The NBER working paper “How People Use ChatGPT” (September, 2025) provides unprecedented insights into how ChatGPT is actually used globally.

Citation: Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., & Wadman, K. (2025). How people use ChatGPT (NBER Working Paper No. 34255). National Bureau of Economic Research. https://doi.org/10.3386/w34255

Here are some takeaways and their methodology:

Takeaways

Scale and Growth

ChatGPT achieved extraordinary adoption: By July 2025, it reached 700 million weekly active users (10% of the global adult population) and processed 2.5 billion daily messages. This represents the fastest technology diffusion in history, surpassing even the internet's early adoption rates.

Non-Work Usage Dominates

Personal use grew faster than work use: Non-work messages increased from 53% to 73% between June 2024 and June 2025. This challenges the dominant narrative that AI's primary economic impact comes through workplace productivity.

Consumer surplus is massive: The authors cite Collis and Brynjolfsson's estimate that US users would need $98 compensation to give up generative AI for a month, implying $97 billion in annual consumer surplus.

Three Primary Use Cases

Nearly 80% of usage falls into three categories:

1. Practical Guidance (29%): Personalized advice like custom workout plans, tutoring (10% of all messages are educational)
2. Seeking Information (24%): Factual queries similar to web search
3. Writing (24%): Most common work task, but two-thirds involves editing existing text rather than creating new content Programming is surprisingly small: Only 4.2% of ChatGPT messages involve coding, contradicting assumptions about AI's primary technical applications.

Decision Support vs Task Automation

"Asking" dominates over "Doing": 49% of messages seek advice/information ("Asking") versus 40% requesting task completion ("Doing"). Among work messages, Asking messages received higher quality ratings and grew faster.

Knowledge work applications: 81% of work-related messages involve "obtaining/interpreting information" or "making decisions and solving problems" - suggesting ChatGPT functions more as a research assistant than task automator.

Demographic Patterns

Gender gap closed dramatically: From 80% male users in early 2023 to 52% female users by June 2025, representing complete reversal of the gender disparity.

Youth dominance: 46% of messages come from users under 26, though this concentration has slightly decreased over time.

Global expansion in lower-income countries: Adoption grew fastest in countries with $10,000-40,000 GDP per capita, indicating AI access is democratizing globally.

Education and occupation matter for work use: Users with graduate degrees are 48% likely to send work-related messages versus 37% for those without bachelor's degrees. Professional occupations show 50-57% work usage versus 40% for non-professional roles.

Quality and Satisfaction

User satisfaction improved over time: "Good" interactions became 4x more common than "Bad" by July 2025, up from 3x in late 2024.

Writing tasks dominate professional use: 40% of work messages involve writing, with management/business users reaching 52%. This reflects writing as a universal white-collar skill.

Economic Implications

Home production impact equals workplace impact: The dominance of non-work usage suggests AI's economic value extends far beyond measured workplace productivity into unmeasured household efficiency and welfare.

Decision-making enhancement in knowledge work: The prevalence of information-seeking and advisory use cases indicates ChatGPT primarily augments human judgment rather than replacing human tasks, particularly valuable in knowledge-intensive occupations where better decisions drive productivity.

These findings fundamentally reframe how we should think about AI's economic impact - less about job displacement and automation, more about enhanced decision-making and consumer welfare across both work and personal contexts.

Methodology

The researchers used privacy-preserving methodology to analyze ChatGPT usage data. Here's how they did it:

Three Primary Datasets

1. Growth Dataset

• Source: All consumer ChatGPT plans (Free, Plus, Pro) from November 2022 to September 2025

• Content: Daily message counts, user metadata (country, age in 5-7 year buckets, subscription type)

• Exclusions: Business/Enterprise users, under-18 users, opted-out users 2. Classified Messages Dataset

• Sample size: ~1.1 million randomly selected conversations from May 2024-June 2025

• Method: One message per conversation, weighted to reflect total daily volume

• Additional subset: 130,000 users with up to 6 messages each for detailed analysis 3. Employment Dataset

• Source: External vendor with publicly available employment data

• Size: ~130,000 users matched to occupation/education categories

• Restrictions: Aggregated data only, minimum 100 users per category

Privacy-Preserving Classification Pipeline

The critical innovation: No human researcher ever saw actual user messages. Instead, they used automated LLM classifiers:

Step 1: PII Removal

• Messages first processed through "Privacy Filter" tool to remove personally identifiable information Step 2: Automated Classification

• Five different LLM-based classifiers applied to each message:

  • Work vs. Non-work
  • Asking/Doing/Expressing intent
  • Conversation topics (24 categories)
  • O*NET work activities (332 categories)
  • Interaction quality

Step 3: Context Inclusion

• Classifiers considered up to 10 previous messages for context
• Messages truncated to 5,000 characters maximum
• Used GPT-5-mini for most classifications, GPT-5 for interaction quality

Data Clean Room for Employment Analysis

Secure separation: Employment data held by external vendor, never directly accessed by researchers

Query restrictions:

• Only pre-approved aggregate queries allowed
• Minimum 100 users required for any reported category
• Committee approval required for each analysis Privacy controls: Individual records never visible, only statistical summaries above threshold

Validation Methodology

Human validation: Researchers validated classifiers using WildChat dataset (publicly available third-party chatbot conversations)

Agreement metrics:

• Work classification: 83% model-human agreement (κ=0.83)
• Intent classification: 74% agreement (κ=0.74)
• Topic classification: 56% agreement (κ=0.56) Quality checks: Cross-referenced automated sentiment analysis with actual user thumbs-up/down feedback on 60,000 interactions

Sampling and Weighting

Representative sampling: Messages weighted to reflect actual daily volume patterns, not just random selection

Exclusion criteria consistently applied:

• Users who opted out of training data sharing
• Self-reported under-18 users
• Deleted conversations and banned accounts
• Logged-out users (for consistency, though available from March 2025)

Methodological Strengths and Limitations

Strengths:

• Unprecedented scale and access to actual usage data

• Strong privacy protections exceeding academic standards

• Multiple validation approaches

• Global, representative sample Limitations:

• Consumer plans only (excludes enterprise users who might have different patterns)

• Classification accuracy varies by task

• External employment matching limited to subset of users

• Potential biases in automated classification, though validated against human judgment This methodology represents a significant advance in studying digital behavior while maintaining privacy. The automated classification approach could become a template for analyzing sensitive user data in other contexts.