How People Use ChatGPT

Executive Summary

This paper measures how people use ChatGPT at consumer scale and over time, without humans reading any user messages. By July 2025, ChatGPT reached ~700M weekly active users sending ~18B messages/week; usage is increasingly non-work, and the most common topics are Practical Guidance, Seeking Information, and Writing. The authors argue ChatGPT’s economic value stems from decision support, especially in knowledge-intensive work.

Key Technical Advancements

Privacy-preserving automated classification pipeline: Messages are scrubbed with an internal Privacy Filter and then labeled by LLMs across several taxonomies (work vs. non-work, topic, Asking/Doing/Expressing intent, and O*NET activity). No researcher viewed message text.

Conversation topic taxonomy: A 24-category classifier aggregated into seven topics reveals that Practical Guidance, Seeking Information, and Writing account for ~77–80% of conversations; Writing is the top work topic (~40%).

Intent rubric (Asking/Doing/Expressing): ~49% Asking, 40% Doing, 11% Expressing overall; at work, Doing ≈56% with Writing comprising nearly one-third of work messages.

O*NET mapping of work activities: Work usage concentrates on obtaining, documenting/interpreting information, making decisions/solving problems, and thinking creatively, with similar patterns across occupations.

Secure employment/education linkage: Aggregated demographics were analyzed only inside a data clean room with strict ≥100-user thresholds.

Practical Implications & Use Cases

Technical impact: Expect rising demand for decision-support workflows (Ask → reason → advise) and editing pipelines: two-thirds of Writing requests are to modify user text (edit, critique, translate) rather than generate from scratch. Tooling that makes “paste-edit-refine-export” seamless will see outsized use.

UX/UI implications: Since Asking dominates and rates higher in user feedback than Doing, optimize for clarifying follow-ups, comparisons, and rationale surfacing; add affordances for customization and iteration on advice and plans.

Strategic implications: Non-work usage grew from 53%→73% (Jun 2024→Jun 2025), highlighting consumer surplus beyond workplace productivity. Prioritize features that enhance everyday guidance, search-adjacent lookup, and writing assistance, and consider growth in low-/middle-income countries and among younger users.

Challenges and Limitations

Sample scope: Analysis covers consumer plans (Free/Plus/Pro); Business/Enterprise/Education plans and logged-out users are excluded for most of the period—limit generalization to enterprise contexts.

Automated labeling: Results depend on LLM classifiers (validated against public data), which can introduce classification errors despite strong alignment with human judgments.

Aggregation & privacy thresholds: Demographic insights are coarse and thresholded (≥100 users), which prevents fine-grained reporting but protects privacy.

Future Outlook & Considerations

The data show Asking is growing faster than Doing and interaction quality (good:bad) improved from ~3:1 to >4:1 by July 2025—suggesting continued gains from reasoning, explanation, and follow-up UX. Teams should evaluate decision-support and writing-assist features, and consider privacy-first analytics approaches mirrored here for trustworthy measurement.

Conclusion

ChatGPT is primarily used for decision support and writing, not coding or companionship. Practical Guidance, Seeking Information, and Writing dominate usage; Writing leads at work, and non-work use is now the majority. For product and design leaders, this points to investing in guided advice, editing workflows, and transparent reasoning; for engineers, it underscores the value of classification and safety layers that enable privacy-respecting insights at scale.