What over 2.5 Billion Daily Messages Reveal About ChatGPT Use

Evidence from a Harvard, Duke, and OpenAI analysis of real-world interactions (mid-2025)

Usage Patterns at a Glance

> 2.5B
Daily Messages
By mid-2025, users sent about 2.5 billion messages per day.
29,000
Messages per Second
Rough throughput across the day.
> 700M
Weekly Active Users
By mid-2025, roughly 10% of the world’s adults use ChatGPT weekly.
Over 70%
Non‑Work Share
Up from 53% a year earlier.
≈10%
Tutoring Content
About one in ten messages involve tutoring.
56%
Doing at Work
Share of work messages that ask the model to perform a task.
52%
Management Writing Focus
Share of management work messages that are writing‑related.
16%
66+ Work Mix
Share of messages that are work‑related among users aged 66+.
80%
Early Adopter Skew
Early users had typically masculine names; the gap has since closed.
1M+
Study Scale
Conversations analyzed with privacy‑preserving methods.
100+ users
Clean Room Threshold
Only aggregated results for groups of 100 or more.

How We Interact: Asking vs Doing vs Expressing

Nearly half of all messages are 'Asking' (49%), seeking advice or information; 40% are 'Doing', where the model is asked to perform tasks; and 11% are 'Expressing'. This split shows that most people use ChatGPT to inquire and get help on specific tasks rather than to simply vent or reflect.

Topic Shifts Over Time

Seeking Information grew from 14% to 24%, while Writing declined from 36% to 24%. This indicates a shift toward using ChatGPT as a conversational search and explanation engine, even as writing remains a major use.

Writing Tasks: Edit vs Create

About two-thirds of writing tasks involve modifying user-provided text, with one-third creating content from scratch. Users are leveraging ChatGPT more as an editor, translator, and critiquer than as a sole author.

Low-Frequency Uses (Myths vs Reality)

Contrary to common narratives, programming accounts for just 4.2% of messages, relationships/personal reflection 1.9%, and games/role-play 0.4%. These are notable but niche compared to everyday practical and informational use.

Common Work Activities (O*NET Mapping)

In work contexts, tasks concentrate on knowledge handling: Getting Information (19.3%), Interpreting Information (13.1%), and Documenting/Recording (12.8%). This aligns with ChatGPT’s role as a decision co‑pilot that supports analysis and communication.

Key Insights

ChatGPT is used more for everyday guidance and information than for coding or companionship. Asking and Doing dominate interaction styles; at work, knowledge and writing tasks lead. Writing support is primarily editorial, not generative, underscoring AI’s augmentation role. Non‑work use has surged, pointing to broad consumer value beyond the workplace.

Busting the Myths

Myth
1
Myth 1: “Everyone’s using AI to code”
Reality check on programming usage.
Reality check on programming usage.
Myth
2
Myth 2: “AI is our therapist or companion”
Assessing relationship and role‑play content.
Assessing relationship and role‑play content.
Myth
3
Myth 3: “AI replaces human creativity”
What writing tasks actually look like.
What writing tasks actually look like.

What People Really Do with ChatGPT

Top Use
1
Practical Guidance
Around 29% of interactions: personalized advice, ideas, and help for everyday tasks. Tutoring alone is roughly 10% of all messages.
Around 29% of interactions: personalized advice, ideas, and help for everyday tasks. Tutoring alone is roughly 10% of all messages.
Rising
2
Seeking Information
Grew from 14% to 24%: conversational search with context.
Grew from 14% to 24%: conversational search with context.
Core
3
Writing
Declined from 36% to 24%: drafting and especially editing communications.
Declined from 36% to 24%: drafting and especially editing communications.

Conversation Topics and User Intent

29%
Practical Guidance Share
Most common topic in ChatGPT conversations.
40%
Writing in Work-Related Messages
Share of work-related messages focused on Writing.
4.2%
Technical Help (Programming)
Share of messages related to computer programming.

Distribution of Conversation Topics (All Messages)

This bar chart shows the distribution of high-level conversation topics in ChatGPT messages as of June 2025. Practical Guidance, Seeking Information, and Writing together account for 77% of all usage, with Writing declining over time and Seeking Information increasing.

User Intent Breakdown

This pie chart illustrates user intent in ChatGPT messages: 49% are Asking (seeking information or advice), 40% are Doing (requesting output or task completion), and 11% are Expressing (sharing views or feelings).

Share of Writing Subcategories

This chart breaks down Writing conversations into subcategories. Editing/Critiquing, Translation, and Argument/Summary Generation (modifying user text) make up two-thirds of Writing messages, indicating that most writing requests involve refining existing content.

Key Insights

Practical Guidance, Seeking Information, and Writing dominate ChatGPT usage. Most writing requests involve editing or modifying user-provided text. The platform is used primarily for information-seeking and decision support, with technical help and self-expression representing smaller shares.

Occupation-Based Usage Patterns

57%
Highest Work-Related Usage
Share of work-related messages among computer/math occupations.
52%
Writing in Management/Business
Share of work-related messages focused on Writing in management/business occupations.

Work-Related Message Share by Occupation

This chart compares the share of work-related messages across broad occupation categories. Computer/Math and Management/Business users are most likely to use ChatGPT for work purposes.

Distribution of Work-Related Conversation Topics by Occupation

Writing is especially common in management/business and non-professional occupations, while technical help is most prevalent among computer-related users.

Key Insights

Work-related ChatGPT usage is highest among computer/math and management/business professionals. Writing tasks dominate in management and non-professional roles, while technical help is concentrated in technical occupations.

Demographic Patterns in ChatGPT Usage

46%
Share of Messages by Young Adults
Messages sent by users aged 18-25.
2025
Gender Parity Achieved
Year when active users with feminine names outnumbered those with masculine names.

Gender Distribution Over Time

This line chart shows the shift in gender distribution among active ChatGPT users. Initially, 80% of users had typically masculine names, but by June 2025, feminine names slightly outnumber masculine, indicating a closing gender gap.

Work-Related Message Share by Education Level

This chart shows that work-related message share increases with education: 37% for users with less than a bachelor's degree, 46% for bachelor's, and 48% for graduate education.

Key Insights

ChatGPT usage has become more gender-balanced over time. Young adults are the most active users, and higher education correlates with increased work-related usage. These patterns reflect the platform's broadening appeal and integration into professional and personal contexts.

Privacy‑First Research Pipeline

1
Global Sample
Analyze over a million conversations across available regions.
2
De‑Identification
Remove personal information before any classification.
3
Automated Classification
AI models classify messages; no human reviews raw content.
4
Clean Room Queries
Run pre‑approved queries in a secure environment; results only return for groups of 100+ users.
5
Aggregated Patterns
Report population‑level trends while protecting individual privacy.

Why This Approach

Reason
1
Scale
Surveys can’t capture billions of real interactions; automated analysis reflects actual behavior.
Surveys can’t capture billions of real interactions; automated analysis reflects actual behavior.
Reason
2
Privacy
No raw messages exposed; clean‑room queries and thresholds protect individuals.
No raw messages exposed; clean‑room queries and thresholds protect individuals.

Limits and Caveats

1
Coverage
Consumer users only; excludes business/education accounts.
Consumer users only; excludes business/education accounts.
2
Age and Auth
Under‑18s and logged‑out users excluded.
Under‑18s and logged‑out users excluded.
3
Classifier Errors
Automated classifiers can err, especially in nuanced or multilingual chats.
Automated classifiers can err, especially in nuanced or multilingual chats.
4
Model Drift
Model changes over time can influence behavior and classifications.
Model changes over time can influence behavior and classifications.
5
Demographic Inference
Some attributes (e.g., gender) inferred from names and are imperfect.
Some attributes (e.g., gender) inferred from names and are imperfect.
6
Publication Status
Working paper; not yet peer‑reviewed.
Working paper; not yet peer‑reviewed.

Real‑World Applications and Impact

🏢
Organizations
Prioritize decision‑support workflows and writing assistance for better analysis and communication.
🎓
Education
Leverage tutoring (≈10% of all messages) while teaching AI literacy: verify, evaluate, iterate.
🌍
Society & Policy
Consumer surplus extends to home production; support equitable, multilingual access as global adoption grows.

Quality of Interactions and User Satisfaction

4:1
Good-to-Bad Ratio (July 2025)
Good interactions outnumber bad ones by more than four to one.

Interaction Quality: Good vs Bad Responses Over Time

The share of 'Good' interactions (user satisfaction) has increased over time, with good responses now more than four times as common as bad ones.

Good-to-Bad Ratio by Topic

Self-Expression topics receive the highest satisfaction, while Multimedia and Technical Help have the lowest good-to-bad ratios.

Key Insights

User satisfaction with ChatGPT's responses has improved over time, especially in expressive and creative domains. Technical and multimedia queries see lower satisfaction, highlighting areas for further improvement.

Questions to Consider

Reflect
1
Measuring Value
If AI’s biggest impact is on unpaid work and personal tasks, how should we measure its economic value?
If AI’s biggest impact is on unpaid work and personal tasks, how should we measure its economic value?
Reflect
2
Changing Expertise
As AI improves decision support, how will the nature of human expertise evolve?
As AI improves decision support, how will the nature of human expertise evolve?
Reflect
3
Ubiquity of AI Advisors
What happens when AI advisors become as common as web searches?
What happens when AI advisors become as common as web searches?

Co‑Pilot, Not Autopilot

The data show that ChatGPT is woven into daily life as a digital Swiss Army knife. People rely on it for guidance and information more than for niche tasks like coding or role‑play. In writing—especially at work—AI augments rather than replaces creativity: two‑thirds of tasks refine what users already wrote. Treating AI as a thinking partner aligns with these patterns: ask better questions, compare options, and use the model to clarify and communicate more effectively.