Understanding yourself / 10 min read / November 14, 2026

What people talk about in conversation, counted

Social and personal topics took about two thirds of the talking time in the 1997 study that counted them, sampled every thirty seconds by observers. A 2025 recount put it near 85 percent. What a chat export changes is the method.

A flat illustration of eight stacked message bars, six of them violet and two grey, beside a dotted vertical line with a dot level with each bar, for the article What people talk about in conversation, counted.

Social and personal topics took up about two thirds of the time people spent talking, in the one study that put a number on it. The number came from observers sitting in public places with a notepad, recording what one speaker was talking about once every thirty seconds.

The figure is quoted constantly and almost always without its method. It was not produced by reading anything. It was produced by listening to 45 conversations, most of them in two university cafeterias, and sampling the subject every half minute.

That matters now, because a chat export is the first kind of conversational record where the topics do not have to be sampled at all.

Two thirds of the talking time went on social and personal topics

The study is Robin Dunbar, Anna Marriott and Neil Duncan, “Human conversational behavior”, published in Human Nature in 1997. Its own summary says that conversations in relaxed social settings “consist predominantly of exchanges of social information (mostly concerning personal relationships and experiences)”, that most of those exchanges carry information about the speaker or about third parties, and that “very few involve critical comments or the soliciting or giving of advice”.

The two thirds number is not in that summary. It is reported by a later paper that went back over the same question, and which counts Dunbar among its authors. That paper, a 2025 revisit published in Language and Cognition, says the 1997 study “estimated that ‘gossip’, understood loosely as conversation about social and personal topics, accounted for about two-thirds of time spent on conversation”. So the figure used here is secondary: it is the later authors’ reading of the earlier result, not a line lifted from the 1997 paper itself.

“Gossip” is the researchers’ own word for the category, used loosely and without the everyday sting. It covers talk about people, including talk about yourself. The 1997 summary is explicit that criticism of other people was rare in what the observers heard.

Forty five conversations, sampled every thirty seconds by someone listening

The method is the part worth knowing. According to the 2025 revisit’s description of it, the 1997 researchers analysed 45 casual conversations held in public settings, recording the general topic of each speaker’s utterances by hand, speaker by speaker. They used scan sampling: they noted the topic at one moment, then repeated that every 30 seconds.

The revisit also lists what that design cost. The coders were, in its words, “surreptitious eavesdroppers”, so they might not have heard the full content of some conversations, and they had to identify and record topics in real time while still listening. In some cases the coders had to rely on their own reading of what a speaker intended in order to pick one category, and the classifications rested on a single coder’s intuition at one of three locations.

The sample was narrow as well. More than three quarters of the conversations were held by students in two university cafeterias, in Liverpool and in London.

The same research group had already reported a separate limit on conversation from field observation: Dunbar, Duncan and Nettle, writing in Human Nature in 1995, found an upper limit of about four on the number of people who can interact in one spontaneous conversation, and attributed it to the mechanics of producing and hearing speech. That was a claim about talking out loud, not about threads.

Fourteen subject areas, grouped afterwards into ten

The topics were not invented as the coders went along. The 2025 paper records that the original classification had fourteen subject areas, later reclassified into ten major categories: personal relationships, personal experiences, future social activity, future non-social activity, sport and leisure, culture, art and music, politics, religion, morals and ethics, work and academic, and technical or instructional.

The 1997 authors said they “delineated the topics a priori to reflect functionally relevant categories”. The social share was then a sum: the time spent on whichever of those lower level topics had been grouped together as social. The 2025 paper points out that the 1997 study never gave an explicit definition of social against non-social, because sorting conversation into those two bins was not its main purpose.

The 1997 summary does report a handful of differences between the sexes in how much time went on particular topics, and the authors interpreted those differences as reflecting, in their words, “females’ concerns with networking and males’ concerns with self-display”. That is an interpretation offered by the authors of a 1997 observational study, and it is worth holding at arm’s length rather than repeating as a fact about anyone.

A recount of a large spoken corpus landed near 85 percent

The 2025 revisit asked the 1997 question again with different tools. Instead of listening in public, it worked from a corpus of informal Polish conversations: 669 conversations and more than 2.6 million word tokens, built from live recordings made in private as well as public places, with speakers ranging across ages and levels of education. Two annotators coded the lines, and for the broad social against non-social split they agreed on 71 percent of the dataset, leaving 197,621 lines to analyse.

Its answer was higher than two thirds. The Oxford record for the paper states the result plainly: against the 1997 estimate that two thirds of conversation gravitates around social matters, “our findings indicate an even larger portion, approximately 85% being of a social nature”.

The revisit also cites an intermediate figure from the same line of work: Dahmardeh and Dunbar put the proportion of time devoted to social topics in Iranian conversations at 76 percent in 2017. Three numbers, three methods, one direction.

Change the definition of social and the share moves a long way

The 2025 authors make the caveat themselves, and it is the most useful sentence in the paper: the proportion of social to non-social content “depends mostly on how ‘social content’ is defined”. They chose a deliberately inclusive definition, in which talking about yourself counts as social, on the grounds that self-disclosure is how reputations get built.

The size of that choice is measurable. In their own earlier preliminary study, leaving out the single category of “personal experiences” left only 50.9 percent of conversation content classified as social, against the roughly 85 percent their full analysis reports. One definitional decision moves the headline by more than thirty points.

So two thirds and 85 percent are not two measurements of the same quantity that happen to disagree. They are two different questions, asked of two different samples, by two different methods, both answered honestly.

A chat export lets topics be counted instead of sampled

Every figure above is an estimate from a sample. In 1997 a person wrote down one topic every thirty seconds and may have missed the rest. In 2025 two coders read transcribed lines and discarded the 29 percent they disagreed on. Both are careful. Neither is a complete count.

A chat export is different in kind. It is the whole written record, in text, with every message present and in order. The topics in it can be counted from the text itself rather than estimated by somebody listening. That is a change in method, not a confirmation of the 1997 finding, and it is worth saying clearly that nothing in the 1997 work was about written messages. A typed thread is a different medium: people choose what to put in writing, replies arrive minutes or days apart, and a great deal of the same relationship happens out loud somewhere else. What a count of an export measures is the written part.

Subject matter is also the part of a conversation that survives being anonymised. Portraitor’s privacy policy says that before anything is transmitted your browser replaces names, email addresses, telephone numbers, postal addresses and credentials with anonymous placeholders on your own device, and then states the limit of that in the same breath: “This removes who is speaking. It does not remove what was discussed.” A topic, by definition, is what was discussed. If you want the mechanics of that replacement, what happens to names in a chat analysis covers it, and reading only your own side of a conversation covers the narrower case where the export is of your own writing.

What a count of topics does not settle

A proportion is not a meaning. Knowing what share of a thread is about people tells you nothing about why it is that share, and this page is not going to tell you what your own mix says about you, because none of the research above supports that step. The 1997 and 2025 results are population level estimates of what conversation is made of, not readings of individuals.

The limits stack up. The 1997 figure rests on 45 conversations from a narrow sample, coded in real time by someone who was still listening. The 2025 figure rests on one language and one corpus, and its authors say the answer depends mostly on a definition they chose. Neither looked at written messages at all. And a topic label is a judgement in both studies: the 2025 coders could not agree on 29 percent of lines even for the two way split.

An export has its own gaps. It is one participant’s copy of something several people made together, which is why what happens to everyone else in a group chat is a fair question to ask first. For the broader list of things a written record genuinely cannot show, the limits of chat analysis is the honest version. Topic is also only one axis of a conversation: the tendency of language to lean positive is about the valence of the words rather than the subjects, and the cooperative rules behind how people phrase things is about manner rather than matter.

Frequently asked questions

Frequently asked questions

What do people talk about most in conversation?

Social and personal topics: other people, relationships and personal experience. The 1997 study by Dunbar, Marriott and Duncan was read as putting that at about two thirds of talking time, and a 2025 recount of a Polish conversation corpus put it at roughly 85 percent. Both figures depend heavily on where the line between social and non-social is drawn.

Did the two thirds figure come from text messages?

No. It came from people listening to spoken conversations in public places in the mid 1990s and writing down the topic every thirty seconds. Nothing in that study involved written or typed messages.

Why did the later study get 85 percent instead of two thirds?

Mostly because it defined social content more inclusively and used a different method. Its authors say the proportion depends mostly on how social content is defined, and report that in their own earlier preliminary study, leaving out the single category of personal experiences left only 50.9 percent of content classified as social.

Can the topics in a chat export be counted exactly?

The text of an export can be counted directly rather than sampled, which is what makes it different from an observational study. What such a count means about the people in the conversation is a separate question, and the research described here does not answer it.

What did the researchers mean by gossip?

It was their technical shorthand for conversation about social and personal matters, not an accusation. The 1997 summary notes that very few of the exchanges the observers heard involved critical comments or the giving of advice.

The count is older than the medium, and the medium is the first one that can be counted without listening in. Portraitor works from the written record.

Where we come in

Your conversations, read the way you would read someone else’s.

Portraitor takes a chat export you already have, anonymises it in your browser, and returns a structured reading of how you come across. Identifying details are removed in your browser before anything is sent.

See how it works →
Sources
  1. Dunbar, R. I. M., Marriott, A., & Duncan, N. D. C. (1997). Human conversational behavior. Human Nature, 8(3), 231-246.
  2. Szala, A., Wacewicz, S., Placiński, M., Poniewierska, A., Schmeichel, A., Stefanczyk, M., Żywiczyński, P., & Dunbar, R. (2025). How much conversation content is actually social: human conversational behaviour revisited. Language and Cognition, 17.
  3. Oxford University Research Archive record for Szala et al. (2025), How much conversation content is actually social: human conversational behaviour revisited.
  4. Dunbar, R. I. M., Duncan, N. D. C., & Nettle, D. (1995). Size and structure of freely forming conversational groups. Human Nature, 6(1), 67-78.
  5. Portraitor, Privacy Policy, on masking and what it does not remove.