How AI Uses Data

You don't need to be a data scientist to understand how AI works with data. This knowledge helps you evaluate AI outputs and make better decisions about when to trust AI-generated insights.

AI Learning: The Simple Version

Think of AI like learning to recognize cats:

TRAINING PHASE (What AI "Studies")
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📷 Millions of cat photos → 🧠 AI finds patterns → ✅ "This is what cats look like"
📷 Millions of non-cats  → 🧠 AI finds patterns → ❌ "This is NOT a cat"

APPLICATION PHASE (What AI Does)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📷 New photo → 🧠 Compare to patterns → 🐱 "95% likely a cat"

Key insight: AI doesn't "understand" cats. It recognizes patterns in data that tend to indicate "cat."

The Three Types of AI Data Usage

1. Training Data (What AI Learns From)

This is the historical data used to teach AI to recognize patterns.

Example: A customer churn prediction AI might be trained on:

• 5 years of customer purchase history
• Customer support ticket data
• Subscription renewal records
• Demographics and behavior patterns

What you should ask: "What data was this AI trained on, and is it relevant to our situation?"

2. Input Data (What AI Analyzes Now)

This is the new data you give AI to analyze.

Example: You give the churn prediction AI data about a current customer:

• Their recent purchases (down 40%)
• Support tickets (3 this month)
• Account age (2 years)

What you should ask: "Is the input data complete and accurate?"

3. Output Data (What AI Produces)

This is the prediction, recommendation, or insight AI generates.

Example: The AI outputs:

• "85% probability of churn within 30 days"
• "Recommended action: Personal outreach from account manager"

What you should ask: "Does this output make sense given what we know?"

How AI Finds Patterns

AI excels at finding patterns humans might miss, but it can also find meaningless patterns.

Pattern Type	Example	Is It Useful?
Real correlation	Customers who call support 3+ times are 5x more likely to cancel	✅ Yes—actionable
Spurious correlation	Sales increase on days the CEO wears blue	❌ No—coincidence
Hidden bias	AI recommends men for promotions more often (trained on historical data with gender bias)	⚠️ Dangerous—reflects past discrimination

AI Limitations Every Professional Should Know

1. AI Only Knows What It Was Trained On

Scenario	Problem
AI trained on US data	May not work well for Asian markets
AI trained on 2019 data	Doesn't know about post-pandemic behaviors
AI trained on large companies	May give poor advice for startups

2. AI Doesn't Understand Context

AI sees correlations, not causation or context.

Example: An AI might predict that employees who work long hours are more productive. But it doesn't understand:

• They might be working long hours because they're struggling
• Burnout could be destroying long-term productivity
• The "productivity" metric might be flawed

3. AI Can Be Confidently Wrong

AI often provides confidence scores (like "85% certain"), but:

• High confidence doesn't mean correctness
• AI can be very confident about very wrong answers
• Confidence scores should prompt verification, not replace it

The AI-Data Quality Connection

Remember the DAMA data quality dimensions from Module 2? They matter even more with AI.

Quality Issue	Impact on AI
Inaccurate data	AI learns wrong patterns, makes wrong predictions
Incomplete data	AI misses important factors, gives partial picture
Outdated data	AI gives advice based on past that no longer applies
Biased data	AI amplifies and perpetuates existing biases

The Iron Rule: AI trained on garbage data produces garbage outputs—just more confidently and at scale.

Questions to Ask About AI Outputs

Before trusting any AI-generated insight, ask:

DATA QUESTIONS
□ What data was this AI trained on?
□ When was the training data collected?
□ Is the training data representative of our situation?

RELEVANCE QUESTIONS
□ Does the AI's domain match our use case?
□ Were there any major changes since training (market shifts, pandemic, etc.)?
□ Is the input data we provided complete and accurate?

OUTPUT QUESTIONS
□ Does this output make logical sense?
□ Does it align with what we know from other sources?
□ What would happen if the AI is wrong?

Real-World Example: Marketing AI

Scenario: An AI tool recommends increasing ad spend on Facebook by 200% because it predicts 3x ROI.

Critical thinking questions:

1. What historical data was this trained on? (Answer: Pre-iOS 14 privacy changes)
2. Has anything changed since then? (Answer: Yes—Facebook tracking is now limited)
3. What's the cost if this prediction is wrong? (Answer: Significant wasted budget)

Decision: Test the recommendation with a small budget increase first before committing to 200%.

Key Insight: AI is a powerful tool for finding patterns in data, but it doesn't replace human judgment. Understanding how AI uses data helps you leverage its strengths while protecting against its weaknesses.

Next: Learn how to verify AI outputs for accuracy, bias, and reliability. :::