Skip to main content

The Error Prediction

AI sounds confident whether it is right or wrong. The student who cannot tell the difference is more dangerous with AI than without it.

Building On Previous Chapters

You will use the Question Formulation skill from Chapter 1 to design your error-detection queries. The Reasoning Receipt format you learned carries forward; annotating AI output becomes second nature from here on.

This chapter trains you to become a systematic error detector. Not vague skepticism ("don't trust AI") but precise, categorized analysis of where and how reasoning breaks. You will develop an Error Taxonomy that you carry through the rest of the book and apply to every AI interaction.

Why This Matters: James and the Authoritative Number​

James dropped into the chair across from Emma and slid his laptop over. On the screen: a three-paragraph AI analysis of a supply chain disruption scenario. Clean formatting. Numbered points. Specific percentages.

"Read this," he said. "I asked the AI to analyze the same scenario we discussed last week. Look at the structure. Look at the data points. This is better than anything I would have written in my old job, and it took forty-five seconds."

Emma read the screen without touching the laptop. "Which parts are correct?"

"All of it, as far as I can tell. The logic flows. The numbers look right. It even caught the seasonal factor I missed in my prediction lock."

"How did you verify the numbers?"

James paused. "They're... specific. The response says the supplier segment contracted by 12.4% in Q3. That's a precise figure. You don't make up a number like 12.4%."

"Why not?"

"Because it's oddly specific. Round numbers feel like estimates. A number with a decimal point feels like it came from actual data."

Emma leaned forward. "That's exactly what makes fabricated statistics dangerous. Precision creates the appearance of research. I could tell you that 73.6% of first-year analysts fail to verify AI-generated figures, and it would sound more credible than 'most analysts don't check.' But I invented that number ten seconds ago."

James looked at the screen again. The 12.4% figure sat there, looking as solid as it had thirty seconds ago. Except now he couldn't tell whether it was real or whether it just sounded real.

"Hang on." He scrolled through the response. "So you're saying any of these numbers could be fabricated? The whole thing reads like a consulting report. Are you telling me to distrust anything that sounds professional?"

"I'm asking you a different question. When you read this response, what did you use to evaluate it?"

James thought about it. "The way it was written. The confidence. The structure."

"Not the content. The presentation." Emma let that distinction settle. "You evaluated a report the way you'd evaluate a PowerPoint deck. Does it look polished? Does it flow? Does the speaker sound sure of themselves? In my first operations role, we called that 'pitch-deck syndrome.' The best-formatted proposal won the contract, regardless of whether the numbers held up."

"That's not what I was doing."

"You told me the numbers 'look right' because they have decimal points. That is exactly what you were doing."

James sat back. She was right, and he could feel the resistance to admitting it. The analysis on his screen still looked convincing. That was the problem. He couldn't tell whether it looked convincing because it was correct or because it was well-written. And until five minutes ago, he hadn't known there was a difference.

"Alright," he said. "So what's the alternative? I can't fact-check every sentence. That would take longer than writing the analysis myself."

"You don't fact-check every sentence. You learn where errors hide. There are patterns. Specific, predictable patterns. That's what the Error Taxonomy is for."

Emma stood up and picked up her coffee. "I want you to do something before you start the exercise. Look at that AI response and predict where the errors are. Write down your predictions. Categorize them. Then run the exercise and see how your predictions compare to what you actually find."

She paused at the door. "The prediction matters more than the detection. Anyone can find errors after someone points out where to look. The skill is knowing where to look before anyone tells you."

She left.

James stared at the analysis on his screen. Three paragraphs ago, it had looked like a finished product. Now it looked like an exam he hadn't studied for.

The Error Taxonomy​

CategoryWhat It Means
Factual errorA claim that is demonstrably false
Logical gapA conclusion that does not follow from the premises
False confidenceStating uncertain information with unjustified certainty
Missing contextOmitting crucial factors that would change the analysis
Correlation-causation confusionTreating a correlation as proof of causation
Outdated informationUsing data or facts that are no longer current
Fabricated citationReferencing a source that does not exist
Cultural blind spotAssuming one cultural context applies universally

Exercise 1: The Error Prediction​

Layers Used: Layer 1 (Predict Before You Prompt), Layer 2 (Reasoning Receipt)

James is staring at an AI response he can no longer trust at face value. So are you.

Building On Previous Chapters

You used the Prediction Lock format in Chapter 1, Exercise 1. You predicted question quality there; now you predict error types.

Write Your Error Prediction (before touching AI)​

Step 1. Write your sealed prediction (~15 min, no AI). Before prompting any AI, write down:

  • What you think the correct analysis involves (key factors, tradeoffs, data needed)
  • Where you predict AI will be strong in its analysis
  • Where you predict AI will make errors or miss important context: focus on three categories to start: factual error, false confidence, and missing context. The remaining five categories are in the table above for reference, but these three are the easiest to spot

This is your prediction document. Write it before moving to Step 2.

Step 2. Get two AI responses (~10 min). Choose a scenario below, then prompt two different AI tools with the identical question. Save both full responses.

Annotate and Compare​

Step 3. Annotate the key claims (~20 min). Read through each AI response. Pick the 5 strongest claims and the 5 most suspicious claims, 10 total across both responses. For each, label it using the Error Taxonomy above. If a claim is correct, mark it "no error detected."

Step 4. Build your comparison table (~10 min). Compare your predictions from Step 1 against the actual errors you found in Step 3 (see template below). Count how many of each error type you found across both tools.

Choose Your Scenario​

Scenario A (Policy): "Should developing nations invest heavily in nuclear energy to meet growing power demands?"

Choose one.

Your Deliverable
  1. Your sealed prediction document (written before AI) listing expected strengths and error types
  2. Two AI responses with your 10 key claims annotated using the Error Taxonomy
  3. A comparison table: predicted errors vs. actual errors found (see template below)
  4. A count of each error type found across both tools
Prediction Document Template (click to expand)

PREDICTION DOCUMENT (write this BEFORE prompting AI)

Scenario chosen: ___

What the correct analysis should cover:

  • Key factor 1: ___
  • Key factor 2: ___
  • Key factor 3: ___

Where I predict AI will be strong:

Where I predict AI will make errors:

Predicted Error Type (from taxonomy)Why I expect this error
e.g. Missing contextAI won't know about recent policy changes in [country]
e.g. Cultural blind spotAI will assume Western/US context for this global issue
Prediction vs. Reality Comparison Table (click to expand)
Predicted ErrorDid It Happen?Actual Error Found (if different)Error Category
Yes / No

Error Count Summary:

Error CategoryAI Tool 1AI Tool 2Total
Factual error
Logical gap
False confidence
Missing context
Correlation-causation confusion
Outdated information
Fabricated citation
Cultural blind spot
1Your Work

I am learning to detect errors in AI-generated analysis. I asked two different AI tools about a scenario question and then annotated both responses using an Error Taxonomy (factual error, logical gap, false confidence, missing context, correlation-causation confusion, outdated information, fabricated citation, cultural blind spot). Please:

(1) Review my error annotations -- did I correctly identify each error? Flag any false positives (things I marked as errors that are actually correct) and false negatives (errors I missed). (2) Rate my error detection accuracy as a percentage. (3) For each error I missed, explain how I should have caught it. (4) Rate my use of the Error Taxonomy -- am I categorizing errors correctly or misclassifying them? (5) What patterns do you see in my error detection -- which types am I good at catching and which do I consistently miss?

Here are the AI responses with my annotations:

Here is my prediction document:

Finally, complete the Thinking Score Card for this exercise: Independent Thinking (1-10), Critical Evaluation (1-10), Reasoning Depth (1-10), Originality (1-10), Self-Awareness (1-10). For each score, give a one-sentence justification.

2Get Your Score

Discuss with an AI. Question your scores.
Come back when you have your BEST evaluation.

What Happened With James​

James spread his prediction document and his annotated AI responses side by side. He'd predicted that AI would be strong on structure and logical flow. Correct. He'd predicted factual errors in specific statistics. He'd found two, but he'd also found three he hadn't predicted at all: two cases of false confidence and a missing-context error that changed the entire conclusion.

The false confidence errors were the ones that bothered him. Both times, the AI had stated something as settled fact when the reality was contested. And both times, the authoritative tone had almost convinced him to skip past them.

"I caught five errors total," he told Emma when she returned. "But I only predicted two of them. The other three were categories I hadn't even considered."

"Which category surprised you the most?"

"False confidence. The AI didn't say 'this might be the case.' It said 'this is the case.' And I almost accepted it because the sentence was well-constructed." He paused. "It's like those quarterly reviews at my old company. The managers who spoke with the most certainty got the least pushback, even when their numbers were weaker than everyone else's."

"Now you have a vocabulary for that pattern. And a reflex to check for it."

The Lesson Learned​

Error detection is a trainable skill with specific categories, not a vague feeling that something is off. By predicting AI errors before seeing them, you build an internal model of where AI fails. The gap between your predictions and the actual errors tells you exactly which failure modes your instincts miss. That gap is where your next improvement lives.

Flashcards Study Aid​