PaperlayerView exact revision on arXiv

Source-linked AI summary

How People Use ChatGPT: Conversation-Level Evidence from India, Nigeria, Brazil, and Pakistan

Shreyasi Roy Chowdhury, Kiran Garimella

arXiv:2609.38279v1cs.CYcs.HCcs.SI

TL;DR

Platform reports use fixed taxonomies and inferred demographics and release only aggregates, leaving no conversation-level, demographically grounded view of ChatGPT use across Global South countries. The authors analyze 202,590 complete ChatGPT conversations from 1,252 users in India, Nigeria, Brazil, and Pakistan, paired with self-reported age and gender, using platform classifiers, unsupervised topic discovery, and qualitative reading. Personal use dominates in every country and coursework is about as common as work; local topic discovery surfaces country-specific uses, while Expressing rose to roughly a fifth of conversations in every country.

  • Problem

    Platform reports use fixed taxonomies and inferred demographics and release only aggregates, leaving no conversation-level, demographically grounded view of ChatGPT use across Global South countries.

    Source: §1 · p. 1

  • Method

    The authors analyze 202,590 complete ChatGPT conversations from 1,252 users in India, Nigeria, Brazil, and Pakistan, paired with self-reported age and gender, using platform classifiers, unsupervised topic discovery, and qualitative reading.

    Source: §1 · p. 2, §4 · p. 5

  • Results

    Personal use dominates in every country and coursework is about as common as work; local topic discovery surfaces country-specific uses, while Expressing rose to roughly a fifth of conversations in every country.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    Personal use dominates in every country, and coursework is about as common as work, showing workplace-productivity narratives capture only part of these users’ activity.

    Source: §1 · p. 2

    The convenience sample consists of digitally literate Clickworker users willing to share their histories, so prevalence estimates describe this population rather than national populations.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Public understanding of how people use LLM-based conversational AI assistants comes primarily from aggregate platform reports by OpenAI and Anthropic, which apply fixed taxonomies and inferred demographics to hundreds of millions of users and release only summary statistics that outside researchers cannot re-analyze. We provide a complementary, conversation-level view: complete ChatGPT exports comprising 202,590 conversations from 1,252 users across India, Nigeria, Brazil, and Pakistan, paired with self-reported age and gender and spanning December 2022 to February 2026. To our knowledge this is the first conversation-level, demographically grounded comparison of ChatGPT use across multiple non-Western countries. We ask what these users use ChatGPT for (purpose), what they talk about (topics), and how they engage with it (mode of interaction), using the platform's own classifiers, unsupervised topic discovery, and a thematic analysis of expressive conversations. Personal use accounts for 55-64% of conversations in every country and coursework is about as common as work, so workplace productivity describes a minority of use. Unsupervised topic discovery surfaces country-specific uses that the OpenAI taxonomy folds into generic categories: health and wellness in India and Brazil, Urdu-English translation in Pakistan, current affairs in Nigeria, religious questions in Nigeria and Pakistan, and self-reflection in Brazil. Over three years, the share of conversations that seek information declined only modestly and the share that delegate a task did not grow, while conversations in which users express themselves rose from a few percent to roughly a fifth or more in every country. The same product is thus attached to different local needs in each country, and understanding what adoption means requires conversation-level, country-sensitive measurement alongside global aggregates.

1 Introduction

This study addresses the lack of conversation-level, demographically grounded evidence about ChatGPT use in the Global South by analyzing complete exports from four countries. It finds that use extends beyond workplace productivity, varies with local needs, and increasingly includes expressive conversations.

1 Introduction

  • Because conversational AI supports disclosure and multi-turn interaction, measuring use requires attention to expressive engagement as well as topics and tasks.The authors argue that global taxonomies also need bottom-up discovery to reveal what categories contain in a given country.Source: §1 · pp. 1–2
  • Prior evidence is largely platform-reported and aggregate, with fixed taxonomies and inferred demographics, leaving no conversation-level, demographically grounded view of ChatGPT use in the Global South.The study asks what users use ChatGPT for, what they discuss, and how they engage with it.Source: §1 · p. 2
  • The study analyzes complete ChatGPT exports from 1,252 users in India, Nigeria, Brazil, and Pakistan, paired with self-reported age and gender.The corpus contains 202,590 conversations spanning December 2022 through February 2026; it is a convenience sample, so estimates describe this population rather than national populations.Source: §1 · p. 2
  • Personal use dominates in every country, while coursework is about as common as work, making workplace productivity a minority account of use.The analysis pairs task-purpose classification with self-reported demographic information.Source: §1 · p. 2
  • Unsupervised topic discovery surfaces locally prevalent uses that the global taxonomy groups into generic categories, including health, translation, religion, online earning, and self-reflection.The study pairs the 24-category taxonomy with country-specific topic discovery.Source: §1 · p. 2
  • Expressing rose to roughly a fifth of conversations in every country, while information seeking declined only modestly and task delegation did not grow.The rise also appears within individual users, rather than only reflecting changes in who was active.Source: §1 · p. 2

2 Related Work

Prior work describes LLM use through platform aggregates, workplace-focused studies, public conversation corpora, and research on affective or culturally situated AI. This paper positions conversation-level donation data from four Global South countries as a complement to those approaches.

Measuring real-world LLM use at scale

  • Platform studies classify large-scale use with fixed categories but release aggregate results rather than conversations, while public corpora rely on self-selected users and interfaces.Data donation offers a different route by collecting users’ own platform records with consent.Source: §2 · pp. 2–3

Work, coursework, and everyday assistance

  • Workplace-impact research measures productivity and occupational exposure, but conversation-level prevalence of non-market uses such as education and household assistance remains unmeasured.Related work treats household production and free digital goods as value that standard market accounting may not capture.Source: §2 · pp. 2–3

Geography, language, and culturally situated AI use

  • Research on geography, language, and culturally situated AI describes linguistic and cultural differences, while country-level reports still use fixed taxonomies and leave conversation-by-conversation use undocumented.The cited literature also discusses model failures involving religion, food, and naming.Source: §2 · p. 3

Affective, relational and disclosive use of conversational AI

  • Affective-use research examines relationship-focused products or affect detected at scale, leaving ordinary disclosures within general-purpose assistant conversations less studied.Examples include users raising health worries or religious questions while asking for assistance.Source: §2 · p. 3

Synthesis

  • Across these strands, evidence remains aggregate, workplace-centred, or platform-defined and is drawn almost entirely from high-income countries.The paper addresses this gap with conversation-level, demographically grounded data from India, Nigeria, Brazil, and Pakistan.Source: §2 · p. 3

3 Dataset

The dataset comprises complete, de-identified ChatGPT exports from recruited users in India, Nigeria, Brazil, and Pakistan, paired with age and gender self-reports. It offers longitudinal conversation-level coverage, but its convenience sample does not support country-wide prevalence estimates.

3.1 Data Collection

  • Participants were recruited through Clickworker and uploaded complete ChatGPT exports after a local client-side script stripped names and contact details.They provided age and gender, were compensated in accordance with local minimum wage, and the study received IRB approval.Source: §3.1 · p. 3
  • The corpus contains 202,590 conversations from 1,252 users, spanning December 2022 through February 2026.India contributes the largest share of users and conversations; Nigeria has the highest median engagement at 91 conversations per user.Source: §3.1 · p. 3
  • The sample’s gender and age composition varies across countries: India and Pakistan skew young and male, while Brazil is older and has more women than men.The dataset overview and demographic table summarize country-level coverage and user composition, respectively.Source: §3.1 · p. 3, Table 1 · p. 4, Table 2 · p. 4
  • The sample is a convenience sample of active, digitally literate Clickworker users willing to share their histories, not a demographically representative sample of national populations.Prevalence estimates are conditional on this population; fewer participants hold salaried office jobs, which may affect the observed work share.Source: §3.1 · p. 4

4 Methods

The study applies platform-derived classifiers alongside country-specific unsupervised topic discovery and qualitative reading to analyze purpose, topics, and interaction mode. It also examines within-user trends, demographic differences, and language, with validation and sampling choices shaping how results should be interpreted.

4 Methods

  • Four classifiers and language detection measure topic, country-specific clusters, intent, and task purpose; each research question uses a defined subset of these measurements.Trend analyses aggregate monthly and exclude months with fewer than 20 conversations; classifier validation against human labels is reported in Appendix B.Source: §4 · p. 4
  • The 24-category platform taxonomy supports comparison with published aggregates, while its coarse categories can obscure distinctions among local uses.The authors pair it with a taxonomy-independent unsupervised pipeline and validate classifiers against human labels, while noting that platform-defined categories and models shape measurement.Source: §4 · p. 4
  • Country-specific topic discovery embeds the first ten user–assistant turns, forms 500 micro-clusters, and merges them into roughly 50 topics per country.The resulting topics cover 91–95% of conversations per country and are grouped into ten cross-country themes; cluster-level and theme-level shares have different granularities.Source: §4 · p. 5
  • Intent classification labels user messages as Asking, Doing, or Expressing, defining Expressing as emotional sharing, statements, or reflections and adding Other for residual cases.Messages use the previous five turns as context; the majority label determines a conversation’s label, with ties retaining multiple labels.Source: §4 · p. 5
  • Task-purpose labels classify each conversation as Work, Coursework, or Personal.Work concerns tasks related to the user’s profession; Coursework covers coursework or academic help, and Personal covers neither category.Source: §4 · p. 5
  • An author read 100 randomly sampled Expressing conversations per country in full and inductively grouped recurring patterns into themes.This thematic reading uses one reader and no codebook or inter-coder agreement; it characterizes forms of expression rather than estimating prevalence.Source: §4 · p. 5
  • Within-user trend analyses use 530 users with at least 30 conversations spanning at least 12 months, representing 78% of classified conversations.The analyses compare intent by calendar quarter across first-use cohorts and each user’s first versus last three active months, with user-clustered bootstrap intervals.Source: §4 · p. 5
  • Demographic comparisons use user-clustered bootstrap intervals and Benjamini–Hochberg correction, with surviving patterns interpreted descriptively rather than causally.Language detection is reduced to English versus not for India, Nigeria, and Pakistan, and Portuguese versus not for Brazil, because romanized and code-mixed text can be misassigned.Source: §4 · p. 5

5 Results

Across four countries, personal use dominates, while unsupervised topic discovery surfaces locally specific uses that broad categories obscure. Asking remains the most common interaction mode, but Expressing has grown steadily, including within individual users.

5.1 RQ1: Work, Coursework, and Personal Use

  • Personal conversations make up 61.5% of India’s, 63.7% of Brazil’s, 55.0% of Nigeria’s, and 55.4% of Pakistan’s conversations; work and coursework are similarly prevalent overall.The work and coursework shares are 20.8% and 19.6%, respectively, and personal use has increased in every country.Source: §5.1 · p. 5, Figure 1 · p. 6

5.2 RQ2: Topics

  • The OpenAI taxonomy shows country differences in topic prevalence, while work conversations commonly involve writing and technical help and coursework is dominated by STEM subjects.Writing and technical help together account for roughly 45–50% of work conversations in every country; writing alone is the largest work category at 28–42%.Source: OpenAI taxonomy · p. 6, Topics within work, coursework, and… · p. 6
  • Gender and age patterns are interpreted descriptively: most reported topic differences do not survive correction, and the sample’s uneven group sizes constrain broader generalization.After FDR correction, the clearest age differences are higher Education/Academic prevalence among younger users in Brazil and higher Programming/Tech prevalence among younger users in Pakistan.Source: OpenAI taxonomy · pp. 6–7, Topic Discovery through Unsupervised… · p. 7
  • Unsupervised clusters reveal country-specific uses, including health and wellness in India and Brazil, Urdu–English translation in Pakistan, current affairs in Nigeria, and self-reflection in Brazil.Health and wellness accounts for 8.0% of India’s and 9.2% of Brazil’s conversations; Urdu–English translation is Pakistan’s largest cluster at 7.2%.Source: Figure 2 · p. 6, Topic Discovery through Unsupervised… · p. 7, Figure 3 · p. 7

5.3 RQ3: Asking, Doing, Expressing

  • Expressing grew in every country to roughly a fifth of conversations by the observation window’s end, while Asking declined gradually and Doing stayed roughly stable except for a rise in Nigeria.Within 449 users, Expressing rose 13.7 percentage points between their first and last three active months, while Asking fell 10.6 points and Doing fell 3.0 points overall.Source: §5.3 · p. 7, Figure 5 · p. 8
  • Topic and interaction mode align: programming, content writing, resume drafting, and translation skew toward Doing, while self-reflection, emotional venting, religion, and personal messages concentrate in Expressing.Work-related conversations are disproportionately Doing in every country; demographic differences in intent are described as tendencies within the sample.Source: §5.3 · pp. 7–8, Figure 4 · p. 8
  • Doing conversations use the dominant language more often than Expressing conversations in every country, but most of this aggregate gap reflects the topics each mode covers.Across the 15 largest Expressing clusters per country, the median within-topic language gap is +0.8 percentage points; some clusters have larger gaps, while others show none.Source: §5.3 · pp. 8–9
  • Expressive conversations rarely contain only emotional expression; they usually combine disclosure with a request for information, advice, or a task tied to the disclosed context.Recurring subjects include health concerns, relationships, academic pressure, religious questions, and money.Source: Figure 6 · p. 8, §5.3 · p. 9

6 Discussion

Across these four countries, ChatGPT use extends well beyond workplace productivity, with locally distinct topics and a growing expressive mode. The discussion argues that global categories need country-sensitive discovery, while noting important sample, measurement, and outcome limits.

RQ1: Beyond the workplace productivity frame

  • Personal conversations account for 55–64% of use in every country, while coursework is roughly as prevalent as work.The work observed is shaped by the participant pool, and personal use remains the majority in every age band.Source: §6 · p. 9

RQ2: Local context shapes what the same tool becomes

  • Health and wellness is prominent in India and Brazil and top-five in Pakistan; Urdu–English translation leads in Pakistan, while religious queries, online earning, and self-reflection vary by country.Religious queries are prominent in Nigeria and Pakistan; online earning clusters are prominent in India and Pakistan, and self-reflection is especially visible in Brazil.Source: §6 · p. 9
  • Fixed global categories aid comparability but can flatten locally meaningful uses, so the discussion recommends pairing them with country-sensitive discovery.The authors also note that convenience samples cannot settle whether country differences reflect national contexts or sample composition.Source: §6 · p. 10

RQ3: Disclosure plus request

  • 13.7 pp: Expressing share rose from users’ first three active months to their last three, and 81% of users individually expressed more over time.Doing stayed roughly flat and Asking declined slowly; the data cannot distinguish user learning from a more conversational model eliciting disclosure.Source: §6 · p. 10
  • Expressive conversations often pair personal disclosure with a request, and their lower dominant-language share is mostly explained by topic composition across major clusters.Across the 15 largest Expressing clusters per country, the median language-share gap from same-cluster Asking and Doing conversations was +0.8 pp, with larger gaps in some topics.Source: §6 · p. 10

What centering these users changes

  • Because expressive use can combine intimate disclosure with task completion, the discussion argues that safety evaluation should address that overlap and multilingual use.The authors also flag commercial privacy concerns around intimate disclosures and leave open whether disclosure improves the help users receive.Source: §6 · pp. 10–11

Limitations

  • The study’s prevalence estimates are conditional on a convenience sample, its classifiers rely on platform models, and observed behavior does not establish whether answers are accurate or helpful.The observation window also coincides with model and feature changes, so user learning cannot be fully separated from changes in what the system elicits.Source: §6 · p. 11

7 Conclusion

The study analyzes donated ChatGPT conversations from four countries to compare purpose, topics, and interaction modes. It finds that personal and locally specific uses are common, while expressive conversations grew and workplace productivity captures only part of use.

7 Conclusion

  • Personal use is the majority in every country, coursework is about as common as work, and expressive conversations grew to a fifth or more of use.Information seeking declined only modestly, while task delegation did not grow.Source: §7 · p. 11
  • The analysis covers 202,590 conversations donated by 1,252 users in India, Nigeria, Brazil, and Pakistan, paired with self-reported age and gender.The authors describe it as the first conversation-level, demographically grounded comparison across several Global South countries.Source: §7 · p. 11
  • Unsupervised topic discovery surfaces country-specific uses, including health, religion, translation, online earning, and self-reflection, that fixed categories fold into generic labels.The authors conclude that understanding adoption requires conversation-level, country-sensitive measurement alongside global aggregates.Source: §7 · p. 11

Generative AI use

The appendix describes the per-country topic-modeling pipeline and its cross-country aggregation, and discloses the paper’s use of generative AI tools.

Generative AI use

  • The authors used Gemini and Claude for ideation, framing, and writing assistance, and Claude Code for data analysis; they say all outputs were checked.The disclosed assistance applied to parts of the paper.Source: Generative AI use · p. 11

A BERTopic Pipeline Details

  • Conversation embeddings use the first ten user–assistant turns, capped at 5,000 characters per message, then are L2-normalized.The embeddings are 3,072-dimensional; the authors say opening exchanges typically establish the task and limit cost by embedding only the first ten turns.Source: A BERTopic Pipeline Details · p. 13
  • For each country, MiniBatch K-means creates 500 narrow micro-clusters before cosine-distance hierarchical clustering merges them toward approximately 50 interpretable topics.The cutoff is selected by visual inspection of each country’s dendrogram, balancing comparability and granularity.Source: A BERTopic Pipeline Details · p. 13
  • GPT-4o-mini labels clusters using representative documents and keywords; ambiguous labels are refined, and the final country models cover 91–95% of conversations.The pipeline yields 50, 45, 36, and 53 topics for India, Nigeria, Brazil, and Pakistan; unassigned conversations are excluded from BERTopic analyses but retained elsewhere.Source: A BERTopic Pipeline Details · p. 13
  • The ten-theme cross-country view groups cluster labels by keyword matching, a lossy aggregation that retains country-level cluster resolution for cluster figures.Some cluster labels match multiple themes.Source: A BERTopic Pipeline Details · p. 13

B Classifier validation

The two conversation-level classifiers were checked against human labels, with stronger reported reliability for task purpose than for the Expressing category.

Classifier validation

  • One author, blind to model predictions, labeled a stratified sample of about 50 conversations per country; 194 valid items were pooled after ambiguous cases were dropped.The pooled labels were compared with model labels using accuracy, macro-F1, and Cohen’s κ.Source: B Classifier validation · p. 13
  • Work/Coursework/Personal classification reached pooled accuracy 0.88 and κ = 0.82; Expressing had recall 0.78 but precision 0.35.The validation used human gold labels on a stratified sample; the table caption describes pooled metrics and identifies Expressing over-prediction as the residual error.Source: B Classifier validation · p. 13, Table 3 · p. 13
  • Asking and Doing reached F1 scores of 0.84 and 0.77, while Expressing was often assigned to conversations humans labeled Asking or Doing.The authors attribute the mismatch to self-disclosure interleaved with concrete requests, and treat Expressing prevalence as a noisy upper bound.Source: B Classifier validation · p. 13

D Ethics and Data Protection

The study collected consented conversation exports, removed direct identifiers before upload, and restricts access and release because personal histories may remain identifying.

Consent and recruitment

  • Participants consented before uploading, confirmed the export was their own, and the protocol received institutional review board approval.Participants were told names and contact details would be removed on their device and that participation could be withdrawn.Source: D Ethics and Data Protection · p. 14

Compensation

De-identification

  • The browser-side script removes names, email addresses, and other direct identifiers, but cannot reliably remove life events, workplaces, or references to other people.The authors therefore treat the scrubbed corpus as sensitive.Source: D Ethics and Data Protection · p. 14

Storage and access

  • Uploads were stored on access-controlled institutional storage; vendor API calls for classification and embedding used terms stating submitted data were not used to train models.The data were not shared with the crowd-work platform.Source: D Ethics and Data Protection · p. 14

Release policy

  • Conversation text is not publicly released; academic researchers may apply for a defined subset under an agreement barring re-identification, redistribution, and commercial use.Per-conversation labels, prompts, analysis code, and aggregate tables are released to support reanalysis of quantitative results.Source: D Ethics and Data Protection · p. 14

Examples in the paper

  • Quoted conversation examples were paraphrased and stripped of exact ages, amounts, drug names, place names, and rare combinations of attributes.The examples appear in Sections 5.3 and 6.Source: D Ethics and Data Protection · p. 14

E Research Questions and Methods

Table 9 links each research question to its measurements, validation, and results location.

Research Questions and Methods

  • Table 9 maps each research question to the measurements answering it, how each measurement is validated, and where results appear.The table is reproduced in the appendix for space.Source: E Research Questions and Methods · p. 14

F Additional Figures

The appendix supplements the main results with per-country breakdowns of purpose, topics, and Asking/Doing/Expressing intent, including demographic and temporal views.

F Additional Figures

  • The appendix organizes supplementary figures around purpose, topics, and intent, with Table 9 mapping research questions to measurements, validation, and results locations.The figures provide per-country, per-demographic, and per-topic breakdowns.Source: F Additional Figures · pp. 14–15, Table 9 · p. 16

Purpose: per-task topics and trends

Topics: full distributions and demographic conditioning

  • Topic breakdowns show gender and age patterns in unsupervised themes, while supervised gender differences are selected by magnitude and none listed survives correction.The figures also track coarse-topic shares over time for non-work conversations, from early dominance of writing and technical help to a later rise in practical guidance and information seeking.Source: Topics: full distributions and… · p. 21, Figure 12 · p. 22, Figure 13 · p. 23

Intent: demographic and topical conditioning

  • Expressing rises within users from their first three active months to their last three in every country, while early- and later-starting users’ quarterly trends move together.The increases are 15.7 pp in India, 16.2 pp in Brazil, 10.5 pp in Nigeria, and 9.6 pp in Pakistan; the two cohort curves correlate at r = 0.98. Figure 16 shows monthly within-user shares for more active users.Source: Intent: demographic and topical… · p. 25, Figure 16 · p. 26
  • Asking, Doing, and Expressing vary with coarse topic: programming and translation skew Doing, health and how-to Asking, and self-expression and personal communication Expressing.The appendix also breaks down intent by work versus non-work, country, age, and gender.Source: Intent: demographic and topical… · p. 25, Figure 17 · p. 26, Figure 18 · p. 27
  • Expressing has the lowest dominant-language share in every country, and Table 8 reports a median difference of +0.8 pp across 60 leading clusters.The appendix also presents Table 7’s demographic-topic differences whose user-clustered bootstrap intervals exclude zero; asterisks mark differences surviving Benjamini–Hochberg correction.Source: Table 8 · p. 15, Table 7 · p. 15, Figure 21 · p. 28
Loading 2609.38279v1…