Step-by-Step Teaching

Step 1: What Problem Are We Solving?

The core question: How do we actually measure how people use AI for work?

Before this data release, researchers had limited tools to answer:

• What kinds of work tasks do people bring to AI?
• Does AI use differ between countries?
• Is AI use becoming more or less specialized over time?

Key takeaway: Without structured data, these questions are impossible to answer rigorously.

---

Step 2: What Is the Dataset?

OpenAI released what they call "OpenAI Signals" data.

⚠️ Important exclusion: Enterprise messages and Codex use are not included. This matters when interpreting results.

---

Step 3: Understanding the Framework — What Are IWAs?

To make sense of the data, OpenAI used a standardized system called Intermediate Work Activities (IWAs).

What are IWAs?

• IWAs come from O*NET, a well-established occupational framework
• They group occupation-specific tasks into broader categories shared across many jobs
• Example categories include:
  • Preparing instructional materials
  • Developing marketing content
  • Programming
  • Analyzing business data

Why does this matter?

Using a common framework allows researchers to:

1. Compare ChatGPT usage with real-world employment data
2. Connect AI usage to existing economic indicators
3. Track changes consistently over time

🧠 Think of it this way: IWAs are like a universal language for describing work — they let you compare a teacher's tasks with a marketer's tasks on the same scale.

---

Step 4: Key Finding #1 — Work Use Is Becoming Less Concentrated

The finding:

Across 74 countries, the top 10 activities dropped from 56.0% of work messages in Q1 2025 to 51.1% in Q3 2026.

What does this mean?

Q1 2025:  [████████████████████░░░░░░░░░░░░░░░] 56% → Top 10 activities dominate
Q3 2026:  [██████████████████░░░░░░░░░░░░░░░░░] 51% → More spread across activities

• People are using ChatGPT for a wider variety of work tasks over time
• Countries are specializing in different types of work-related AI use
• This trend suggests AI adoption is maturing and diversifying

---

Step 5: Key Finding #2 — ChatGPT Use ≠ The Actual U.S. Economy

When researchers compared ChatGPT work usage to actual U.S. job activities, they found substantial differences.

Why the gap?

Approximately 43% of U.S. employment-weighted tasks fall into categories like:

• Moving materials
• Operating vehicles
• Maintaining health/medical records

These are physical or highly regulated tasks that people do NOT bring to ChatGPT — and they were excluded from the reported categories.

The lesson:

ChatGPT usage data reflects what people choose to use AI for, not a mirror of the entire economy. Physical and hands-on work remains largely outside AI chat tools.

---

Step 6: How Can Researchers Use This Data?

The dataset enables several research directions:

---

Step 7: Privacy Protections — How Is This Done Ethically?

A critical component of this research is user privacy. OpenAI uses several safeguards:

1. ✅ Automated classification systems (no human reads conversations)
2. ✅ Identifying information is removed before classification
3. ✅ Only aggregated statistics are published
4. ✅ Results for small groups are withheld
5. ✅ Statistical noise is added to protect individual messages

🔐 This is an example of privacy-preserving research design — balancing scientific value with ethical responsibility.

---

Summary: The Big Picture

PROBLEM:  How do we measure AI's role in work?
              ↓
SOLUTION: Connect ChatGPT usage → IWA framework (O*NET)
              ↓
DATA:     165 activities, ~100 countries, Q1 2025–Q3 2026
              ↓
FINDINGS: • AI work use is diversifying over time
          • Countries specialize differently
          • ChatGPT ≠ full picture of economy (physical work excluded)
              ↓
PURPOSE:  Enable researchers to study AI + labor economics rigorously

---

Check Your Understanding

1. What is an IWA and why is it useful for this research?
2. What does it mean that work use became "less concentrated" over time?
3. Why does ChatGPT usage data NOT perfectly represent the U.S. economy?
4. Name two privacy protections used in this dataset.
5. What types of work are notably absent from ChatGPT usage data, and why?