What Do People Ask AI? An Analysis of 10,000 Real Conversations

American studies from major AI companies have already shown what people use chatbots for in English. Europe, across 22 languages, is a different story — one we can tell from our own data. Alongside our AI chat, we run a free ChatGPT alternative in 22 languages — and it’s from this free version that we sorted 10,000 conversations into 15 topic categories to find out what Europeans actually want from artificial intelligence. Spoiler: far more work emails than philosophical debates.

A note on privacy up front: we worked exclusively with aggregated, anonymized data — no IP addresses, no publishing the content of a single conversation (every example in this article is a paraphrase). Paying customers’ data never enters our studies. Details are in the methodology at the end of the article.

Key findings at a glance

The big picture: emails, text fixes, and homework

The winner is clear-cut: writing a work or official text — an email to your boss, a complaint to a supplier, a request to an office, a résumé. A sixth of all conversations. Add in proofreading your own texts (11.6%), personal messages and greetings (3.0%), and translations (2.2%), and we arrive at the study’s central thesis:

A third of all AI conversations (33.1%) are about working with text. More than general knowledge, school, and how-to guides combined. In Europe, artificial intelligence isn’t an oracle or a friend — it’s an office assistant. This fits neatly with our time-of-day study: traffic peaks on Tuesday at 11 a.m., and drops by half on weekends.

The extreme is Portuguese and Dutch speakers: a third of their conversations are pure editing — “here’s my text, fix it” (32.7% and 28.6% for editing alone, respectively). Estonians and Greeks, by contrast, barely work with text at all — for them, AI is more of a search engine and advisor.

National profiles: what each country handles differently

The overall numbers hide the most interesting part — every market uses AI in its own way. Here are the topics where individual languages stand out the most; the number in parentheses shows how many times more often the topic appears than the European average:

A few profiles are worth calling out:

Germans use AI like an encyclopedia. General knowledge and trivia make up 25.5% of German conversations — 2.6 times the average and the highest of any language. Combined with the finding from our time-of-day study (79.6% of German conversations are a single long message), a consistent picture emerges: precise question, precise answer, done.

English is the language of homework. School and study make up 18.6% of English-language conversations — double the average. The English-language version draws a global audience of students: essays, test questions, explanations of course material.

Health: a Baltic and Balkan domain

Lithuanians ask AI about health more than anyone in Europe — every seventh conversation. And it’s not general tips: a typical query is interpreting one’s own lab results or the side effects of a specific medication. The whole Baltic and Balkan region discusses health with AI two to two-and-a-half times more often than Western Europe — a pattern that could reflect access to doctors as much as trust in technology. The French and the Poles barely discuss health with AI at all.

Bureaucracy: AI as a clerk for Central and Southern Europe

Notice periods, severance entitlements, benefits, inheritance. Slovenians and Greeks ask AI about law and bureaucracy in 15–17% of conversations — and Czech is also high (10.5%, 4th place), as is Slovak. At the other end are English and Dutch (under 2%). Where bureaucracy is complicated and legal help is expensive, people ask the machine instead.

The French phenomenon: AI as a stage

This is the study’s most striking outlier: creative writing and roleplay make up 21.8% of French conversations — 5.5 times the European average — more than work writing. French users (judging by content, often young anime fans) have the AI play favorite characters and hold long dialogues with them — for them, AI isn’t a tool but a stage. We independently spotted the same phenomenon in our politeness study, where French roleplay complicated measuring how people treat AI like a person. For other languages, creative writing stays around 1–5%.

Bonus: sexual content is rarer than its reputation suggests

AI chatbots have a reputation as tools for erotic roleplay. The data says otherwise: after a strict audit (see methodology), explicitly sexual content makes up 2.2% of all conversations. French leads again (11.0%), ahead of Polish (7.9%) — for French, this is an explicit branch of the same roleplay phenomenon. In six languages we didn’t find it at all, and Czech sits just above zero at 0.6%. Treat this figure as a lower-bound estimate of behavior on a regular, moderated chat — dedicated “AI companion” services would likely show different numbers.

And one more category worth noting: 5.5% of people are just testing the AI — a greeting with no follow-up, “what’s 2 + 2,” “do you speak Czech?” An independent dictionary-based measurement in our politeness study arrived at the same figure (5.4%), which we find methodologically reassuring: two different methods, the same result.

How we measured this

• Sample: 10,000 anonymized records from the free ChatGPT alternative by GuideGlare (no-registration version), June 23 – July 24, 2026, 22 language versions; after removing 6 technical duplicates, we worked with 9,994 conversations — the same sample as our etiquette and time-of-day studies. Classification covered all 9,994 conversations. Rankings include languages with at least 50 conversations (Danish and Finnish were excluded). Sample sizes range from 53 conversations (Slovenian) to 2,775 (English); full counts by language are in the downloadable data.
• Taxonomy: we built the 15 substantive categories plus “testing AI,” “other,” and “unclassifiable” empirically — by manually coding a stratified sample of 629 conversations, with definitions and edge-case rules for ambiguous situations (e.g. “translate this health-related text” is a translation, not health; a request for a finished official letter is writing, a question about a legal clause is law).
• Classification: a language model assigned each conversation to exactly one category based on its dominant intent (from the user’s messages). We checked quality with a blind verification: a stronger model independently reclassified 300 random conversations without knowing the original label — 81.7% agreement, with disagreements concentrated almost entirely between neighboring categories (typically “write a new text” vs. “fix my text”). The ranking of the major categories is unaffected; we don’t interpret differences of up to ~2 percentage points between neighbors in the ranking.
• Audit of the sensitive category: all 336 conversations labeled as sexual content underwent an additional strict review by a stronger model. 66% held up — the rest was mostly romantic roleplay without explicit content, which we moved to creative writing. The published NSFW figures are therefore post-correction; before it, the category would have come out at 3.4%.
• Statistics: percentages are shares within a given language; for all values we compute Wilson 95% confidence intervals (full figures in the downloadable data). The “times more often than average” figure is the ratio of a category’s share in a language to its share in the full sample; we only comment on deviations of ≥1.5× with sufficient sample sizes.
• Privacy: we worked exclusively with aggregates — no IP addresses, no user identification, no publication of conversation content; every example in the article and in the taxonomy definitions is a paraphrase.
• Limits: one category per conversation (multi-purpose conversations get their dominant topic); classification is based on a short excerpt of the conversation; 3.1% of conversations remained unclassifiable (fragments, incoherent text) and we leave them in the base. Language isn’t country — especially for Spanish, Portuguese and English, a significant share of users write from other continents. The sample consists of users of one free tool; the numbers describe behavior on our chat, not the population of an entire country.
• Data downloads and citation: download the complete aggregated data (18 categories × 22 languages × confidence intervals) as a CSV. You’re free to reuse the figures and charts in articles, blogs, and presentations — just credit “Source: GuideGlare” with a live link to this page; details in our citation guidelines. We’re happy to send print-quality charts — just email info@explicaire.com.

Try it yourself

The data for this study comes from our free ChatGPT alternative — it works in all 22 languages covered by this study, needs no registration, and the full version is GuideGlare’s AI chat. Whether you need to write an email to your boss like a sixth of Europe, or want to chat with your favorite character like the French. And if you’re curious when everyone finds the time for all this, read our companion study When Do We Talk to AI.