist.
  • A quotation attributed to a well-known figure that they never said.
  • A statistic such as “87% of practitioners report satisfaction” with no source.
  • A claim that “the Canadian AI Act of 2022 requires annual audits” when no such act exists.
  • A credential like “Dr. A. B. Chen, senior researcher at the Stanford AI Lab” for a person who does not exist.

Examples of factual error (synthetic):

  • Stating that the UN was founded in 1946 (it was 1945).
  • Attributing a quote to the wrong author.
  • Using the wrong date for a known event.
  • Incorrectly describing the findings of a real study.

Overlap: A fabrication may also contain factual errors (e.g., an invented source with an incorrect date). The categories are not mutually exclusive.

6.5 Fabricated References and Quotations

Fabricated references are among the most common and consequential errors in AI-generated content. They are dangerous because they look legitimate—they follow the formatting conventions of real citations—and they can mislead readers into believing that a claim is supported by authoritative evidence.

Key rule: Citation-like formatting is not proof that a source exists. Always verify the source independently.

Warning signs of fabricated references:

  • Nonexistent publication: The journal, book, or report cannot be found in any catalogue or database.
  • Mismatched author and title: The author’s known areas of expertise do not align with the topic.
  • Implausible journal: The journal name sounds credible but does not exist, or the article is not in that journal.
  • Incorrect date: The date is inconsistent with the author’s career, the topic’s emergence, or the publication history.
  • Malformed identifier: The DOI, ISBN, or other identifier is invalid, does not resolve, or points to a different work.
  • Source does not support the attributed claim: Even if the source exists, the specific claim may not appear in it.

Fabricated quotations:

  • Quotation marks around text that the source never contained.
  • A page number that does not exist in the source.
  • A quotation that is a paraphrase presented as a direct quote.
  • A quotation attributed to an author or organisation with no verifiable source.
🧩 Quick check: suspicious reference

Consider this reference: “Thompson, L. (2019). AI and the Future of Work. Oxford University Press, p. 87.”

What steps would you take to verify it?

Show suggested verification steps
  1. Check Oxford University Press’s catalogue for the title.
  2. Search for the author and title in a library database or scholarly search engine.
  3. Look for the book in WorldCat or a similar catalogue.
  4. Check the author’s institutional profile or publication list.
  5. If the book exists, locate the specific passage on p. 87 and verify the claim.

Remember: Formatting alone proves nothing. Verification requires actual evidence.

6.6 Fluency, Formatting, and False Credibility

Generative AI systems produce text that is often fluent, well-organised, and stylistically polished. They can generate headings, tables, bullet points, and citation-like references that mimic the structure of authoritative documents. This creates a powerful illusion of credibility.

Critical insight: Fluency, detail, and structural polish are features of the language model, not indicators of truth or reliability. A passage can be perfectly formatted, richly detailed, and entirely fabricated.

Features that may create false credibility:

  • Confident tone: Absence of hedging or qualification may suggest certainty, but it is actually a stylistic choice.
  • Detailed wording: Specific names, dates, and numbers can feel authoritative, but they may be invented.
  • Headings and structure: Professional-looking organisation can signal expertise, but it is not evidence.
  • Tables and figures: Visual representations of data can be fabricated along with the data.
  • Statistics and percentages: Numbers give an impression of precision, but they may be false precision or invention.
  • Citation-like references: As discussed, these can be completely fabricated.
  • Technical vocabulary: The correct use of domain-specific terms can suggest competence, but it does not guarantee accuracy.
  • Balanced structure: Presenting pros and cons can make a passage seem objective, but the content may still be flawed.

Why this matters: When you read an AI-generated passage, you must consciously override the natural tendency to trust fluent, well-presented text. Do not mistake style for substance.

Adopt a verification mindset

Every time you encounter a claim, ask: “What is the evidence for this claim, and where did it come from?” If the passage itself is the only source of the evidence, you have not yet verified anything.

7. TRACE Framework: Introductory Walkthrough

The TRACE framework is the central error-analysis and verification tool for Unit 10. In Lesson 10.1, we focus on the first element: identifying claims and error signals. Later lessons cover classification, consequence, correction, and communication.

Placeholder: [INSERT AUTHOR-APPROVED TRACE EXPANSION]

Until the exact author-approved expansion is confirmed, we use TRACE operationally as a framework addressing:

Note: The exact element names and definitions are subject to author confirmation. This operational version aligns with the requirements of Unit 10 and prepares you for the precise framework used in Unit 11 and Assignment 5.

TRACE Element 1: Claim and Error-Signal Identification

Element name: [INSERT AUTHOR-APPROVED NAME FOR THE FIRST TRACE ELEMENT]

Operational definition: The process of separating an AI-generated response into its constituent claims and identifying signals that suggest potential error.

Guiding questions:

Example (synthetic): In the opening scenario passage, the first sentence contains a compound claim: a 2022 study exists, it was by Martinez and Kowalski, it was published in the Journal of Educational Technology, it showed a 34% improvement, and it involved 15 controlled trials. Each of these is an atomic claim. The phrase “according to” signals that a source is being invoked—a prime target for verification.

Common mistake: Treating the entire passage as one claim, or overlooking the implicit claims embedded in quotations and references.

Required evidence: The output text itself, plus any available context about its generation (prompt, purpose, audience).

Possible correction: Mark each atomic claim for verification. Separate claims that require different types of evidence (e.g., source existence vs. numerical accuracy).

Escalation or stop condition: If the passage contains multiple fabricated sources or central claims that cannot be verified, escalate to a subject-matter expert or stop relying on the output.

Residual limitation: Identifying claims is the first step, but it does not tell you which claims are true or false. It only tells you what needs to be checked.

🧩 TRACE practice: claim identification

Revisit the opening scenario passage. Using the TRACE claim-identification element, list every atomic claim you can find. Then identify at least three claims that signal potential fabrication (e.g., suspicious references, overly precise numbers).

Show sample claim inventory (partial)

Atomic claims (selected):

  • A 2022 study by Martinez and Kowalski exists.
  • That study was published in the Journal of Educational Technology.
  • The study showed a 34% improvement in STEM outcomes.
  • The study involved 15 controlled trials.
  • The quotation about STEM outcomes is attributed to Martinez and Kowalski (2022, p. 112).
  • ISTE endorsed data-driven instruction in 2021.
  • ISTE described it as a “cornerstone of modern pedagogy.”
  • A meta-analysis by Chen (2020) exists.
  • That meta-analysis concluded that AI-powered tutoring systems improved scores by 28%.
  • The meta-analysis was conducted by the Learning Analytics Research Group at Stanford University.

Signal claims (potential fabrication):

  • The precise 34% figure with 15 trials—highly specific and un-sourced beyond the claim itself.
  • The attribution to Martinez and Kowalski (2022) with a page number—needs source verification.
  • The ISTE quotation—requires checking ISTE’s 2021 publications.
  • The Chen (2020) meta-analysis—requires verifying the existence of that study and the 28% figure.

Note: This is a partial inventory. A complete audit would identify every atomic claim and flag each for verification.

8. Error-Analysis Process Walkthrough: Passage to Claim Inventory to Hallucination and Fabrication Screening

This walkthrough demonstrates the first stage of error analysis for a synthetic AI-generated passage. The goal is to move from a raw output to a structured inventory of claims and a preliminary screen for hallucinations and fabrications.

  1. Stage 1: Read and annotate. Read the passage through once without attempting to evaluate truth. Highlight or underline every sentence that appears to make a claim, reference a source, or present a fact, interpretation, or recommendation.
  2. Stage 2: Separate atomic claims. For each sentence, break it into its atomic propositions. Write each atomic claim as a separate, self-contained statement.
  3. Stage 3: Classify each claim. For each atomic claim, note its type: factual, quantitative, causal, comparative, interpretive, predictive, evaluative, normative, legal, professional, historical, or technical. This helps determine what kind of evidence is needed.
  4. Stage 4: Flag possible hallucinations and fabrications. Look for warning signs: overly precise numbers, named sources that are not well-known, quotation marks around long or specific text, references that seem too perfect, and any claim that would be difficult to verify quickly.
  5. Stage 5: Determine verification priority. Prioritise claims that are central to the argument; consequential; surprising or counterintuitive; quantitative; current or time-sensitive; legal, policy-related, or professional; attributed to a named source; likely to be acted upon.
  6. Stage 6: Document and carry forward. Record your claim inventory, your provisional classifications, and your verification priorities.
Key insight

This process is not about getting the “right answer” immediately—it is about systematically separating what the output says from what you know to be true. The separation itself is the essential skill.

9. Worked Example: Analysing a Synthetic AI-Generated Paragraph

This worked example demonstrates the full claim-identification and hallucination/fabrication screening process on a synthetic paragraph about renewable energy policy.

📌 Passage — “Renewable Energy Policy in Canada” (synthetic extract)

According to a 2023 report by the Canadian Climate Institute, “federal investment in renewable energy infrastructure has reduced greenhouse gas emissions by 22% since 2018, while creating approximately 45,000 new jobs in the clean energy sector” (Canadian Climate Institute, 2023, p. 31). The report also notes that these investments have been most effective in provinces with existing hydroelectric capacity, such as British Columbia and Quebec. In contrast, a 2022 policy brief from the Fraser Institute argued that the economic costs of the transition have been underestimated, with an estimated loss of 12,000 jobs in the oil and gas sector (Fraser Institute, 2022). The Canadian government has stated that it remains committed to achieving net-zero emissions by 2050, a target that aligns with the Paris Agreement.

All names, organisations, statistics, and publication details are synthetic. Do not rely on them as factual.

Step-by-step analysis

1. Annotate and read

Passage contains several factual-sounding statements, a direct quotation, two named sources (Canadian Climate Institute, Fraser Institute), and a claim about government commitment.

2. Separate atomic claims

Total: 11 atomic claims.

3. Classify each claim

4. Flag possible hallucinations/fabrications

5. Determine verification priority

6. Preliminary assessment

This passage is structurally fluent and cites multiple sources, but it contains several high priority claims that are either fabricated or require verification. Do not accept the 22% reduction, the 45,000 jobs, or the 12,000 jobs figures without independent confirmation. The named sources may not exist, and even if they do, the numbers may be misrepresented.

Recommendation: Treat all quantitative claims and all source attributions as unverified until proven otherwise. Proceed to the verification stage (Unit 11) for definitive conclusions.

Show worked example summary table
Worked example: claim inventory and priority
Atomic claim Type Verification priority Possible fabrication?
2023 Canadian Climate Institute report exists Factual High Yes (source may be fabricated)
Report contains quotation Factual High Yes (quotation may be fabricated)
Emissions reduced by 22% since 2018 Quantitative High Yes (number may be fabricated)
45,000 new jobs created Quantitative High Yes (number may be fabricated)
Page 31 reference Factual Medium Yes (page may be fabricated)
Most effective in BC and Quebec Interpretive Medium Possible (depends on source)
Fraser Institute 2022 brief exists Factual High Yes (source may be fabricated)
Brief argued costs underestimated Interpretive Medium Possible (depends on source)
12,000 jobs lost in oil and gas Quantitative High Yes (number may be fabricated)
Government committed to net-zero 2050 Factual Low Unlikely (easily verifiable)
2050 target aligns with Paris Agreement Factual/Interpretive Low Unlikely (easily verifiable)

10. Guided Learning Activities

🧩 Activity 10.1.A: Claim Inventory
  • Purpose: Practice separating a synthetic AI-generated passage into atomic claims.
  • Learning outcome: Decompose compound claims; distinguish hallucination, fabrication, and factual error.
  • Estimated time: 20–30 minutes.

Materials: The opening scenario passage (above) or a supplied alternative.

Instructions:

  1. Read the passage carefully.
  2. Create a numbered list of every atomic claim you can identify.
  3. For each claim, note its type (factual, quantitative, interpretive, etc.).
  4. Identify any claims that appear to be compound and separate them.
  5. Flag any claims that you suspect may be hallucinations or fabrications, with a brief reason.
Show model claim inventory (opening scenario)

Opening scenario claim inventory (selected):

  1. A 2022 study by Martinez and Kowalski exists. (Factual — possible fabrication)
  2. The study was published in the Journal of Educational Technology. (Factual — possible fabrication)
  3. The study showed a 34% improvement in STEM outcomes. (Quantitative — possible fabrication)
  4. The study involved 15 controlled trials. (Quantitative — possible fabrication)
  5. The quotation about STEM outcomes is attributed to Martinez and Kowalski (2022, p. 112). (Factual — possible fabrication)
  6. ISTE endorsed data-driven instruction in 2021. (Factual — possible fabrication)
  7. ISTE described it as a “cornerstone of modern pedagogy.” (Factual/Interpretive — possible fabrication)
  8. A 2020 meta-analysis by Chen exists. (Factual — possible fabrication)
  9. The meta-analysis concluded that AI-powered tutoring systems improved scores by 28%. (Quantitative — possible fabrication)
  10. The meta-analysis was conducted by the Learning Analytics Research Group at Stanford University. (Factual — possible fabrication)
  11. Educators should consider integrating adaptive platforms into all core curriculum areas. (Normative/Recommendation)
  12. Teachers should receive adequate training in interpreting analytics. (Normative/Recommendation)
  13. The benefits of adaptive technology are most pronounced with professional development and support. (Interpretive)

Flags: Claims 1–10 all involve named sources, specific numbers, or quotations that cannot be verified from the passage itself. These are the highest-priority claims for verification.

🧩 Activity 10.1.B: Hallucination or Another Problem?
  • Purpose: Distinguish hallucination from fabrication, factual error, unsupported claim, omission, and uncertainty.
  • Learning outcome: Distinguish hallucination from fabrication, factual error, and unsupported claim.
  • Estimated time: 15–20 minutes.

Instructions: Classify each synthetic statement below. Choose one of: Hallucination, Fabrication, Factual Error, Unsupported Claim, Uncertain, or Insufficient Information. Explain your reasoning.

  1. “The 2021 World Health Organization report stated that coffee consumption reduces the risk of type 2 diabetes by 15%.” (No source is provided, and you have not checked the WHO report.)
  2. “The capital of Canada is Toronto.” (You know this is incorrect.)
  3. “According to Dr. Jane Smith of the University of X, the new treatment is effective in 89% of cases.” (You have not verified whether Dr. Smith or the study exists.)
  4. “The AI system generated a response claiming that the 2020 Olympic Games were held in Tokyo, but it did not mention that they were postponed to 2021 due to the pandemic.”
  5. “The system states that ‘the average temperature in July is 22°C’ without providing a source or data range.”
Show classification and reasoning
  1. Unsupported claim — The claim is plausible but lacks evidence. It is not demonstrably false, but without the source, it cannot be verified.
  2. Factual error — The statement is demonstrably false. Ottawa is the capital of Canada.
  3. Insufficient information — We do not know whether Dr. Smith, the university, or the study exist. This could be a fabrication, a factual error, or a supported claim. More investigation is needed.
  4. Omission — The statement is technically true (the Games were in Tokyo) but omits the crucial context that they were postponed to 2021. This is a material omission.
  5. Uncertain — Without a source or data range, we cannot determine if the claim is accurate. It is not necessarily false, but it is unverifiable as presented.

Key point: The same statement may be classified differently depending on what you know and what evidence is available. The categories are not always mutually exclusive.

🧩 Activity 10.1.C: Fabricated-Reference Warning Signs
  • Purpose: Practise identifying warning signs in citation-like references.
  • Learning outcome: Identify fabricated sources and recognise warning signs.
  • Estimated time: 15–20 minutes.

Instructions: Examine the following synthetic references. For each, list any warning signs that suggest it may be fabricated.

  1. “Thompson, P. (2020). AI and Ethics: A Practical Guide. Springer, p. 45.”
  2. “Kim, S.-H., & O’Brien, M. (2019). ‘The role of generative models in educational assessment.’ Journal of Learning Analytics, 12(3), 45–67.”
  3. “According to the 2021 UNESCO report, ‘digital literacy is a fundamental human right in the 21st century’ (UNESCO, 2021, p. 89).”
  4. “Chen, L. (2018). Foundations of Machine Learning. MIT Press, p. 212.”
  5. “International Energy Agency. (2022). Global Energy Review 2022. Paris: IEA Publications.”
Show warning signs
  1. Suspicious: The title is very generic; Springer is a real publisher, but the book may not exist. The page number is specific and needs verification.
  2. Suspicious: The volume and page numbers look plausible, but the journal name and volume/issue may not match. The authors and topic are plausible but need verification.
  3. Suspicious: The quotation is grand and generic (“fundamental human right”)—a red flag for fabrication. UNESCO reports are real, but this specific statement may not exist.
  4. Suspicious: MIT Press is real, and the title is plausible, but the book may not exist or may not contain that content on p. 212.
  5. Suspicious: The IEA does publish annual reviews, but this specific title and year may not exist, or the content may be misrepresented.

General warning: Every citation should be verified independently. The absence of obvious errors does not prove that a source exists.

🧩 Activity 10.1.D: Credibility-Feature Audit
  • Purpose: Identify features that create false credibility and explain why they do not prove accuracy.
  • Learning outcome: Explain why fluency, formatting, and citation-like structures do not establish reliability.
  • Estimated time: 15–20 minutes.

Instructions: Using the opening scenario passage (or another synthetic AI-generated passage), identify at least five features that make the passage appear credible. For each feature, explain why it does not actually prove that the content is accurate.

Show sample audit

Features that create false credibility:

  • Confident tone: The passage states facts without hedging (“has been shown to improve,” “concluded that”). Why this does not prove accuracy: Confidence is a stylistic choice, not a measure of truth.
  • Detailed references: The passage includes author names, publication dates, journal titles, and page numbers. Why this does not prove accuracy: These details can be entirely fabricated. They are textual patterns, not evidence of existence.
  • Specific statistics: The 34% and 28% figures. Why this does not prove accuracy: Numbers can be invented or misrepresented. Precision does not equal truth.
  • Quotation marks: The passage uses quotation marks around a statement attributed to Martinez and Kowalski. Why this does not prove accuracy: Quotation marks indicate a direct quote, but the quote may be fabricated or taken out of context.
  • Organisational endorsements: The passage mentions ISTE and Stanford University. Why this does not prove accuracy: Named organisations can be used to lend authority, but the specific claims may not be endorsed by those organisations.
  • Professional language: The passage uses terms like “adaptive learning platforms,” “data-driven instruction,” and “meta-analysis.” Why this does not prove accuracy: Technical vocabulary indicates familiarity with a field, not correctness of the specific claims.

Key takeaway: These features are indicators of possible credibility, but they are not evidence. Verification requires independent confirmation from authoritative sources.

11. Exercises

📝 Exercise 10.1.1: Distinguish Claims from Non-Claims
  • Difficulty: Introductory
  • Estimated time: ~10 min
  • Learning outcome: Distinguish claims from non-claims.

Task: For each of the following statements, indicate whether it is a claim (an assertion that can be examined) or a non-claim (question, command, exclamation, or pure opinion without assertoric content). If it is a claim, note its type (factual, quantitative, interpretive, etc.).

  1. “The experiment was conducted in 2022.”
  2. “Could the experiment be replicated?”
  3. “Please review the data for errors.”
  4. “The results suggest a correlation between the variables.”
  5. “Wow, that is an impressive finding!”
  6. “In my opinion, the policy should be revised.”
  7. “The policy increased compliance by 15% according to the report.”
Show answers and explanations
  1. Claim (factual). It asserts a fact about when the experiment was conducted.
  2. Non-claim (question). It asks a question; it does not assert anything.
  3. Non-claim (command/request). It instructs someone to do something.
  4. Claim (interpretive). It asserts that the results suggest a correlation—an interpretation of evidence.
  5. Non-claim (exclamation). It expresses emotion, not an assertion.
  6. Claim (normative/opinion). Even though it begins with “in my opinion,” it still makes an assertoric claim about what should happen.
  7. Claim (quantitative and factual). It asserts a 15% increase and cites a report.
📝 Exercise 10.1.2: Separate Compound Claims
  • Difficulty: Introductory
  • Estimated time: ~10 min
  • Learning outcome: Decompose compound claims into atomic claims.

Task: For each compound statement, separate it into its atomic claims. List each atomic claim as a separate numbered item.

  1. “The 2020 study by Dr. Patel, published in the Journal of Medical Ethics, found that the new protocol reduced patient wait times by 18%.”
  2. “According to the 2019 OECD report, Canada’s investment in early childhood education has increased by 12% since 2015, and this has been associated with improved literacy outcomes.”
  3. “The committee recommended, based on the 2021 survey data, that the university should implement a hybrid teaching model for the 2023 academic year.”
Show sample decompositions
    • A 2020 study by Dr. Patel exists.
    • The study was published in the Journal of Medical Ethics.
    • The study found that the new protocol reduced patient wait times.
    • The reduction was 18%.
    • A 2019 OECD report exists.
    • The report states that Canada’s investment in early childhood education has increased.
    • The increase is 12%.
    • The increase has occurred since 2015.
    • The increase has been associated with improved literacy outcomes.
    • A committee exists.
    • The committee made a recommendation.
    • The recommendation was based on 2021 survey data.
    • The survey data exists.
    • The recommendation is that the university should implement a hybrid teaching model.
    • The implementation would be for the 2023 academic year.

Note: Some claims are implicit (e.g., “the survey data exists”). It is often useful to surface these implicit claims to ensure they are examined.

📝 Exercise 10.1.3: Classify Hallucination and Fabrication
  • Difficulty: Intermediate
  • Estimated time: ~10 min
  • Learning outcome: Distinguish hallucination from fabrication, factual error, and unsupported claim.

Task: Classify each synthetic scenario as Hallucination, Fabrication, Factual Error, or Unsupported Claim. Briefly explain your reasoning.

  1. An AI system generates a citation to a book that does not exist, along with a quotation from that book.
  2. An AI system states that the Earth is flat, when it is not.
  3. An AI system claims that “studies show” a particular treatment is effective, but no studies are cited and none can be found.
  4. An AI system generates a reference to a real journal article but attributes it to the wrong author.
Show classification and reasoning
  1. Fabrication — The book and quotation are invented and presented as real. This is a clear fabrication.
  2. Factual error — The claim is demonstrably false. It is incorrect, not invented.
  3. Unsupported claim — The claim may be true or false, but no evidence is provided. It is unsupported, not necessarily fabricated.
  4. Factual error — The journal article exists, but the attribution is wrong. This is an error in the representation of a real source.
📝 Exercise 10.1.4: Identify Fabricated Quotation Risks
  • Difficulty: Intermediate
  • Estimated time: ~10 min
  • Learning outcome: Identify fabricated sources and recognise warning signs.

Task: The following quotation appears in a synthetic AI-generated passage. Identify at least three reasons to suspect it may be fabricated.

“According to Dr. Emily Watson of the Harvard School of Public Health, ‘the correlation between air pollution and respiratory disease is now so well established that policy inaction constitutes a public health emergency’ (Watson, 2021, p. 47).”
Show warning signs
  • Dramatic language: “policy inaction constitutes a public health emergency” is strong and attention-grabbing—a common feature of fabricated quotes.
  • Vague source: “Dr. Emily Watson of the Harvard School of Public Health” is specific enough to sound authoritative, but the reader has no way to verify that this person exists or said this.
  • Precise page number: The page number (p. 47) adds a veneer of authenticity, but it may be entirely invented.
  • Lack of context: There is no indication of which document or publication the quote is from—only the author and year.
  • Timing: “2021” is recent enough to seem current, but the reader cannot check the source.

Key point: A single quotation can contain multiple fabricated elements. Verify the existence of the author, the publication, and the specific wording.

📝 Exercise 10.1.5: Identify Suspicious Citation Features
  • Difficulty: Intermediate
  • Estimated time: ~10 min
  • Learning outcome: Identify fabricated sources and recognise warning signs.

Task: Examine the following synthetic citation. List any features that make it suspicious and explain why they are warning signs.

“Martinez, A., & Lee, S. (2022). ‘The future of AI in education: A meta-analysis of 50 studies.’ Journal of Educational Technology Research, 45(2), 112–130.”

Show warning signs
  • Broad, attractive title: “The future of AI in education” is a very general, attention-grabbing title—common in fabricated citations.
  • “Meta-analysis of 50 studies”: This is a very specific claim that would require significant effort to verify. It is the kind of detail that can be invented to lend authority.
  • Journal name: The Journal of Educational Technology Research sounds plausible, but the reader should verify its existence and whether this article appears in it.
  • Volume and page numbers: 45(2), 112–130 is a plausible citation format, but the numbers may be invented. Check if the journal actually has volume 45, issue 2, and pages 112–130.
  • Authors: Martinez and Lee are common surnames—easy to invent. Verify their institutional affiliations and publication history.

Key point: A citation can look perfect and still be fabricated. Verification is the only reliable test.

📝 Exercise 10.1.6: Explain Why Fluency Does Not Establish Reliability
  • Difficulty: Intermediate
  • Estimated time: ~10 min
  • Learning outcome: Explain why fluency, formatting, and citation-like structures do not establish reliability.

Task: Write a short paragraph (4–6 sentences) explaining why a fluent, well-organised AI-generated passage with detailed references and precise statistics is not necessarily reliable. Use your own words, and include at least two specific examples of features that may create false credibility.

Show sample explanation

Sample answer: Fluency and organisation are features of the language model’s training, not markers of truth. A generative AI system can produce a passage that is grammatically flawless, logically structured, and filled with specific names, dates, and statistics—all of which may be entirely fabricated. For example, a citation with a precise page number and journal name may refer to a non-existent source, and a statistic such as “34% improvement” may be invented to add apparent precision. These features exploit our natural tendency to trust well-presented information, but they are stylistic, not evidentiary. Responsible users must always verify the underlying claims, regardless of how polished the presentation appears.

12. Homework Problems

Formative · Ungraded · Recommended preparation for Assignment 5 and for Unit 11.

🧠 Problem 10.1.1: Claim Identification

Task: Identify the claims in the following sentence. List each atomic claim separately.

“The 2021 survey conducted by the Canadian Digital Literacy Network found that 67% of respondents aged 18–24 reported using generative AI tools for academic purposes, and this finding is consistent with earlier research on technology adoption among younger demographics.”

Show sample answer
  • A 2021 survey was conducted by the Canadian Digital Literacy Network.
  • 67% of respondents aged 18–24 reported using generative AI tools for academic purposes.
  • This finding is consistent with earlier research on technology adoption among younger demographics.
🧠 Problem 10.1.2: Atomic vs. Compound Claims

Task: For the sentence below, identify which parts are atomic claims and which parts make it a compound claim. Then separate the atomic claims.

“Dr. A. B. Chen’s 2020 textbook, AI for Everyone, published by Oxford University Press, argues that AI literacy should be a core component of secondary education.”

Show sample decomposition

Atomic claims:

  • Dr. A. B. Chen exists (or existed).
  • Dr. Chen wrote a textbook in 2020.
  • The textbook is titled AI for Everyone.
  • The textbook was published by Oxford University Press.
  • The textbook argues that AI literacy should be a core component of secondary education.

Compound nature: The sentence packages all these claims into one syntactic unit. Each atomic claim requires separate evidence.

🧠 Problem 10.1.3: Hallucination Classification

Task: Classify the following scenario. Is it a hallucination, fabrication, factual error, or unsupported claim? Explain.

“An AI system generates a response that cites a 2022 report from the International Monetary Fund (IMF) stating that global GDP growth will be 3.5% in 2024. The IMF did release a 2022 report, but it does not contain this specific forecast.”

Show classification and reasoning

Classification: Fabrication (or, more specifically, a fabricated detail embedded in an otherwise real source). The IMF report exists, but the specific claim (the 3.5% forecast) is invented and attributed to the real source. This is a fabrication because the detail is represented as real but is not present in the source.

Alternative view: It could also be considered a hallucination, since the system generated content that is ungrounded. The key is that the source exists but the claim is misattributed.

🧠 Problem 10.1.4: Fabrication Identification

Task: In the following synthetic passage, identify at least one likely fabrication and explain why you suspect it.

“According to a 2023 study by Dr. Maria Santos and Prof. James Okafor, published in the Journal of Global Health, the implementation of community-based health programmes in sub-Saharan Africa reduced child mortality by 42% between 2015 and 2022. The study, which involved 30 villages across five countries, concluded that the programmes were highly effective.”

Show likely fabrication

Likely fabrication: The specific statistic “42% reduction” across “30 villages across five countries” is highly precise and would require significant evidence. The named authors, journal, and publication year are all plausible but unverified. The most suspicious element is the precise percentage combined with the detailed scope—this is a common pattern of fabrication.

Other concerns: The Journal of Global Health may or may not exist. Dr. Santos and Prof. Okafor may or may not be real researchers. Each of these details must be verified.

🧠 Problem 10.1.5: Reference Warning Signs

Task: Examine this synthetic reference: “World Health Organization. (2020). Global Health Statistics 2020. Geneva: WHO Press, p. 104.” List three warning signs that should trigger verification.

Show warning signs
  • Generic title: Global Health Statistics 2020 is a very broad title; WHO does publish health statistics, but this specific publication may not exist.
  • Precise page number: p. 104 is specific but may be invented. If the source exists, the page must be checked.
  • Lack of ISBN or DOI: The reference does not include an identifier that would make verification easy.
  • Common pattern: WHO is a real organisation, and publishing statistics is plausible, but this is a common pattern for fabrication—real organisation + plausible title + invented details.
🧠 Problem 10.1.6: Quotation Verification

Task: The following quotation appears in an AI-generated passage. What steps would you take to verify it?

“As Dr. K. Nakamura stated in his 2019 book The Ethics of AI, ‘We must ensure that AI systems are transparent, accountable, and aligned with human values’ (Nakamura, 2019, p. 78).”

Show verification steps
  1. Search for Dr. K. Nakamura and the book The Ethics of AI in library catalogues, publisher catalogues, and scholarly databases.
  2. If the book exists, locate it and verify the quotation on p. 78.
  3. Check if the quotation is actually present and whether the wording matches.
  4. If the book does not exist or the quotation is not present, the quotation is likely fabricated.
  5. Consider whether the quotation is paraphrased or taken out of context.
🧠 Problem 10.1.7: Credibility-Feature Analysis

Task: The following sentence appears in an AI-generated response. Identify three features that might make it seem credible, and explain why each feature does not prove the claim is true.

“According to a comprehensive 2023 meta-analysis published in the Journal of Educational Psychology, active learning strategies improve student performance by an average of 22% across 45 studies, with the strongest effects observed in STEM disciplines.”

Show credibility features and limitations
  • Feature: “comprehensive 2023 meta-analysis” — This sounds rigorous and current. Limitation: The meta-analysis may not exist, or it may not have found a 22% improvement.
  • Feature: “published in the Journal of Educational Psychology” — A real, reputable journal. Limitation: The specific article may not exist in that journal, or the claim may be misrepresented.
  • Feature: “22% across 45 studies” — Specific and seemingly data-driven. Limitation: The number may be fabricated, or the studies may not support that conclusion.
  • Feature: “strongest effects observed in STEM” — Adds nuance and specificity. Limitation: This detail may also be fabricated or unsupported.
🧠 Problem 10.1.8: Transfer Problem

Task: Find a short (2–4 sentence) AI-generated passage from your own work, a public source, or use the opening scenario from this lesson. Produce a claim inventory that includes:

  • Each atomic claim numbered and listed.
  • A classification of each claim (factual, quantitative, interpretive, etc.).
  • At least two claims flagged as possible hallucinations or fabrications, with brief reasoning.
  • At least two credibility features identified and explained.
Show sample transfer solution

Passage (reusing opening scenario): “According to a 2022 study by Martinez and Kowalski published in the Journal of Educational Technology, ‘the integration of adaptive learning platforms has been shown to improve student outcomes in STEM disciplines by an average of 34% across 15 controlled trials’ (Martinez & Kowalski, 2022, p. 112).”

Claim inventory:

  1. A 2022 study by Martinez and Kowalski exists. (Factual — possible fabrication)
  2. The study was published in the Journal of Educational Technology. (Factual — possible fabrication)
  3. The study showed a 34% improvement in STEM outcomes. (Quantitative — possible fabrication)
  4. The study involved 15 controlled trials. (Quantitative — possible fabrication)
  5. The quotation is attributed to Martinez and Kowalski (2022, p. 112). (Factual — possible fabrication)

Flagged claims for verification: All five claims require verification because they involve named sources, a specific statistic, and a page reference.

Credibility features:

  • Citation format (author, date, journal, page) — looks authentic but may be fabricated.
  • Use of quotation marks and page number — adds specificity but can be invented.

13. Mini-Project: Claim and Fabrication Audit

📁 Project: Claim and Fabrication Audit
  • Purpose: Synthesise the skills of claim identification, hallucination and fabrication screening, and credibility-feature analysis in a systematic audit of a synthetic AI-generated passage.
  • Learning outcomes: All Lesson 10.1 outcomes.
  • Estimated effort: 75–100 minutes.

Passage to analyse

Use the synthetic passage below, or choose one of the alternative passages from the unit materials. If using a live AI system, generate a 5–7 sentence passage on a synthetic topic (e.g., “The impact of AI on remote work productivity,” “Trends in online learning engagement,” etc.) using public or authorised information only. Do not use real personal, confidential, or sensitive data.

📌 Passage — “AI and Productivity in the Workplace” (synthetic extract)

A 2022 report by the Digital Transformation Institute concluded that “organisations that have implemented AI-powered productivity tools have seen a 28% increase in employee output, with the most significant gains observed in knowledge-intensive sectors such as finance and technology” (DTI, 2022, p. 45). The report, which surveyed 200 companies across 12 countries, also noted that 73% of employees reported improved job satisfaction when AI tools were used to automate repetitive tasks. These findings align with a 2021 study by Chen and Patel, published in the Journal of Organisational Behaviour, which found a 22% reduction in task completion time among teams using AI-assisted project management software (Chen & Patel, 2021, p. 112). However, some researchers have cautioned that the long-term effects of AI integration on workplace culture and employee autonomy remain poorly understood.

All names, organisations, statistics, and publication details are synthetic. Do not rely on them as factual.

Required deliverable

Produce a structured audit document that includes the following sections:

  1. Passage purpose and intended audience (2–3 sentences).
  2. Claim inventory — Numbered list of every atomic claim.
  3. Atomic claims — Separate and list all atomic claims from compound statements.
  4. Compound claims — Identify any compound claims and show how you separated them.
  5. Interpretive statements — Identify any interpretive claims and explain why they are interpretive.
  6. Recommendations — Identify any recommendations or normative claims.
  7. Possible hallucinations — List at least three claims that may be hallucinations, with brief reasoning.
  8. Possible fabrications — List at least two claims that may be fabrications, with brief reasoning.
  9. Citation concerns — Identify any citation-like references that require verification.
  10. Quotation concerns — Identify any quotations that require verification.
  11. Claims requiring verification — List the claims that most urgently require verification, with priority ranking (high/medium/low).
  12. Consequence of error — For each high-priority claim, describe the possible consequence if the claim is wrong.
  13. Priority order — List the top 5 claims to verify first, in order of priority.
  14. Escalation or stop condition — Identify any conditions that would require escalation to a subject-matter expert or stopping use of the passage.
  15. Preliminary assessment — A 2–3 sentence summary of your overall assessment of the passage’s reliability.

Self-assessment criteria

  • Did you identify at least 10 atomic claims?
  • Did you separate all compound claims into their atomic components?
  • Did you flag at least 3 possible hallucinations and 2 possible fabrications?
  • Did you provide reasoning for each flag?
  • Did you prioritise claims for verification?
  • Did you identify an escalation or stop condition?
Show staged hints

Hint 1: Start with the claim inventory. Read the passage once without judgment. Then go sentence by sentence and ask: “What is being asserted here?” Write each assertion as a complete, standalone sentence. This gives you your raw claim inventory.

Hint 2: Separate compound claims. Look for sentences with multiple clauses, especially those with “and,” “but,” “or,” or embedded references. Break them into atomic claims.

Hint 3: Flag possible hallucinations. Look for claims that are specific, consequential, and would require external evidence. Pay special attention to statistics, named sources, and precise dates or page numbers.

Hint 4: Determine verification priority. Ask: “If this claim is wrong, what would be the consequence?” Claims with high consequence and high plausibility of being wrong should be prioritised.

Show illustrative partial example (based on a different passage)

Illustrative example (Passage G — different from the main passage):

Passage: “According to a 2021 UNESCO report, 67% of countries have adopted AI ethics frameworks, and this has led to a 15% increase in public trust in AI systems (UNESCO, 2021, p. 34).”

Claim inventory (partial):

  • A 2021 UNESCO report exists.
  • 67% of countries have adopted AI ethics frameworks.
  • Adoption of AI ethics frameworks has led to a 15% increase in public trust.
  • The 15% increase is attributed to the UNESCO report (p. 34).

Possible hallucinations:

  • The 67% figure — highly specific and would require cross-country data.
  • The 15% increase — specific and causal, requiring evidence of a causal link.

Possible fabrications:

  • The UNESCO report itself — may not exist, or may not contain these claims.
  • The page reference (p. 34) — may be invented.

Verification priority:

  • High: Verify the UNESCO report exists and contains these claims.
  • High: Verify the 67% figure.
  • High: Verify the 15% increase and the causal claim.
  • Medium: Verify the page reference (if the source exists).

Escalation condition: If the UNESCO report cannot be found, or if the report does not contain these claims, the passage should be escalated or rejected.

14. Self-Check Quiz

8 questions · Formative · Ungraded. Expand each item to see feedback.

Question 1: Which of the following is a claim as defined in this lesson?

Select one:
Show answer and explanation

Answer: b.

“The experiment was conducted in 2022” is a factual claim that can be examined and verified. The others are a command, a question, and an exclamation—none of which assert a proposition.

Question 2: What is an atomic claim?

Select one:
Show answer and explanation

Answer: b.

An atomic claim is a single proposition that can be examined independently. It may be true or false, and it may be part of a larger argument, but it is the smallest unit of assertion that can be evaluated on its own.

Question 3: What is the key distinction between fabrication and factual error?

Select one:
Show answer and explanation

Answer: b.

Fabrication involves invented details presented as real (e.g., a nonexistent source), while a factual error is an incorrect statement about reality (e.g., a wrong date). Both are problems, but they have different characteristics and may require different responses.

Question 4: Which of the following is not a warning sign of a fabricated reference?

Select one:
Show answer and explanation

Answer: a.

A real DOI that resolves to the correct publication is a sign that the reference is not fabricated (though the attribution of the claim may still be wrong). The other options are all warning signs that should trigger verification.

Question 5: Why does fluent, well-formatted language not establish the reliability of an AI-generated passage?

Select one:
Show answer and explanation

Answer: a.

Fluency is a feature of how the language model generates text; it is not evidence of truth or accuracy. A passage can be perfectly fluent and entirely fabricated. Reliability requires independent verification of the claims.

Question 6: What is the first step in the TRACE framework as introduced in this lesson?

Select one:
Show answer and explanation

Answer: c.

The first step in the TRACE framework is the identification of claims and error signals. This means separating the output into its constituent claims and flagging any features that suggest potential error. Classification, correction, and communication come in later stages.

Question 7: A claim that is plausible but lacks sufficient evidence is best described as:

Select one:
Show answer and explanation

Answer: c.

An unsupported claim is one that may be true but lacks sufficient evidence. It is not necessarily fabricated or false; it simply cannot be verified with the information provided.

Question 8: Which of the following is a compound claim?

Select one:
Show answer and explanation

Answer: c.

This sentence contains multiple atomic claims: a study exists, it was by Dr. Singh, it was published in The Lancet, it was published in 2020, it found a reduction, and the reduction was 12%. Options A, B, and D are atomic claims.

15. Assignment Connections

Assignment 5: Verification and Fact-Checking Assignment

Unit 10 begins direct preparation for Assignment 5: Verification and Fact-Checking Assignment (10% of final grade, submitted at the end of Module 4). In Lesson 10.1, you have developed skills that are essential for Assignment 5:

  • Passage segmentation: You have learned to read an AI-generated passage and separate it into its constituent claims.
  • Claim inventory: You have practised creating systematic inventories of atomic and compound claims.
  • Error classification: You have begun to classify claims as possible hallucinations, fabrications, factual errors, or unsupported claims.
  • Fabrication analysis: You have learned to identify warning signs in citations, quotations, and statistics.
  • Verification priorities: You have started to prioritise claims based on consequence and other factors.

Retain your work: The claim inventories, error flags, and verification priorities you develop in this lesson are valuable evidence for Assignment 5. Keep copies of your exercises, homework, and mini-project outputs. You will build on them in Lessons 10.2, 10.3, and 10.4, and in Unit 11’s detailed verification methods.

Do not submit any work from this lesson as Assignment 5. The assignment will require a full verification of an AI-generated passage, including source verification, evidence evaluation, and corrected wording. Lesson 10.1 provides the foundational skills; Unit 11 will provide the verification methods.

📝 Assignment 1: Reflective Journal

Consider addressing the following prompts in your journal:

  1. Changed understanding of hallucination: How has your understanding of “hallucination” changed after this lesson? What did you previously think it meant, and what do you now understand it to be?
  2. Claim decomposition: Describe your experience of dividing a compound passage into atomic claims. What did you find challenging, and what did you learn from the process?
  3. Fabrication and credibility: Reflect on a time when you encountered a plausible but fabricated claim (in an AI output or elsewhere). What made it seem credible, and how might you approach such a claim differently now?
  4. Trust and verification: How has this lesson affected your trust in AI-generated content? What practical changes might you make to your verification habits?
Show reflection-planning checklist
  • Did you define hallucination in your own words?
  • Did you describe at least one specific experience with claim decomposition?
  • Did you identify a concrete example of a fabricated claim and explain why it seemed credible?
  • Did you articulate a specific change in your verification practice?
  • Did you connect your reflection to the course learning outcomes?

16. Key Takeaways

Remember these points
  • A claim is an assertion that can be examined. Not every sentence is a claim; questions, commands, and exclamations are generally not claims.
  • Compound claims must be separated into atomic claims. Each atomic claim requires its own evidence.
  • Hallucination is ungrounded content presented plausibly. It is not a mental state; it is a pattern in the output.
  • Fabrication involves invented details presented as real. It is distinct from factual error, which is a misstatement of reality.
  • Fabricated references and quotations are common. Do not trust citation-like formatting; verify the source independently.
  • Fluency, detail, and structure create false credibility. These are features of the language model, not evidence of truth.
  • The first step in TRACE is identifying claims and error signals. You cannot analyse what you have not first identified.
  • Verification priority is determined by consequence, centrality, and other factors. Not all claims need the same level of scrutiny.

17. Key Terms

Claim
A statement that presents something as true, probable, likely, appropriate, or supported. A claim is an assertion that can be examined, questioned, verified, or refuted.
Atomic claim
A claim containing one principal proposition that can be examined independently.
Compound claim
A statement containing several propositions that may require different evidence, each of which can be separated into atomic claims.
Hallucination
Generated content that is presented plausibly but lacks adequate grounding or alignment with reality, supplied evidence, or authoritative sources.
Fabrication
Generating details that are represented as real but are invented or unsupported—e.g., nonexistent sources, invented statistics, or false quotations.
Factual error
A claim that conflicts with reliable evidence—a wrong date, incorrect name, false numerical value, or inaccurate description.
Unsupported claim
A claim that may be plausible but lacks sufficient evidence; it is not necessarily false, but it cannot be verified with the available information.
Credibility feature
A textual or stylistic characteristic—such as confident tone, detailed references, or precise statistics—that may create an impression of reliability but does not, in itself, prove accuracy.
Verification priority
The order in which claims should be verified, determined by factors such as consequence, centrality, specificity, and source concern.
TRACE (operational)
A framework for error analysis: identifying claims and error signals, classifying error types, assessing consequence and verification priority, correcting or qualifying content, and communicating evidence and uncertainty.

18. Looking Ahead: Lesson 10.2

Lesson 10.2: Factual Error, Omission, Framing, and Contextual Mismatch will build on the claim-identification skills you have developed here. You will learn to distinguish false claims from unsupported claims, identify material omissions, analyse misleading framing, and detect contextual mismatches including outdated information. You will also explore how a statement can be technically true yet still materially misleading.

After completing Lesson 10.2, you will be better prepared to evaluate the full range of errors that can appear in AI-generated content—not just outright fabrications, but also subtle omissions and framing choices that distort understanding.

Continue to Unit 11 for detailed source-verification methods, including the CLAIM framework, after you complete all four lessons in Unit 10.

Continue to Lesson 10.2: Factual Error, Omission, Framing, and Contextual Mismatch Return to Unit 10 Home

19. Sources & Further Reading

[VERIFICATION REQUIRED: Additional sources, page numbers, and jurisdiction-specific references will be confirmed when the final textbook is supplied.]

⚠️ Note

All examples, scenarios, and references in this lesson are synthetic. Do not rely on them as factual without independent verification. Some textbook content, definitions, examples, and figures are marked with placeholders and will be supplied by the course coordinator.

👤