Table of Contents
- 1. Lesson Overview
- 2. Learning Outcomes
- 3. Before You Begin
- 4. Retrieval Practice
- 5. Opening Scenario
- 6. Core Lesson Content
- 6.1 Evidence, Assumption, Inference, and Conclusion
- 6.2 Invalid Inference
- 6.3 Correlation vs. Causation
- 6.4 Overgeneralisation
- 6.5 False Precision
- 6.6 Distinguishing Forms of Confidence
- 6.7 Uncertainty and Its Communication
- 7. TRACE Walkthrough — Classifying Error Type
- 8. Error-Analysis Process Walkthrough
- 9. Worked Example
- 10. Guided Learning Activities
- 11. Exercises
- 12. Homework Problems
- 13. Mini-Project
- 14. Self-Check Quiz
- 15. Assignment Connections
- 16. Reflective Journal Prompts
- 17. Key Takeaways
- 18. Key Terms
- 19. Looking Ahead
- 20. Sources & Further Reading
1. Lesson Overview
In Lessons 10.1 and 10.2, you learned to identify and classify factual claims, fabrications, omissions, and contextual mismatches. Now we turn to the reasoning that connects evidence to conclusions. AI-generated outputs often contain inferences—leaps from what is known (or assumed) to what is claimed as a result. These inferences may be invalid, overgeneralised, or based on a confusion of correlation with causation. Moreover, AI systems often present their conclusions with false precision—numbers and probabilities that are more exact than the evidence supports—and use language that exudes confidence regardless of the underlying evidence.
This lesson teaches you to dissect the reasoning in an AI-generated output: to separate observation from evidence, assumption from inference, and prediction from conclusion. You will learn to detect invalid inferences, correlation-causation errors, and overgeneralizations. You will also examine false precision and the different forms of confidence—linguistic, statistical, model-based, and human—and learn to communicate uncertainty in a way that is proportionate to the evidence.
These skills are essential for responsible AI use: a claim may be factually correct in isolation, but if the reasoning that connects evidence to conclusion is flawed, the overall output can be misleading or dangerous. By the end of this lesson, you will be able to evaluate not only what is claimed, but how it is supported (or not).
2. Learning Outcomes
By the end of this lesson, you should be able to:
- Distinguish observation, evidence, assumption, inference, prediction, recommendation, and conclusion in AI-generated text.
- Identify invalid inferences where the conclusion does not follow from the premises.
- Recognise correlation-causation errors and explain why they are problematic.
- Detect overgeneralizations that extend claims beyond the available evidence.
- Identify false precision in numerical and categorical claims.
- Distinguish linguistic confidence, statistical confidence, model probability, subjective human confidence, and evidential support.
- Communicate uncertainty using language that is proportionate to the evidence.
- Apply the TRACE framework's error-classification component to identify reasoning errors.
3. Before You Begin
- Assigned reading: Chapter 7, sections on “Inference,” “Uncertainty,” and “Precision.”
- Prerequisite knowledge: You should be comfortable with claim identification (Lesson 10.1) and the distinction between factual errors, omissions, and framing (Lesson 10.2). Review the definitions of assumption and inference from the course glossary.
- Optional approved AI tool: You may use an institutionally approved AI system with synthetic, public, or authorised information. Do not enter personal, confidential, proprietary, or sensitive data.
- Supplied-output alternative: All activities include synthetic outputs for analysis.
- Privacy and integrity reminder: Do not upload confidential or personally identifiable information.
- Verification reminder: An AI system cannot authenticate its own claims; you must independently evaluate reasoning and evidence.
- Estimated preparation time: 15–20 minutes to review prerequisites.
4. Retrieval Practice
These questions draw on Lessons 10.1 and 10.2. Answer briefly, then check the feedback.
Question 1: In your own words, why is it important to separate compound claims into atomic claims before evaluating reasoning?
Show feedback
Sample answer: Compound claims contain multiple propositions that may each require different evidence or different forms of reasoning. If you treat them as a single unit, you risk overlooking a weak inference or an unsupported assumption embedded in one part. Separating them allows you to examine each link in the chain of reasoning independently.
Question 2: How does an unsupported claim differ from a factual error? Why does this matter for reasoning analysis?
Show feedback
Sample answer: A factual error is a claim that conflicts with reliable evidence; it is demonstrably wrong. An unsupported claim may be true or false, but it lacks sufficient evidence to establish it. In reasoning analysis, an unsupported claim is a red flag for a weak inference—it may be a leap of faith. A factual error, on the other hand, invalidates the premise entirely. Both undermine the conclusion, but they require different corrective actions.
Question 3: How can a material omission affect the validity of an inference drawn from AI-generated content?
Show feedback
Sample answer: An inference is only as strong as the evidence it is based on. If crucial information is omitted (e.g., a conflicting study, a limiting condition, or an alternative explanation), the inference may appear valid when it is actually incomplete. Omitting such information can lead to overgeneralization or false certainty. Thus, identifying omissions is a prerequisite for sound inference evaluation.
5. Opening Scenario: A Synthetic Report on AI and Productivity
Read the following synthetic AI-generated report excerpt. It contains several reasoning errors, including invalid inference, correlation-causation confusion, overgeneralization, and false precision. As you read, note any claims that seem to leap beyond the evidence.
A recent survey of 500 companies found that those which had implemented AI-powered project management tools reported a 15% increase in on-time project delivery compared to those that did not. This clearly demonstrates that AI tools cause improved project efficiency. Moreover, the same survey revealed that 78% of employees in these companies felt more engaged in their work. Therefore, AI tools not only boost productivity but also enhance employee morale.
Based on these findings, we can confidently predict that within the next two years, all companies that adopt such tools will see a 20% increase in overall profitability. Additionally, the data suggests that the most effective implementation strategy is to roll out the tools company-wide within the first six months, as this yields the highest returns.
Note: All statistics and survey details are synthetic. Do not rely on them as factual.
Initial Questions
- What evidence does the passage present? What assumptions does it make?
- What inferences does the passage draw from the survey data?
- What causal claims are made, and are they justified?
- What predictions and recommendations are offered? On what basis?
Show guided analysis (initial)
Observations:
- Evidence: A survey of 500 companies; reported 15% increase in on-time delivery; 78% employee engagement.
- Assumptions: The survey is representative; the reported increases are accurate; no other factors explain the changes.
- Inferences: The 15% increase is caused by AI tools (causal leap). Engagement is also caused by AI tools. These will lead to a 20% profitability increase within two years (prediction).
- Recommendation: Roll out tools company-wide within six months—based on unsupported causal reasoning.
- False precision: 15%, 78%, 20% all appear exact but are based on a single survey; predictions are overly specific.
Key takeaway: The passage makes strong causal and predictive claims without ruling out alternative explanations, without acknowledging uncertainty, and with a level of precision that is likely unjustified.
6. Core Lesson Content
6.1 Evidence, Assumption, Inference, and Conclusion
- Observation / Evidence:
- Data, facts, or reports that are directly observed or obtained from a source.
- Assumption:
- Something treated as true or accepted for the purpose of reasoning without being established fully.
- Inference:
- A conclusion drawn from evidence or premises; a logical step from what is known to a new claim.
- Conclusion:
- The final claim that the argument is intended to support.
In AI-generated content, these elements are often intermingled. A fluent passage may present an inference as if it were direct evidence, or hide assumptions behind confident language.
Example:
- Evidence: “In a survey of 500 companies, those using AI tools reported 15% higher on-time delivery.”
- Assumption: The survey is accurate and representative; no other factors (e.g., better management) affected on-time delivery.
- Inference: “Therefore, AI tools improve delivery times.” (Causal inference.)
- Conclusion: “AI tools are effective at improving project outcomes.”
Why it matters: To evaluate an AI output, you must separate these components. Ask: “What is directly observed? What is assumed? What is inferred? What is the final conclusion?”
Consider: “The stock price rose after the CEO’s speech. The speech must have caused the rise.” Identify the evidence, assumption, inference, and conclusion.
Show decomposition
- Evidence: The stock price rose after the CEO’s speech.
- Assumption: The speech was the only (or main) cause of the rise; no other market factors were at play.
- Inference: The speech caused the rise (causal inference).
- Conclusion: The CEO’s speech was effective.
Warning: This is a classic post hoc ergo propter hoc fallacy—after this, therefore because of this. Correlation in time does not equal causation.
6.2 Invalid Inference
Invalid inference: A conclusion that does not logically follow from the premises or evidence.
Invalid inferences can take many forms:
- Leap in logic: The conclusion is stronger than the evidence permits.
- Missing link: A necessary intermediate step is skipped.
- Confusing necessary and sufficient conditions: Assuming that because A is required for B, A guarantees B.
- Appeal to authority without evidence: “Experts agree” without citing the evidence.
Example (synthetic): “The new software reduced errors by 10% in one department. Therefore, it will reduce errors by 10% in all departments.” This inference ignores differences in context, team skills, and task complexity.
6.3 Correlation vs. Causation
Common AI error: Generative AI systems often infer causation from observed patterns because they are trained on text that frequently makes such leaps. They may present a correlation as if it were a proven causal relationship.
Examples:
- Correlation: Ice cream sales and drowning deaths both increase in summer. Does ice cream cause drowning? No—heat is a confounder.
- Reverse causation: “Companies with higher profits adopt more AI.” An AI output might claim “AI adoption causes higher profits,” but it could be that profitable companies can afford AI.
- Coincidence: Two trends happen at the same time without any causal link.
How to detect: Look for language like “leads to,” “causes,” “results in,” “drives,” “improves.” Ask: “Has the system ruled out alternative explanations? Has it established a mechanism?”
6.4 Overgeneralisation
Overgeneralisation: Extending a claim beyond the scope of the evidence—applying findings from a specific sample, context, or timeframe to a much broader population or setting.
Examples:
- “A study of 50 university students found that AI tutoring improved grades. Therefore, AI tutoring improves grades for all students in all subjects.”
- “Our pilot program in one region succeeded. Therefore, the program will succeed nationwide.”
- “The model performed well on the test data. Therefore, it will perform well in real-world deployment.”
Why it happens: AI systems may generate broad claims because their training data often contains generalisations. They lack the awareness to qualify a claim as context-dependent.
Detection: Look for universal terms (“all,” “every,” “always,” “never”) and ask: “Does the evidence support this breadth?”
6.5 False Precision
False precision: Presenting more numerical or categorical exactness than the evidence supports—e.g., exact percentages, precise dates, or detailed rankings based on vague or insufficient information.
Signs of false precision:
- Exact percentages from a small or non-random sample: “67.3% of respondents agree.”
- Precise forecasts without a validated model: “Sales will increase by 12.4% next quarter.”
- Detailed rankings with no criteria: “The best AI tool is X, followed by Y and Z.”
- Numerical confidence expressed as a single number (e.g., “95% certain”) without a statistical basis.
Why it is dangerous: False precision creates an illusion of rigor and certainty, which can lead decision-makers to overestimate the reliability of the information.
Example: “Based on our survey of 30 people, we are 92.5% confident that the new policy will increase satisfaction.” The sample size is too small to support that level of precision.
6.6 Distinguishing Forms of Confidence
AI-generated text often uses language that expresses high confidence, but this is linguistic confidence, not evidential support. It is essential to distinguish several concepts:
| Type | Definition | Example |
|---|---|---|
| Linguistic confidence | The wording—assertive, certain, definitive—regardless of the actual evidence. | “It is clear that…” |
| Model probability (or score) | A technical value (like a softmax probability) associated with a generated token or output. Reflects the model’s internal statistics, not the truth of the content. | A softmax score of 0.92 for a generated answer. |
| Statistical confidence | A defined concept from inferential statistics (e.g., confidence interval, p-value) that quantifies uncertainty in a well-specified statistical model. | A 95% confidence interval of [0.5, 1.2]. |
| Human confidence | A person’s subjective degree of belief, which may be informed by expertise, experience, and evidence. | “I feel confident that this is correct.” |
| Evidential support | The actual quality and relevance of evidence supporting a claim—this is what should determine our confidence. | Multiple peer-reviewed studies confirming a finding. |
Example: An AI system might say, “It is certain that 87% of doctors recommend this treatment.” The linguistic confidence is high, but the claim may be fabricated. The model probability score is a statistical property of the model, not a reflection of the real-world claim. A human’s confidence should be based on verifying the source and evidence.
6.7 Uncertainty and Its Communication
Uncertainty: The state of being unsure or having limited knowledge about a claim, due to missing evidence, conflicting evidence, ambiguity, model variability, incomplete context, changing conditions, measurement limitations, interpretation, future events, or unknown system behaviour.
Responsible AI users communicate uncertainty rather than hiding it. When you use or share AI-generated content, you should qualify claims that are not fully supported.
Appropriate uncertainty language:
- “Available evidence suggests…”
- “The supplied source states…”
- “This claim requires verification.”
- “Evidence is incomplete.”
- “Sources disagree.”
- “The result depends on [condition].”
- “The estimate is approximate.”
- “No conclusion can be reached from the available information.”
- “Qualified review is required.”
Avoid vague disclaimers that do not identify the actual uncertainty, such as “results may vary” without specifying the conditions.
Example of responsible communication: “Based on a single survey of 500 companies, there is some evidence that AI tools are associated with improved on-time delivery, but causation has not been established. The 15% figure is an estimate and may not generalise to other industries.”
Rewrite the following overconfident statement to reflect the actual uncertainty: “Our data proves that the new training program increases productivity by 20%.”
Show sample rewrite
Sample rewrite: “Our data suggest that the new training program may be associated with a productivity increase, but the 20% figure is based on a small sample and requires further verification. The results should be interpreted cautiously and may not apply to all contexts.”
7. TRACE Walkthrough: Classification of Error Type
In Lesson 10.1 we focused on the first TRACE element—identifying claims and error signals. Now we apply the second element: classifying error type. In this lesson, we focus on reasoning-related errors: invalid inference, correlation-causation, overgeneralization, and false precision.
After identifying claims and error signals, assign each claim to one or more error categories: hallucination, fabrication, factual error, unsupported claim, omission, misleading framing, contextual mismatch, outdated information, invalid inference, overgeneralization, correlation-causation, false precision, or uncertainty.
Guiding questions:
- Does the conclusion follow logically from the evidence?
- Is a causal relationship claimed without ruling out alternative explanations?
- Is the claim extended beyond the scope of the evidence?
- Is the numerical precision justified by the data?
- What type of uncertainty is present, and is it acknowledged?
Example: In the opening scenario, the claim “AI tools cause improved project efficiency” is a correlation-causation error and an invalid inference. The prediction of “20% increase in profitability” is false precision and overgeneralization.
Common mistake: Treating an invalid inference as a factual error. The evidence may be accurate, but the conclusion drawn from it is not justified.
Required evidence: The original evidence, the assumptions, and the chain of reasoning.
Possible correction: Reframe the conclusion as a tentative inference, not a proven fact. Add qualifications and acknowledge alternative explanations.
Escalation or stop condition: If the core conclusion is based on an invalid inference and the evidence is insufficient, escalate the need for expert review or stop using the conclusion as a basis for decisions.
8. Error-Analysis Process Walkthrough: Evidence → Assumption → Inference → Precision → Conclusion → Uncertainty
This walkthrough shows how to systematically analyse reasoning in an AI-generated passage, using the opening scenario as a running example.
- Stage 1: Identify the evidence and observations. List what is directly reported or observed. In the scenario: survey of 500 companies; 15% increase in on-time delivery; 78% employee engagement.
- Stage 2: Identify assumptions. What is taken for granted? Examples: survey is representative; increases are due to AI tools; no other changes occurred; the measurements are accurate; the effects will persist.
- Stage 3: Identify inferences. What conclusions are drawn from the evidence? “AI tools cause improved delivery.” “AI tools enhance morale.” “Profits will rise 20%.”
- Stage 4: Evaluate inference validity. Does the conclusion follow? For the causal claim, no—correlation is mistaken for causation. For the prediction, no—it extrapolates beyond the data without justification.
- Stage 5: Check for overgeneralization. Are the conclusions applied more broadly than the evidence? The scenario applies to “all companies” and “all employees” without basis.
- Stage 6: Examine numerical precision. Are the numbers (15%, 78%, 20%) justified? The survey provides some basis, but the precision is likely overstated given the single source and lack of margin of error.
- Stage 7: Identify confidence language and distinguish from evidence. Language like “clearly demonstrates,” “confidently predict,” “will see” expresses high linguistic confidence, but the evidential support is weak.
- Stage 8: Communicate uncertainty. Draft a revised version that acknowledges the limitations: “Based on a single survey, there is a correlation between AI tool adoption and reported on-time delivery, but causal inference is not warranted. The 15% figure is an estimate; further study is needed.”
- Stage 9: Determine consequence and priority. The causal claim and profit prediction have high consequence—they could lead to expensive decisions. They should be prioritised for verification and expert review.
9. Worked Example: Analysing Reasoning in a Synthetic AI Output
Consider the following synthetic AI-generated paragraph about the effect of remote work on team collaboration.
A 2023 survey of 200 remote teams found that teams using video conferencing daily reported 40% higher collaboration scores than those using it weekly. This demonstrates that daily video meetings are essential for effective collaboration. Furthermore, the survey showed that 85% of employees in high-collaboration teams felt more connected to their colleagues, suggesting that video conferencing directly causes a sense of belonging. Companies should therefore mandate daily video meetings to ensure team cohesion and improve project outcomes.
Step-by-step analysis
1. Evidence
Survey of 200 remote teams; 40% higher collaboration scores for daily vs. weekly; 85% feel connected.
2. Assumptions
The survey is representative; collaboration scores are valid; no other differences between daily and weekly groups; causation is assumed; the findings apply to all companies.
3. Inferences
- Daily meetings cause higher collaboration (correlation-causation).
- Video conferencing causes a sense of belonging (correlation-causation).
- Mandating daily meetings will improve outcomes (overgeneralization and invalid inference).
4. Invalid inference detection
The conclusion “daily meetings are essential” does not follow—the data only show an association. Other factors (team size, company culture) could explain the difference.
5. Overgeneralization
The recommendation applies to “all companies” without considering industry, team maturity, or other variables.
6. False precision
“40% higher,” “85%” are presented as exact, but the survey method is not described; margins of error are unknown.
7. Confidence language
“Demonstrates,” “essential,” “directly causes,” “should therefore mandate” are highly confident and unjustified.
8. Revised uncertainty communication
“A 2023 survey found an association between daily video conferencing and higher collaboration scores among 200 remote teams. However, this is a correlation; causation has not been established. The 40% difference is a survey estimate and may be influenced by other factors. Mandating daily meetings should be evaluated based on context and further research.”
9. Consequence and priority
High consequence—mandating meetings could impact productivity and morale. The causal claims should be verified with more robust studies before acting.
Show complete error classification table
| Claim | Error type(s) |
|---|---|
| Daily video meetings cause higher collaboration | Correlation-causation, invalid inference |
| Video conferencing causes sense of belonging | Correlation-causation, invalid inference |
| Mandating daily meetings will improve outcomes | Overgeneralization, invalid inference |
| 40% higher, 85% | False precision (likely, given survey details omitted) |
| “Demonstrates,” “essential,” “directly causes” | Linguistic confidence mismatched with evidential support |
| Recommendation for all companies | Overgeneralization, contextual mismatch |
10. Guided Learning Activities
- Purpose: Practise distinguishing components of reasoning.
- Learning outcome: Distinguish observation, evidence, assumption, inference, and conclusion.
- Estimated time: 15 minutes.
Instructions: For each statement below, classify it as Evidence, Assumption, Inference, or Conclusion. (Some may fit more than one; choose the best fit.)
- “A survey of 100 students found that those who used the AI tutor scored 10% higher.”
- “The students who used the AI tutor were more motivated.”
- “The AI tutor must have caused the improvement.”
- “The survey is representative of all college students.”
- “AI tutors are effective learning tools.”
Show classifications
- Evidence — a reported observation.
- Evidence (if directly reported) or Inference (if concluded from other data).
- Inference (causal claim drawn from the evidence).
- Assumption — taken for granted.
- Conclusion — the overall claim the argument supports.
- Purpose: Identify invalid inferences, overgeneralizations, and correlation-causation errors.
- Learning outcome: Identify invalid inference, overgeneralization, and correlation-causation.
- Estimated time: 20 minutes.
Instructions: Read the synthetic passage below and list at least three reasoning errors, classifying each as invalid inference, overgeneralization, or correlation-causation.
“Our analysis of 50 startups shows that those using cloud computing services grew revenue by 30% faster than those that did not. This proves that cloud computing is the key to startup success. Therefore, any new startup should invest heavily in cloud infrastructure. In addition, we observed that startups with cloud computing also had higher employee satisfaction, indicating that the technology directly improves morale.”
Show sample analysis
- Correlation-causation: “proves that cloud computing is the key” – the correlation may be due to other factors.
- Overgeneralization: “any new startup” – the sample is limited to 50 startups, not all.
- Invalid inference: “cloud computing directly improves morale” – no mechanism is established.
- Purpose: Detect unjustified numerical specificity.
- Learning outcome: Identify false precision.
- Estimated time: 15 minutes.
Instructions: Identify examples of false precision in the following statements. Explain why the precision is unjustified.
- “Based on our survey of 40 people, we are 94.7% confident that the product will be successful.”
- “Sales will increase by exactly 8.3% next quarter.”
- “The new policy will reduce costs by $1,234,567 per year.”
- “Our algorithm is 99.9% accurate.” (No details on test set or error types.)
Show sample explanations
- Sample size of 40 is too small to support such a precise confidence level; the figure is likely fabricated.
- Predicting an exact percentage for an entire quarter is unrealistic without a validated forecasting model.
- Such a specific dollar amount implies detailed cost data that are probably unavailable.
- Accuracy claims require a clear definition of accuracy, a representative test set, and error analysis; 99.9% is often a marketing claim rather than a verified metric.
- Purpose: Distinguish linguistic confidence, statistical confidence, model probability, human confidence, and evidential support.
- Learning outcome: Distinguish forms of confidence.
- Estimated time: 15 minutes.
Instructions: For each scenario, identify which type(s) of confidence are being expressed and whether they are justified.
- An AI system outputs: “It is certain that 85% of teachers support the new curriculum.”
- A researcher reports: “The 95% confidence interval for the mean effect is [0.5, 1.2].”
- A user says: “I feel confident that this AI summary is correct because it reads well.”
- An AI system returns a softmax probability of 0.92 for a generated answer.
Show sample analysis
- Linguistic confidence (“it is certain”) but lack of evidential support – unjustified.
- Statistical confidence – a proper interval; justified if methodology is sound.
- Human confidence based on fluency, not evidence – unjustified.
- Model probability – a technical score; does not imply factual accuracy.
- Purpose: Practise communicating uncertainty proportionately.
- Learning outcome: Communicate uncertainty using language proportionate to the evidence.
- Estimated time: 15 minutes.
Instructions: Rewrite the following overconfident statement to include appropriate uncertainty language.
“The evidence proves that renewable energy will completely replace fossil fuels by 2030, and this will create 5 million new jobs.”
Show sample rewrite
Sample rewrite: “Some projections suggest that renewable energy could significantly reduce reliance on fossil fuels by 2030, but this depends on policy, technology, and economic factors. Job creation estimates are highly uncertain and vary widely; a figure of 5 million is one scenario and should not be treated as a prediction.”
11. Exercises
- Difficulty: Introductory
- Estimated time: ~10 min
- Learning outcome: Distinguish components of reasoning.
Task: For each item, label it as Evidence, Assumption, Inference, or Conclusion.
- “The test scores improved after the new teaching method was introduced.”
- “The new teaching method was the cause of the improvement.”
- “The students were motivated.” (from survey)
- “The sample is representative of all schools.”
- “The new teaching method is effective.”
Show answers
- Evidence
- Inference (causal)
- Evidence (if directly observed) or Inference
- Assumption
- Conclusion
- Difficulty: Intermediate
- Estimated time: ~10 min
- Learning outcome: Identify invalid inference.
Task: Identify the invalid inference in the following argument and explain why it is invalid.
“Our analysis shows that companies with diverse leadership teams have higher profits. Therefore, to increase profits, a company must hire more diverse leaders.”
Show solution
Invalid inference: The argument assumes that diversity causes higher profits (correlation-causation). It also assumes that hiring diverse leaders will inevitably lead to higher profits, which ignores other factors. The conclusion overstates the causal link and ignores possible confounds.
- Difficulty: Intermediate
- Estimated time: ~10 min
- Learning outcome: Detect overgeneralization.
Task: Identify the overgeneralization in the following statement and explain why it is an overgeneralization.
“A pilot program in one high school reduced dropout rates by 30%. Therefore, this program will work in all high schools across the country.”
Show solution
The statement extends findings from a single school to all high schools, ignoring differences in demographics, resources, and local conditions. It assumes the same effect will occur universally, which is an overgeneralization.
- Difficulty: Intermediate
- Estimated time: ~10 min
- Learning outcome: Recognise correlation-causation errors.
Task: For each pair, state whether it is likely correlation or causation and explain.
- Ice cream sales and sunglasses sales.
- Number of fire trucks and fire damage.
- Education level and income.
- AI adoption and company profitability.
Show sample answers
- Correlation (both increase in summer).
- Correlation (more fire trucks at larger fires, not cause).
- Correlation (may be causal but also confounded by other factors).
- Correlation (profitable companies may adopt AI; AI may also help).
- Difficulty: Intermediate
- Estimated time: ~10 min
- Learning outcome: Detect false precision.
Task: Identify the false precision in the following statements and explain why it is problematic.
- “Our survey of 30 customers shows that 83.6% prefer our product.”
- “The model’s accuracy is 97.45% based on a small test set.”
- “We predict the project will cost exactly $1,234,567.”
Show explanations
- Small sample size (30) cannot justify such a precise percentage; margin of error is large.
- Accuracy to two decimal places implies high precision, but the test set may be small or unrepresentative.
- Exact dollar amount suggests detailed cost data that are unlikely to be available early in a project.
- Difficulty: Advanced
- Estimated time: ~10 min
- Learning outcome: Distinguish forms of confidence.
Task: For each statement, identify which type of confidence (linguistic, statistical, model, human, evidential) is being expressed and whether it is justified.
- “We are 99% certain that this drug is safe.” (based on expert opinion only)
- “The model assigned a 0.85 probability to the correct answer.”
- “The study found a statistically significant effect (p < 0.05).”
- “I am convinced that the policy will work.”
Show answers
- Human confidence (expert opinion) – may be unjustified if not evidence-based.
- Model probability – technical score, not a measure of truth.
- Statistical confidence – justified if the study is well-designed.
- Human confidence – subjective; may or may not be evidence-based.
- Difficulty: Advanced
- Estimated time: ~10 min
- Learning outcome: Communicate uncertainty appropriately.
Task: Rewrite the following statement to include appropriate uncertainty language and qualification.
“The data prove that the new algorithm will reduce processing time by 50%, and it will be adopted industry-wide within a year.”
Show sample rewrite
Rewrite: “Preliminary data suggest that the new algorithm may reduce processing time, but the 50% figure is based on limited tests and may not generalise. Adoption timelines are uncertain and depend on many factors; a one-year estimate is speculative.”
12. Homework Problems
Identify the evidence and the inference in this statement: “The number of accidents decreased after the new safety training was introduced. The training must be effective.”
Show solution
Evidence: Accidents decreased after training.
Inference: The training caused the decrease.
Note: The inference is invalid without ruling out other explanations.
Explain why this inference is invalid: “All successful companies use AI, so if we use AI we will become successful.”
Show explanation
The inference commits a logical error: it assumes that using AI is sufficient for success, whereas it is only a correlational factor. Many companies that use AI may not be successful, and success depends on many other factors.
Give an example of a correlation that is often mistakenly interpreted as causation, and explain the likely confounding factor.
Show example
Example: Eating ice cream and drowning deaths. Both increase in summer; heat is a confounder.
Identify the overgeneralization in: “Our study of 100 tech companies shows that flexible work hours increase productivity. Therefore, all companies should adopt flexible hours.”
Show identification
The overgeneralization is applying the finding from tech companies to all companies, ignoring differences in industry, role, and work culture.
Identify false precision in: “Based on a survey of 50 people, we predict a 12.7% increase in sales.”
Show explanation
The sample size of 50 is too small to justify a prediction to one decimal place. The precision is likely spurious.
A statement says: “We are 95% confident that the true mean is between 10 and 12.” What type of confidence is this? What does it mean?
Show answer
This is statistical confidence from a confidence interval. It means that if the sampling procedure were repeated many times, 95% of the intervals would contain the true mean. It does not mean there is a 95% probability that the true mean is in that particular interval.
Rewrite this statement to be more uncertain: “The new policy will definitely cut costs by 20%.”
Show rewrite
“The new policy is projected to reduce costs, but the 20% figure is an estimate based on preliminary analysis and may change.”
Read the following passage and list the evidence, assumptions, inferences, and conclusions. Evaluate the validity of the inferences.
“A 2021 study of 150 students found that those who used a mobile learning app scored higher on final exams. The researchers concluded that the app is responsible for the improved performance. They recommend that all students use the app.”
Show analysis
- Evidence: Study of 150 students; app users scored higher.
- Assumption: No other differences between groups; the study is causal.
- Inference: The app caused the improvement (invalid; correlation only).
- Conclusion: All students should use the app (overgeneralization).
- Validity: Inferences are invalid due to correlation-causation and overgeneralization.
A recommendation says: “Given the strong evidence, we should immediately adopt the new system.” What is the problem with this recommendation if the evidence is only correlational?
Show explanation
The recommendation is too strong; it assumes causation and ignores uncertainty. It should be qualified: “Given the correlational evidence, pilot testing the system is advisable, but a causal link has not been established.”
Find an AI-generated passage (from your own use or a public source) or use the opening scenario of this lesson. Identify the evidence, assumptions, inferences, and conclusions. Evaluate the validity of at least two inferences.
Show sample solution (using opening scenario)
Passage: (Opening scenario.)
Evidence: Survey of 500 companies; 15% increase in on-time delivery; 78% engagement.
Assumptions: Survey representative; AI tools caused the changes.
Inferences: AI tools cause delivery improvement and engagement; these will lead to 20% profit increase.
Conclusions: AI tools are effective; all companies should adopt them.
Validity: Inferences are invalid (correlation-causation, overgeneralization, false precision).
Rewrite the following to include appropriate uncertainty: “The analysis shows that the new marketing campaign will increase sales by 25%.”
Show rewrite
“The analysis suggests that the new marketing campaign may increase sales, but the 25% figure is an estimate and should be validated with more data; actual results may vary.”
Identify all reasoning errors in the following passage and classify them.
“A survey of 200 people found that 70% of those who exercise daily report feeling happier. Therefore, daily exercise guarantees happiness. Everyone should exercise daily to ensure happiness.”
Show classification
- Correlation-causation: Exercise is correlated with happiness, but causation is not established.
- Overgeneralization: Applies to “everyone” from a survey of 200.
- False precision: 70% is a precise figure from a self-reported survey.
- Invalid inference: “guarantees” is too strong.
Using the TRACE classification element, classify the errors in the opening scenario of this lesson. Produce a table with each claim, its error type, and a brief justification.
Show sample classification table
| Claim | Error type | Justification |
|---|---|---|
| AI tools cause improved on-time delivery | Correlation-causation, invalid inference | Survey only shows association. |
| AI tools enhance morale | Correlation-causation | No evidence of causation. |
| 20% profit increase within two years | False precision, overgeneralization | Unsupported prediction. |
| All companies should adopt within six months | Overgeneralization, invalid inference | Assumes universal applicability. |
13. Mini-Project: Reasoning and Uncertainty Audit
- Purpose: Systematically analyse the reasoning, precision, and uncertainty in a synthetic AI-generated argument, and produce a revised version with appropriate qualifications.
- Learning outcomes: All Lesson 10.3 outcomes.
- Estimated effort: 90–120 minutes.
Passage for analysis
A recent meta-analysis of 20 studies found that adolescents who spend more than 3 hours per day on social media report 40% higher rates of anxiety compared to those who spend less than 1 hour. This conclusively shows that social media causes anxiety. The analysis also revealed that 65% of adolescents with high social media use have poor sleep quality, suggesting that sleep disruption is the mechanism through which social media affects mental health. Given these findings, we must immediately limit all adolescents to less than 1 hour of social media per day to protect their well-being. Furthermore, this policy will likely reduce anxiety by 30% within the first year.
Required deliverable
A structured audit document containing the following sections:
- Purpose and main conclusion (2 sentences).
- Evidence claims — list all evidence statements.
- Assumptions — list at least three assumptions.
- Inferences — list all inferences drawn from the evidence.
- Alternative explanations — propose at least two alternative explanations for the observed correlations.
- Invalid reasoning — identify any invalid inferences and explain.
- Overgeneralizations — identify any overgeneralizations.
- Correlation-causation issues — identify any causal claims and assess their justification.
- Numerical claims — list all numbers (40%, 65%, 30%, etc.) and evaluate their precision.
- False precision — identify any false precision.
- Confidence language — identify overconfident wording.
- Uncertainty sources — list at least three sources of uncertainty.
- Consequence of error — describe the consequences if the causal claim is wrong.
- Revised wording — rewrite the passage with appropriate uncertainty and qualification.
- Evidence still required — what additional evidence is needed to support the claims?
- Specialist-review need — indicate whether expert review is required.
- Stop condition — identify a condition under which you would stop using this output.
- Preliminary assessment — summary of your assessment.
Self-assessment criteria
- All 18 sections completed.
- Inferences clearly distinguished from evidence.
- At least three types of reasoning errors identified.
- Revised wording appropriately qualifies claims.
- Uncertainty communicated clearly.
Show staged hints
Hint 1: Start with evidence and conclusion. Identify the raw data/observations first. Then identify the final recommendation or claim. Then trace the steps in between.
Hint 2: Look for causal language. Words like “causes,” “leads to,” “results in,” “shows that” are signals for causal inferences. Question them.
Hint 3: Examine numbers. For each number, ask: “Where did this come from? Is the sample size adequate? Is there a margin of error?”
Hint 4: Propose alternative explanations. For each correlation, think of at least one third variable that could explain it.
Show solution architecture
- Purpose and conclusion: The passage argues that social media causes anxiety in adolescents and recommends limiting use.
- Evidence claims: Meta-analysis of 20 studies; 40% higher anxiety; 65% poor sleep quality.
- Assumptions: Meta-analysis is valid; correlation implies causation; no confounding variables; findings generalise.
- Inferences: Social media causes anxiety; sleep disruption is the mechanism; limiting use will reduce anxiety by 30%.
- Alternative explanations: Underlying mental health conditions, family environment, academic stress.
- Invalid reasoning: Causal claims from correlational data; “conclusively shows” is unjustified.
- Overgeneralizations: “All adolescents”; “must immediately.”
- Correlation-causation: Anxiety and social media; sleep quality and social media.
- Numerical claims: 40%, 65%, 30% — all precise but from unspecified methods.
- False precision: 30% reduction prediction is highly specific.
- Confidence language: “Conclusively,” “must,” “will likely reduce.”
- Uncertainty sources: Meta-analysis methodology, confounding, mechanism, generalisability.
- Consequence of error: If the causal claim is wrong, policy could be ineffective or harmful.
- Revised wording: (See example rewrite.)
- Evidence still required: Longitudinal studies, randomised controlled trials, replication.
- Specialist review: Yes — adolescent mental health specialist.
- Stop condition: If the meta-analysis cannot be verified.
- Assessment: Do not use without verification and specialist review.
Show illustrative partial example (based on a different passage)
Passage: “A study found that children who eat breakfast perform better in school. Therefore, breakfast improves academic performance. Schools should provide free breakfast to all students.”
Evidence: Study shows correlation.
Assumptions: Study is causal; no other factors.
Inferences: Breakfast causes better performance.
Alternative explanations: Children who eat breakfast may come from more affluent families with better resources.
Revised wording: “The study found an association between breakfast consumption and academic performance. Causation has not been established. Providing free breakfast may be worth exploring, but further research is needed to confirm any causal effect.”
14. Self-Check Quiz
Question 1: What is an inference?
Show answer and explanation
Answer: b.
An inference is a logical step from evidence to a new claim.
Question 2: Which of the following is an invalid inference?
Show answer and explanation
Answer: a.
It assumes the storm caused the drop without ruling out other factors; a classic correlation-causation error.
Question 3: What is false precision?
Show answer and explanation
Answer: a.
False precision presents more exactness than the evidence warrants.
Question 4: Why is it a problem to confuse linguistic confidence with evidential support?
Show answer and explanation
Answer: d.
Linguistic confidence is about wording; evidential support is about the underlying evidence. They are separate, and fluent language can hide poor evidence.
Question 5: Which of the following is an example of overgeneralization?
Show answer and explanation
Answer: b.
It applies a finding from a limited sample to all startups.
Question 6: How should uncertainty be communicated in a responsible AI output?
Show answer and explanation
Answer: c.
Responsible communication specifies what is uncertain and why.
Question 7: Which of the following is a model probability?
Show answer and explanation
Answer: b.
Model probability is a technical output of the model, not a measure of factual correctness.
Question 8: What is the primary purpose of the TRACE classification element in the context of reasoning errors?
Show answer and explanation
Answer: b.
The classification element of TRACE involves categorising errors, including reasoning errors, to guide further analysis and correction.
15. Assignment Connections
Lesson 10.3 develops skills that are essential for Assignment 5: Verification and Fact-Checking Assignment. You will need to:
- Identify invalid inferences, overgeneralizations, and correlation-causation errors in an AI-generated passage.
- Evaluate whether numerical claims are justified (false precision).
- Distinguish between linguistic confidence and actual evidential support.
- Communicate uncertainty appropriately in your corrections.
Keep your work from the exercises, homework, and mini-project. You will incorporate your reasoning analysis into your final submission for Assignment 5.
Tip: When you correct an AI output, do not simply remove the error; replace it with a qualified statement that reflects the true state of evidence. Use the uncertainty language you practised.
Reflect on your learning in this lesson using these prompts:
- Invalid inference: Describe a time when you made an invalid inference (or saw one) in an AI-generated output. How did you detect it?
- False precision: Why do you think AI systems often generate false precision? How can you guard against being misled by exact numbers?
- Uncertainty communication: How has your understanding of “confidence” changed? What language will you use to express uncertainty in your own work?
- Correlation vs. causation: Give an example from your field (or everyday life) where it is important to distinguish correlation from causation.
Show reflection-planning checklist
- Did you provide a concrete example of an invalid inference?
- Did you discuss why false precision is problematic?
- Did you articulate a specific change in how you will communicate uncertainty?
- Did you connect your reflection to a real-world context?
16. Key Takeaways
- Separate evidence from inference: Evidence is what is observed; inference is the leap to a conclusion. Do not treat them as the same.
- Correlation does not imply causation: Always question causal claims; look for alternative explanations.
- Avoid overgeneralization: Be careful not to extend findings beyond the scope of the evidence.
- Beware of false precision: Exact numbers do not guarantee accuracy; evaluate the method behind the numbers.
- Confidence is not evidence: Linguistic confidence, model probabilities, and human intuition are distinct from evidential support.
- Communicate uncertainty specifically: State what is uncertain and why, using appropriate qualifying language.
- TRACE classification helps: Use TRACE to categorise reasoning errors and guide correction.
17. Key Terms
- Evidence
- Data, facts, or observations that are directly reported.
- Assumption
- Something treated as true without being fully established.
- Inference
- A conclusion drawn from evidence or premises.
- Invalid inference
- A conclusion that does not logically follow from the premises.
- Correlation-causation error
- Treating a correlation as if it implies a causal relationship.
- Overgeneralization
- Applying findings beyond the scope of the evidence.
- False precision
- Presenting numerical exactness that is not justified by the data.
- Linguistic confidence
- The certainty expressed by wording, independent of evidence.
- Model probability
- A technical score from an AI system, not a measure of truth.
- Statistical confidence
- A formal measure from inferential statistics.
- Human confidence
- Subjective belief, which may be evidence-based or not.
- Evidential support
- The quality and relevance of evidence for a claim.
- Uncertainty
- The state of having limited knowledge about a claim.
18. Looking Ahead: Lesson 10.4
Lesson 10.4: TRACE, Correction, and Responsible Communication will integrate everything you have learned: claim identification (10.1), error classification (10.2), and reasoning analysis (10.3) into the full TRACE framework. You will learn to prioritise claims for verification, correct defective content without introducing new errors, and decide when to escalate, withdraw, or stop using an AI output. This final lesson prepares you for Unit 11’s detailed verification methods (CLAIM framework) and for Assignment 5.
After completing all four lessons, you will be ready to tackle the cumulative Unit 10 project and the verification tasks of Unit 11.
19. Sources & Further Reading
- Primary source: COMP 3270, Chapter 7: “Hallucination, Error, and Uncertainty” — particularly the sections on inference, correlation, precision, and uncertainty. [PAGE NUMBER TO BE CONFIRMED]
- Supporting sources:
- Appendix A: Generative AI Glossary
- Appendix C: Verification Toolkit
- Chapter 8: Verification, Evidence, and Epistemic Responsibility (introductory sections)
All examples, scenarios, and references in this lesson are synthetic. Do not rely on them as factual without independent verification.