Researchers sampled the daily lives of 2,197 people and found they spoke about 338 fewer words for every passing year from 2005 to 2019 — with the decline steepest among under-25s, who lost an estimated 451 spoken words a day per year

There is a large difference between finding that people sampled in later years spoke less and watching the same people become quieter over time. A newly published analysis found the first pattern. It did not follow 2,197 individuals from 2005 to 2019.

Valeria Pfeifer and Matthias Mehl drew on passively sampled audio from 22 studies and reported that each later study year was associated with 338 fewer estimated spoken words per day. In a separate age split, the slope was steeper among people under 25, at 451 words per day for each later year. Their short Perspectives on Psychological Science paper describes this as a loss of conversation, while also acknowledging that the data cannot say which conversations disappeared or why.

This is one exploratory analysis, not settled consensus. The signal deserves a closer look, but “lost” should be read as a modelled difference between samples collected in different years, not a word counter falling inside every participant’s life.

The recorder heard fragments, not whole days

The underlying data came from an instrument called the Electronically Activated Recorder, or EAR. Participants wore a small recorder during ordinary life without knowing exactly when it would switch on. Depending on the study, it captured snippets such as 30 seconds every seven or 12 minutes. Across the archive, 2,197 participants generated 631,030 ambient-audio recordings.

Human transcribers isolated the participant’s own speech in valid waking recordings. Software counted every spoken word token, including repeated words. The team calculated an average number of words per snippet and extrapolated that rate across a waking day. Adults were assigned 17 waking hours; participants aged 10 to 17 were assigned 16.

On average, each participant contributed about 164 minutes of actual audio spread across roughly 46 hours of monitoring. That is far richer than asking someone how talkative they felt, and intermittent recording reduces the burden of continuous surveillance. It remains an estimate. A reported 12,000 words per day does not mean the device captured all 12,000.

The pooled mean was 12,792 estimated words per day. Individual values ranged from fewer than 100 to more than 120,000, an enormous spread that the original registered report in the Journal of Personality and Social Psychology discusses openly.

The 338-word slope was added after preregistration

The archive was assembled to revisit a different question: whether men and women speak similar numbers of words in daily life. Its main analysis was preregistered. During descriptive work, the researchers noticed that the new pooled mean was about 3,000 words below the 15,959-word estimate in a widely reported 2007 study using the same method.

They then ran an additional, explicitly exploratory analysis. A Bayesian multilevel model related estimated words per day to the year in which each sample began data collection, coded from 2005. The model estimated 16,632 words per day in 2005 and a yearly slope of minus 338. Its 95 per cent credible interval stretched from minus 652 to minus 25 words, while the annual standardised effect was very small at d = -.04.

The collection as a whole spans 2005 through 2019. There is a small technical wrinkle in how that range is described: the original report says its study-start-year predictor ran from 2005 through 2018, although some of those studies continued gathering data into 2019. The later letter describes the period as 2005 to 2019. The headline preserves that published framing, but the predictor was the sample’s start year.

Most importantly, no person supplied a 14-year speech trajectory. The model compared different people recruited into different projects at different times. It can identify a downward calendar-year association in this archive. It cannot show that an individual who spoke 16,000 words in 2005 spoke roughly 11,000 by 2019.

The under-25 result was the steeper of two estimates

Pfeifer and Mehl divided the sample at age 25, a threshold intended to separate adolescence and emerging adulthood from later adult life. Among 980 people below 25, they estimated 451 fewer daily words for each later year. Among 1,146 people aged 25 and above, the estimate was 314 fewer. The subgroup counts leave 71 of the 2,197 participants outside this particular comparison, apparently because not every record supplied usable age information.

The younger slope was 44 per cent larger, but uncertainty remained wide. The authors reported an 80 per cent credible interval of minus 825 to minus 89 for the younger group, and minus 579 to minus 41 for the older group. Those intervals overlap substantially and are less conservative than the 95 per cent interval reported for the overall trend.

The paper calls this age analysis descriptive. It does not report a repeated-measures decline within young people, and it does not establish that the two age slopes differ reliably from each other. “Steepest among under-25s” therefore means steeper in this two-group model, not that each person under 25 lost 451 words on every birthday.

Twenty-two studies cannot become one stable population

The source data are unusual and valuable precisely because naturalistic audio is hard to collect. They are not a representative annual survey. The 22 samples included schoolchildren, university students, older adults, couples, new fathers, people going through divorce, women after childbirth, cancer patients and partners, and people taking part in studies of illness or recovery. Most participants lived in the United States, with smaller studies elsewhere.

Protocols varied too. Monitoring lasted from two days to more than a week. Snippets differed in length and frequency, and studies sampled different mixtures of weekdays and weekends. The main report performed sensitivity analyses for several recording features, but the calendar-year result was a secondary model rather than a purpose-built test of social change.

A multilevel model accounts for the fact that participants are nested inside studies. It cannot make an older-adult community sample equivalent to a university cohort, or separate year perfectly from whatever populations and places researchers happened to study then. The 2005 end of the archive is particularly thin. Its oldest sample comprised 13 adults with rheumatoid arthritis, followed by other small early projects. Sample composition, location, life context and protocol can travel with calendar year.

The data and analysis materials are public on OSF, which makes the result inspectable. Openness does not remove that structural confounding.

Fewer spoken words are not automatically less connection

The recorders measured quantity. They did not tell the time-trend model whether speech involved a close friend, a stranger, a colleague or a family member. They could not distinguish affection from an argument, a long meeting from an intimate exchange, or necessary instructions from conversation someone genuinely wanted.

The studies also did not measure texting, email, social media or other digital communication consistently across the 14-year span. The authors raise digital substitution as a plausible explanation and note that younger people use more communication technology. They stop short of claiming the data demonstrate it. The older group also had a downward slope, so age as a rough proxy for technology did not explain the whole pattern.

Related experiments can fill in small pieces without solving this study’s causal problem. Silicon Canals previously covered a café experiment in which keeping phones available made meals slightly less enjoyable, even though diners used them for only about a tenth of the time. Another study of conversations between strangers found that people underestimated how connecting a deeper exchange would feel. Both concern the quality and opportunity of interaction, not a population-wide word budget.

A recent two-week study comparing a human texting partner with a supportive AI friend adds a further complication: typed conversation with another person reduced loneliness while the AI and journal conditions did not. Typed words are not a single substitute category. Who is on the other side still matters.

A stronger trend test would repeat the same design

What I find useful here is not the invitation to make everyone say 338 extra words. The paper’s own model does not test that intervention. Its value is the question it exposes: has live speech declined, and if so, which kinds of conversation moved elsewhere?

A cleaner answer would require comparable, broadly recruited samples measured at regular intervals with the same recorder, the same weekday and weekend balance, and the same waking-time assumptions. Logging voice calls and typed communication alongside in-person speech would test substitution directly. Following at least some of the same people would separate individual change from generational and sample change.

For now, the archive shows that people recorded in later studies produced fewer estimated spoken words per day. The cause, the conversations affected and the social cost remain open questions.