Australia’s ChatGPT habits: action-first, multilingual, local

Introduction

If you’ve ever wondered whether people in Australia use ChatGPT the same way as everyone else—or whether it shows up differently depending on local life—new research has a surprisingly concrete answer.

A team analysed real-world ChatGPT interaction logs from Australia using WildChat, a public dataset of anonymised interaction traces. Their findings (from the paper “How People Use ChatGPT in Australia: A WildChat Analysis” at https://arxiv.org/abs/2609.28990) are based on 37,845 Australian conversations collected from April 2023 to July 2025, across 65 detected languages. And the headline isn’t “Australians use ChatGPT a lot.” It’s how they use it: more action-oriented, more work-inclined, and increasingly more self-expression, with plenty of Australia-specific references to institutions, services, laws, and education pathways.

Why This Matters

This matters right now because ChatGPT is no longer just a “nice-to-have assistant.” It’s quietly becoming part of how people get things done—from drafting messages to agencies, to preparing job applications, to figuring out consumer rights, to negotiating study and immigration steps. When an AI system helps with that kind of real-world navigation, small differences in local context—language expectations, jurisdiction, institutional processes—can change whether the help is useful or risky.

Here’s a scenario that’s common in Australia and could be supported (or harmed) by today’s chatbots:
A non-native English speaker renting in NSW wants to complain about a refund/warranty issue. They ask ChatGPT to “help me write a complaint letter” and maybe “tell me what the rules are.” The research shows that Australian users frequently turn ChatGPT toward exactly this kind of task—especially consumer services and legal/government topics. That means evaluation can’t just ask “does the model write well?” It has to ask: does it understand Australian routes, terms, and uncertainty? And does it push users toward verification when stakes are high?

This research builds on earlier large-scale studies that classify ChatGPT use globally by work relevance, intent, and topic. But it pushes the field to stop treating AI usage as one-size-fits-all. The paper’s Australia-focused view suggests local patterns aren’t subtle: compared with a global benchmark (from OpenAI Signals analysed in prior work), the Australian subset is far more task-executing and much more work-oriented. That’s exactly the kind of “context gap” that can explain why some AI tools feel great in one setting and frustrating (or misleading) in another.

What the Researchers Actually Measured: 37,845 Australian WildChat Conversations

This study used WildChat-4.8M, a public dataset containing real-world ChatGPT interaction logs (with opt-in consent and anonymised transcripts). The researchers focused on conversations tagged as Australia via dataset metadata—resulting in 37,845 conversations from April 9, 2023 to July 31, 2025.

How they classified “what users were doing”

Instead of treating ChatGPT use as one vague behaviour, the researchers layered multiple labels on each conversation:

1. Work relevance
   Each request was classified as either work-related or non-work-related, with work including professional, academic, organisational, technical, or productivity-oriented activities.

2. Interaction intent
   Each request was then categorised as one of:

   • Expressing (opinions, feelings, reflection)
   • Asking (seeking explanations, advice, evaluation)
   • Doing (producing/translating/writing/coding/summarising—basically task execution)

3. Topic capability groups (7 high-level groups)
   The model output categories were grouped into:

   • Writing
   • Practical Guidance
   • Technical Help
   • Multimedia
   • Seeking Information
   • Self-Expression
   • Other/Unknown

4. Work activities (for work-related conversations)
   For work requests, they mapped to O*NET Work Activities, which provides a structured taxonomy of jobs and work tasks.

5. Australia-related domain tags
   A conservative classifier flagged conversations that explicitly referenced Australia in ways like institutions, laws, regulators, education systems, companies, cultural references, and public services. A manual check of 100 examples helped develop domain labels, and automated labelling on the full subset had agreement of Cohen’s κ = 0.715 (good agreement for this kind of annotation task).

This multi-layer approach matters because it avoids a common trap: assuming “writing” or “advice” means the same thing everywhere.

Australia’s usage profile: more doing, more work, less “just asking”

The most striking differences show up in work relevance and interaction intent.

Language diversity: English dominates, but Australia is clearly multilingual

In the Australian subset:
- English: 60.7% of conversations
- Other prominent languages included:
- Persian: 11.3%
- Chinese: 5.7%
- Russian: 4.8%
- French: 4.3%
- Yoruba: 3.1%
- Plus a long tail of smaller language proportions.

So even if Australia’s institutions often run in English, the dataset shows people use ChatGPT across many languages—likely for writing support, translation, and access to information.

Work relevance: Australia is comparatively work-oriented

The Australian subset is strongly work-oriented:
- Work-related: 61.1%
- Non-work: 38.9%

Compared with the global benchmark (OpenAI Signals):
- Work-related: 37.3%
- Non-work: 62.7%

Here’s the difference framed clearly:

They also sampled other countries (random 8,000 conversations from each of Canada, US, UK, China) using the same classifier. Australia landed higher than several English-speaking countries:
- Canada: 53.0%
- US: 51.0%
- UK: 40.9%
- But lower than China: 83.8%

So: Australia isn’t “unique” in being work-heavy, but it’s clearly more work-inclined than the global mix.

Intent: Australians ask ChatGPT to do things, not just talk or explain

Intent is where the gap really jumps out. In Australia:
- Doing: 74.8%
- Asking: 13.9%
- Expressing: 11.3%

Global benchmark (OpenAI Signals):
- Doing: 39.8%
- Asking: 42.1%
- Expressing: 18.1%

If you want a plain-English interpretation: in this dataset, Australians often treat ChatGPT as a task engine—someone (or something) that should produce, translate, edit, or generate the next usable step.

That has real design implications: an interface optimised for “open exploration” may feel mismatched when people mostly want execution.

Topic distribution and turn-taking: writing leads, translation spikes, self-expression climbs

When the researchers grouped conversations into the 7 high-level topic categories, the Australian mix looked like this:

• Writing: 33.17%
• Seeking Information: 20.17%
• Self-Expression: 16.73%
• Practical Guidance: 14.73%
• Technical Help: 8.97%
• Multimedia: (lower share overall)
• Other/Unknown: (also lower)

Compared with the global benchmark:
- Australia is less dominated by Practical Guidance (14.7% vs 28.8%).
- Australia is more concentrated in Writing (33.2% vs 28.1%).
- Australia has a notably higher share of Self-Expression (16.7% vs 7.5%).

Translation is the standout fine-grained label inside writing

At a lower level, the most frequent labels included:
- Translation: 9.06%
- Write fiction: 8.48%
- Edit or critique provided text: 8.07%
- Computer programming: 6.07%
- Tutoring or teaching: 5.85%
- How to advice: 5.68%

This combination is important: “writing” in Australia is not only corporate email drafting. It’s also creative production and language work—especially translation.

Not all topics behave the same way at work vs non-work

The paper also measured how topic categories split by work relevance:

• Technical Help: 96.64% work-related
• Multimedia: 80.81% work-related
• Practical Guidance: 79.56% work-related
• Writing: 70.55% work-related
• Seeking Information: 64.13% work-related
• Self-Expression: 94.70% non-work

In other words, self-expression isn’t just “a little personal.” It’s overwhelmingly non-work in this dataset.

Conversation length tracks with “reflective” tasks

They also looked at mean turns (average number of chat turns) for the 20 most frequent low-level topics with at least 100 conversations. Key pattern: longer, iterative tasks produce longer dialogues.

Higher mean turn counts included:
- Tutoring or Teaching: M = 3.13
- Health, Fitness, Beauty, or Self-care: M = 3.13
- Relationships and Personal Reflection: M = 2.96
- Creative Ideation: M = 2.91
- Personal Writing or Communication: M = 2.57

Lower mean turn counts included transactional prompts like:
- Greetings and Chitchat: M = 1.05
- Create an Image: M = 1.08
- Translation: M = 1.35
- Data Analysis: M = 1.39

That suggests people don’t “just ask once” when they need help thinking, learning, or working through judgement. They keep returning for refinement.

The timeline trend: self-expression grows a lot in 2025

Over time, writing stays prominent for most of 2023 and early 2024, often around one-third to one-half of monthly conversations. Self-expression is low for most of 2023–2024, generally below 10%, then rises:

• December 2024: 13.72%
• January 2025: 34.03%
• April 2025: 45.85%
• June 2025: 60.07%

Crucially, this doesn’t mean the dataset becomes “non-task.” Writing, seeking information, practical guidance, and technical help still remain visible—but the shape of usage broadens. It looks like ChatGPT becomes more conversational and reflective as people gain familiarity.

Australia-specific conversations: institutions, laws, education pathways, and consumer disputes

One of the most practically useful parts of this study is the Australia-related domain analysis.

Out of the full Australian subset, the researchers found:
- 1,288 Australia-related conversations
- Split into domains, totalling 9 categories

Top domains included:
- Culture/media: 25.54% (329 conversations)
- Work/employment: 19.10% (246)
- Consumer services: 16.15% (208)
- Education: 14.05% (181)
- Legal/government: 11.18% (144)
- Smaller: travel/mobility (6.75%), housing (3.34%), healthcare (2.72%), immigration (0.93%), other (0.23%)

Culture/media: “Australia” often shows up as symbolism, not geography

Examples included conversations about:
- AFL and the CommBank Matildas
- ANZAC Day, Gallipoli, Kokoda
- bushfires
- Aboriginal and Torres Strait Islander culture
- local festivals like the Melbourne International Comedy Festival

So it’s not just “where am I?” It’s often “how do I talk about Australia in an Australian cultural frame?”

Legal and government: users want jurisdiction-aware explanations

Examples included:
- Privacy Act 1988
- Australian Privacy Principles
- Australian Consumer Law
- Corporations Act 2001
- agencies/regulators like:
- ASIC
- APRA
- AUSTRAC
- ACCC
- ATO
- Fair Trading NSW
- ReportCyber
- and local councils (e.g., Brisbane City Council)

The key implication: people are using ChatGPT like an information interface to Australian rules and pathways.

Consumer services: everyday problems with real consequences

This domain included disputes and guidance around:
- warranties, refunds, complaints
- companies and platforms like Costco Australia and consumer issues involving “.com.au” sites
- examples involving LDV Australia vehicle warranty disputes

This is exactly where misinformation (even “confidently worded” misinformation) can cause harm—financially and procedurally.

Education: AU-specific pathways like ATAR and TAFE

They saw Australia-linked education references such as:
- ATAR
- Selective high school entrance exams
- term calendars
- institutions like Swinburne, UQ, Monash, Charles Darwin University, TAFE NSW
- registered training organisations

This suggests AI use isn’t only for essays; it’s also for navigation of the education system itself.

What these results mean for designers and researchers building for Australia

The paper closes with implications, and they line up with what this data suggests: Australians use ChatGPT as an action system embedded in real life. That changes what “good performance” should mean.

1) Local context awareness isn’t optional anymore

When users ask about laws, consumer rights, education pathways, workplace rules, tenancy issues, or government services, the response depends on jurisdiction and institution type. That means evaluation should be task- and context-specific, not just “generates plausible text.”

In practice, a chatbot should:
- distinguish federal vs state/territory issues
- ask clarifying questions (e.g., location, dates, eligibility)
- encourage checking official Australian sources before acting

2) Translation should be treated as “institution-ready communication,” not just word swapping

Translation is the most frequent fine-grained label. But translation in this context often means:
- writing emails
- drafting complaints
- preparing applications
- communicating with employers, landlords, universities, and regulators

So interfaces should help users verify:
- whether the wording matches institutional expectations
- what assumptions the translation makes
- what details to confirm before submission

3) Safeguards for everyday high-stakes tasks should be built in

Even if a chatbot interface doesn’t look like “medical” or “legal software,” these conversations involve:
- consumer disputes
- privacy issues
- workplace rights
- immigration steps
- healthcare access
- school pathways

The paper’s implication is clear: the “high stakes” label shouldn’t be tied to the user’s intent; it should be tied to the domain of potential impact.

That might mean adding friction like:
- requiring users to confirm dates/locations/eligibility
- clearly marking uncertainty
- escalating to official resources or human help where appropriate

4) Design needs to support transitions from doing to self-reflection

A big trend in the Australian dataset is the growth of self-expression over time. And the turn-count analysis suggests reflective tasks involve more back-and-forth. That’s a design cue: users aren’t always just producing outputs—they sometimes need space to think.

But self-reflection and advice can also create risk. If the system switches modes seamlessly, it should also switch responsibility:
- avoid overclaiming
- adjust response boundaries
- be explicit when topics become emotionally sensitive or personally consequential

5) If you care about Australia, evaluate Australia—not just global averages

The paper explicitly argues for locally grounded evaluation benchmarks. For research teams, that means building scenarios like:
- drafting an Australian consumer complaint
- identifying the correct regulator pathway
- preparing an appeal tied to specific education institutions
- translating institution-facing communication with uncertainty handling

This is where the study linked back to earlier work: it builds on the global WildChat-style analyses (including the framework used by Chatterji et al., 2025) but shows why local adaptation changes the picture.

Key Takeaways

• Australia’s ChatGPT use looks more “action-first” than the global mix: 74.8% of interactions are classified as Doing (vs 39.8% globally).
• Work-related conversations dominate in Australia: 61.1% work-related vs 37.3% in the global benchmark.
• Writing is the biggest topic group (33.17%) and translation is the top fine-grained label within writing (9.06%), reinforcing that multilingual support is a core real-world need.
• Self-expression is growing fast over time: from generally under 10% for much of 2023–2024 to 60.07% in June 2025.
• Conversation length is task-dependent: reflective/teaching/self-care and relationship topics average more turns (e.g., M = 3.13 for tutoring and self-care), while transactional tasks like translation are shorter (M = 1.35).
• Australia-specific domains matter: in 1,288 tagged Australia-related conversations, the top areas were culture/media (25.54%), work/employment (19.10%), consumer services (16.15%), education (14.05%), and legal/government (11.18%).
• Practical implication for builders: don’t evaluate only “general chatbot quality.” For Australia, you need jurisdiction-aware responses, institution-ready multilingual communication, and safeguards for everyday high-stakes tasks.
• Practical implication for users: when asking ChatGPT about Australian rules (consumer law, privacy, immigration, education), treat the output as a first draft—verify with official sources before acting.

If you want, I can also turn these findings into a “design checklist” for Australia-specific AI features (for consumer complaints, rental disputes, education planning, translation for forms, etc.).