Critical Evaluation

Don’t Trust Blindly

In the previous lesson, we explored human judgment and ai limits. Now let’s build on that foundation. AI sounds confident. AI uses authoritative language. AI presents information as fact.

AI is also frequently wrong.

Critical evaluation isn’t being cynical—it’s being responsible. The same technology that produces valuable outputs also produces convincing nonsense.

The Hallucination Problem

What are hallucinations? AI generating information that sounds true but isn’t. Made-up facts, citations that don’t exist, details that are completely fabricated.

Why do they happen? AI is predicting what text should come next, not retrieving verified information. It doesn’t “know” things—it generates plausible-sounding sequences.

The danger: Hallucinations look exactly like accurate information. AI doesn’t flag uncertainty. It presents everything with the same confidence.

Examples:

• Legal briefs citing non-existent court cases
• Academic papers with fabricated sources
• Medical information that’s dangerously wrong
• Historical “facts” that never happened
• Code that looks right but doesn’t work

Categories of AI Error

Factual errors:

• Wrong dates, names, numbers
• Events that didn’t happen
• Misattributed quotes
• Incorrect technical details

Logical errors:

• Flawed reasoning
• Non-sequiturs that sound connected
• Circular arguments
• Missing steps in logic

Bias and stereotypes:

• Reflecting historical discrimination
• Stereotypical associations
• Missing perspectives
• One-sided presentation

Outdated information:

• Training data has a cutoff
• AI doesn’t know recent events
• Old best practices presented as current

Context misunderstanding:

• Missing nuance of your specific situation
• Generic advice that doesn’t apply
• Ignoring constraints you didn’t state

The Verification Workflow

For any consequential AI output:

1. Identify claims that matter What in this output, if wrong, would cause problems?

2. Verify independently Check claims against authoritative sources outside of AI.

3. Consider the source Would this AI have been trained on reliable information about this topic?

4. Cross-reference Does this match what you know from other reliable sources?

5. Check citations If AI provides sources, verify they exist and say what AI claims.

Verification Methods

For factual claims:

• Check primary sources when possible
• Use authoritative references (official documents, established institutions)
• Be suspicious of specifics (dates, quotes, statistics)

For code and technical content:

• Actually run the code
• Check documentation
• Test edge cases
• Don’t assume it works because it looks right

For analysis and reasoning:

• Follow the logic step by step
• Check if conclusions follow from premises
• Look for missing considerations

Quick check: Before moving on, can you recall the key concept we just covered? Try to explain it in your own words before continuing.

For creative content:

• Check for plagiarism if originality matters
• Verify accuracy of any embedded facts
• Review for harmful stereotypes or misinformation

Red Flags

Be extra skeptical when:

• Very specific numbers or statistics (easy to hallucinate)
• Quotes (often fabricated or misattributed)
• Citations (frequently don’t exist)
• Claims about recent events (training data cutoffs)
• Anything that sounds too perfect or convenient
• Legal, medical, or financial specifics
• Edge cases or unusual scenarios

The Confidence Trap

AI confidence ≠ AI accuracy.

AI says: “The Supreme Court ruled in Smith v. Jones (2019) that…” Sounds authoritative. But there might be no such case.

AI says: “Studies show that 73% of users prefer…” Precise. Specific. Possibly completely made up.

Certainty in AI output tells you nothing about accuracy. Treat all AI output with appropriate skepticism.

Building Verification Habits

For quick, low-stakes outputs:

• Quick plausibility check
• Match against your existing knowledge
• Spot-check anything that will be shared

For important outputs:

• Full verification of key claims
• Independent source checks
• Have someone else review
• Don’t rush—errors have costs

For anything affecting others:

• Highest scrutiny
• Multiple verification methods
• Consider what happens if it’s wrong
• Don’t be the person who passed along AI misinformation

Exercise: Verification Practice

Ask AI a question about a topic you know well.

1. Review the output for accuracy
2. Identify any errors (you may find some!)
3. Note how confident the AI sounds, even when wrong
4. Consider: How would you have caught errors on a topic you don’t know well?

This exercise builds appropriate skepticism.

The Responsibility Shift

When you act on AI output, you own the consequences.

“The AI told me” doesn’t exempt you from:

• Spreading misinformation
• Making bad decisions
• Causing harm through incorrect information
• Professional consequences of errors

Verification is your responsibility.

Key Takeaways

• AI confidently produces false information (hallucinations)
• Confidence in AI output tells you nothing about accuracy
• Categories of error: factual, logical, biased, outdated, context misunderstanding
• Verify consequential claims with independent authoritative sources
• Be extra skeptical of: specific numbers, quotes, citations, recent events
• Build verification habits scaled to the stakes
• When you act on AI output, you own the consequences

Next: Building responsible practices into your AI workflow.

Up next: In the next lesson, we’ll dive into Responsible Practices.