The study behind the texting-ruins-grammar headline explained under 5% of the variance; following children for a year found texting linked to better spelling.

If you search this question you will find, on the same results page, a university press release saying texting may lead to poor grammar and a research foundation saying the claim is a myth. Both are describing real studies. They are describing different ones.

Here is what each actually found.

The study behind the headline

Cingel and Sundar (2012), in New Media & Society, surveyed 228 students in grades six to eight in a mid-sized east-coast school district, asking them about their texting habits and giving them an offline, age-appropriate grammar assessment.

Their conclusion, verbatim from the abstract: “Results show broad support for a general negative relationship between the use of techspeak in text messages and scores on a grammar assessment.” They found word-level adaptations negatively related to grammar scores, while structural adaptations were not significant.

That sentence produced most of the coverage this topic has ever had. It is worth looking at what sits underneath it.

The linguist Mark Liberman, writing on Language Log in August 2012, worked through the paper’s own reported statistics. His findings, from the authors’ numbers:

• “Sent adaptation” accounted for about 4.7% of the variance in grammar scores - a figure he characterizes as “pitifully small”.

• On the 22-item assessment, that is somewhat less than one question’s worth of variation.

• The analysis tested at least twenty independent variables, of which one reached significance.

• The survey response rate was 42.1% of 542 distributed.

He raises further problems: unmeasured confounds including socioeconomic status and peer group; the likelihood that participants could infer the study’s hypothesis from a survey they took home; the omission of prior research finding positive associations between textism use and literacy; and his view that the assessment’s questions “generally have very little to do with ‘grammar’ in the traditional sense.”

Liberman’s analysis is an expert commentary on an academic blog, not a peer-reviewed reply, and we are labelling it as such. But the numbers he is working from are the paper’s own, and the study’s design is a cross-sectional survey. Whatever it shows, it cannot show that texting caused anything.

What happened when someone followed children over a year

The obvious way to settle a direction-of-causation question is to measure children twice.

Wood and colleagues at Coventry University and the University of Tasmania, funded by the Nuffield Foundation and published in the British Journal of Developmental Psychology, did exactly that: text messages and formal grammar and spelling tests, at baseline and at twelve-month follow-up, across primary school, secondary school and university samples.

The results run against the panic in a way that is difficult to explain away:

• No link, at either time point, between grammatical violations in children’s texts and reduced understanding of standard grammar.

• Primary school children’s use of ungrammatical word forms at baseline predicted better spelling ability twelve months later.

• Secondary students’ ungrammatical forms, missing punctuation and omitted capitalization showed positive associations with spelling growth.

• Young adults showed weak negative relationships between textism use and grammatical understanding - which were explained by general ability differences.

A companion concurrent study, Wood, Kemp, Waldron and Hart (2014), tested 243 people across three age bands - 89 primary pupils (mean 9.9 years), 84 secondary pupils (mean 12.8), 70 undergraduates (mean 20.8) - using standardized measures of spoken grammar comprehension (TROG-II), written grammatical understanding, orthographic processing, IQ and spelling, against coded text messages.

Their finding for children: “there does not appear to be any sign of a significant negative relationship between grammatical errors made when texting and primary school children’s understanding of standard grammar or knowledge of conventional spelling.”

And their overall statement: “Overall, there is little evidence that ungrammatical texting behavior is linked to grammatical understanding or knowledge of orthographic representations of language in children.”

There is one real negative finding in this work, and it is worth reporting because it is the exception. Among undergraduates, the tendency to make punctuation and capitalization errors in texts accounted for a further 8.0% of the variance in a written orthographic choice task, and the relationship held after controlling for IQ and spelling ability.

So the pattern is close to the opposite of the folk theory. In children, texting behavior is neutral or mildly positive for literacy. In adults, whose literacy is already consolidated, one narrow negative association shows up - in punctuation, not spelling.

The Dutch study that found the same reversal

van Dijk, van Witteloostuijn, Vasić, Avrutin and Blom (2016), in PLOS ONE, tested 55 Dutch primary school children aged 10 to 13 on receptive vocabulary, grammar via sentence repetition, and executive function tasks, alongside their elicited and spontaneous text messages.

Word omission in text messages significantly predicted better grammar performance. Textism use showed only marginal correlations with language measures, which lost significance in regression. Executive function showed no meaningful relationship with texting at all.

Fifty-five children is a small sample, and the study was funded by a university seed grant with no competing interests declared. But it is a third independent group finding a positive or null relationship where the popular claim predicts a negative one.

Why the reversal might be real

There is a mechanism proposed for this and it is not far-fetched. Most common text abbreviations are phonetically driven - they require the writer to break a word into sounds and represent them with letters, which is the same operation phonological awareness training asks of beginning readers. On that account, texting is not sloppy writing; it is unsupervised phonological practice.

We are flagging that as a proposed explanation rather than a demonstrated one. The longitudinal association is measured; the mechanism behind it is an interpretation.

What to make of it

The honest summary is short. There is no good evidence that text messaging damages children’s spelling or grammar. The best-designed study - a year-long longitudinal design with two measurement points across three age groups - found the association running the other way for the youngest children. The study that produced the alarming headline was a one-off survey in which the key variable explained under five per cent of the variance, in an analysis testing twenty variables.

The one qualification worth carrying: among adults, punctuation habits in texts do track written punctuation performance, by a small amount, independent of general ability.

This is a claim about spelling and grammar specifically. Whether reading on a screen changes comprehension is a different measurement with a different result, and whether hyperlinks impose a cost while reading is another one again.

Limitations and disputes

• The strongest critique of Cingel and Sundar is not peer-reviewed. It is a detailed analysis by a named linguist on an academic blog, working from figures reported in the paper. We have labelled it accordingly and have not treated it as equivalent to a published reply.

• All of this evidence is correlational or longitudinal-correlational. Nobody has randomly assigned children to text or not to text, and nobody plausibly could.

• The samples are small. 55 Dutch children; 243 across three UK/Australian age bands; 228 US students.

• The longitudinal study’s journal record was read via the funder’s account of it; we could not retrieve the paper itself, and no volume or page numbers are printed here for it.

• Wood, Kemp, Waldron and Hart (2014) was read from an institutional repository copy that did not carry the journal’s volume and page numbers; those are omitted rather than guessed.

• The phonological-practice explanation is an interpretation, not a finding of these studies.

• This says nothing about whether texting affects anything else - attention, sleep, social skill. Only spelling and grammar.

Toddlers learn less from video than from a person, but a meta-analysis found publication bias, and a responsive adult on a video call closed the gap.

Show a two-year-old an adult hiding a toy behind a chair and the child will go and find it. Show the same child the same adult doing the same thing on a screen, in the next room, live, and the child is markedly worse at it.

That gap has a name in developmental psychology - the video deficit - and it is the evidence base underneath every “no screens before two” guideline you have ever read. It is also more specific, and more qualified, than the guidance built on it.

The size of the gap

Strouse and Samson (2021), in Child Development, pooled 59 reports yielding 122 independent effect sizes on children aged nought to six learning the same material from video versus face-to-face.

The pooled result: g = −0.53, 95% CI [−0.66, −0.41]. Their plain statement of it:

“Across tasks and contexts, children ages 6 years and younger display approximately half of a standard deviation lower scores on learning outcomes after watching a video presentation than a face-to-face presentation of the same information.”

The moderators sharpen it considerably:

Two of those deserve attention. The deficit is not uniform across what is being learned - it is severe for finding a hidden object and modest for learning a word. And it did not matter whether the video was live or prerecorded, which is a problem for any account that leans on the video being “not really there”.

The honest part of the meta-analysis

Most coverage of the video deficit stops at “half a standard deviation”. The authors did not.

Egger’s regression test found significant funnel asymmetry: z = −7.84, p < .001. Trim-and-fill adjustment brought the pooled effect down to g = −0.47 - still clearly negative, but the adjustment was needed. And the median observed statistical power across the included studies was 35.19%, which the authors describe as “relatively low”.

Underpowered studies plus significant publication asymmetry is the signature of a literature that has probably published its positive findings more readily than its null ones. The effect survives correction, which is the important thing. But anyone quoting the half-a-standard-deviation figure without the correction is quoting the uncorrected number.

The authors’ own list of limitations is unusually frank: only seven effect sizes came from children under twelve months, only 33 of 122 from children aged 36 months and over, the videos were mostly researcher-made rather than commercial, and “between-subjects designs reported larger deficits than repeated-measures designs, which may indicate that random assignment did not entirely eliminate individual differences.”

What actually closes the gap

The most useful experiment in this literature is small, and it changes what the deficit is about.

Roseberry, Hirsh-Pasek and Golinkoff (2014), in Child Development, taught 36 toddlers (mean age 26.5 months, twelve per condition) novel verbs in one of three ways: a live adult in the room; an adult on a screen over a video call, responding to the child in real time; or a prerecorded video of that same adult interacting with a different child.

The results, on the test of whether children could extend the new verb to a matching action:

• Video chat: 66.9% looking to the matching action, t(11) = 7.06, p < .001
• Live interaction: 64.2%, t(11) = 5.87, p < .001
• Prerecorded video: 50.1% - chance, t(11) = .05, p = .96

The screen was identical in two of the three conditions. What differed was whether the person on it responded to the child. The authors conclude that “socially contingent interactions are a powerful catalyst for word learning” and define contingency precisely: responses that are “immediate, reliable, and accurate in content.”

This is why the video deficit is better described as a contingency deficit. It is not that pixels are inert. It is that a recording cannot answer a two-year-old.

Twelve children per condition is a small experiment, and the authors flag a design flaw of their own: the prerecorded condition used the wrong child’s name during warm-up, which “may have confused” participants.

The Baby Einstein record

The commercial half of this story is documented well enough to state without hedging, and it is a case study in what happens when a correlation meets a market.

Zimmerman, Christakis and Meltzoff (2007), in the Journal of Pediatrics, surveyed more than 1,000 families in Minnesota and Washington by telephone, in interviews of about 45 minutes, about their infants’ media exposure and vocabulary. For infants aged 8 to 16 months, each hour a day of baby DVD and video viewing was associated with understanding six to eight fewer words. For toddlers aged 17 to 24 months, no association was found.

It is essential to be clear about what that is. It is a cross-sectional association, with exposure and vocabulary both reported by parents, and the authors said so at the time: “We don’t know for sure that baby DVDs and videos are harmful, but the best policy is safety first.”

What happened next is the part usually left out. According to the University of Washington’s own account, Disney’s chief executive Robert Iger contacted the university to complain about the paper and its news release and to ask that the release be retracted. UW president Mark Emmert declined and stood by both.

Two years later, in October 2009, Disney offered refunds of $15.99 for up to four Baby Einstein DVDs per household, for purchases made between 5 June 2004 and 5 September 2009, after the Campaign for a Commercial-Free Childhood threatened a class action over unfair and deceptive practices.

The experimental test came the following year and was cleaner than the survey. DeLoache and colleagues (2010), in Psychological Science, had 12-to-18-month-olds watch a commercial baby DVD several times a week for a month, against a control group and against a group whose parents taught the same target words during ordinary daily activities. The result: children who watched the DVD “did not learn any more words from their monthlong exposure to it than did a control group.” The largest learning came from the parents doing the teaching.

And a finding that explains the product category’s survival: “Parents who liked the DVD tended to overestimate how much their children had learned from it.”

What this does and does not license

It supports a narrow, defensible claim. For children under about three, learning transfers less well from a recording than from a person, most severely for spatial and imitation tasks, less so for words. Commercial baby videos have not been shown to teach vocabulary, and one controlled test found they taught none.

It does not support the broader claim it is routinely used for - that screens as a class harm young children’s learning. The strongest single experiment here found that a screen with a responsive person on it worked as well as that person in the room. The variable is contingency, not the display.

And it does not licence quoting −0.53 as a settled magnitude. The corrected figure is −0.47, from a literature the analysts themselves found to be underpowered and asymmetric.

How screens affect reading in older children and adults is a different measurement with a different answer, set out separately.

Limitations and disputes

• Significant publication bias was detected in the meta-analysis (Egger z = −7.84, p < .001); the trim-and-fill corrected effect is −0.47, not −0.53.
• Median statistical power across the pooled studies was about 35%.
• Under-twelve-month-olds are barely represented - seven effect sizes out of 122.
• The contingency experiment had twelve children per condition and a warm-up flaw the authors disclose.
• The 2007 baby-DVD study is correlational, with parent-reported exposure and parent-reported vocabulary, and its authors did not claim causation.
• We have not printed a total participant count for the meta-analysis; the accessible record reports reports and effect sizes but not an aggregate N.
• The studies mostly used researcher-made videos, so generalisation to commercial children’s media is an inference, not a finding - with the DeLoache experiment the exception, because it used a commercial DVD.
• This is not paediatric advice. It is a description of what the learning-transfer experiments measured.

Most educational technology arrives, gets adopted, and is never properly evaluated. One Laptop per Child is the rare exception. It made a specific, testable promise, it was deployed at scale, and it was subjected to two large randomised controlled trials with a ten-year follow-up.

We therefore know the answer to a question that is usually left open. The answer is not the one the programme expected.

What was promised

Nicholas Negroponte announced the initiative in 2005: a rugged, cheap, low-power laptop, designed for children who could not yet read, distributed to the world’s poorest schoolchildren, with the explicit aim of transforming education. The machine was known from the start as the “$100 laptop”, and that number did most of the rhetorical work.

Two facts about the delivery are reasonably well attested. Around three million machines were eventually manufactured, with Uruguay and Peru accounting for most of them. And the laptop never sold at $100.

We are not printing a figure for what it actually cost. Contemporary press reports and later retrospectives give materially different numbers - one retrospective interview puts it near $130, a 2007 news report put it at $200 - and we could not resolve the discrepancy against a primary record. The verified claim is narrower and sufficient: the announced price was not reached.

The first randomised trial: no effect on maths or language

Cristia, Ibarrarán, Cueto, Santiago and Severín (2017), in the American Economic Journal: Applied Economics, evaluated the Peruvian deployment after 15 months across 318 rural primary schools.

The programme worked as an access intervention. Computers per student in treatment schools rose from 0.12 to 1.18 - a roughly tenfold increase - and this “translated into substantial increases in use of computers both at school and at home.” Whatever else went wrong, the laptops arrived and were used.

The learning result, in the authors’ own words: “No evidence is found of effects on test scores in math and language. There is some evidence, though inconclusive, about positive effects on general cognitive skills.”

That last clause kept a certain amount of hope alive for the better part of a decade.

The ten-year follow-up closed it

Cueto, Beuermann, Cristia, Malamud and Pardo (2024) returned to the same experiment and tracked outcomes over ten years, from 2009 to 2019, across 531 rural schools - 296 treatment, 235 control - using national examination data from 2009 to 2016 plus grade progression and educational attainment.

The precision here is what makes it decisive. This is not a study that failed to find an effect because it was too small to detect one. It is a study that can rule effects out:

The results “rule out positive statistically significant effects larger than 0.05 and 0.03 standard deviations in mathematics and reading, respectively.”

Grade progression showed a negative effect of 1.0 percentage point - individually insignificant, barely significant at the 10% level once multiple testing is accounted for. Effects on years of schooling were negative and insignificant.

The authors also tested the “general cognitive skills” hope from 2017 and it did not survive. Students showed digital skill gains of 0.4 standard deviations - they genuinely learned to use a computer - but “no positive effects on… cognitive skills based on the Raven’s Progressive Matrices test, verbal fluency, and coding tests.”

Their explanation is mundane and probably correct. Teachers reported only small increases in laptop use for academic purposes, despite having received training. The hardware reached the children. The teaching did not change around it.

Romania: the same intervention, and grades went down

The Peru result is sometimes read as a story about deployment failure in a poor rural school system. A separate natural experiment makes that reading harder to sustain.

Malamud and Pop-Eleches (2011), in the Quarterly Journal of Economics, studied Romania’s Euro 200 programme, which in 2008 issued roughly 35,484 vouchers worth about €200 towards a home computer, allocated by a strict income cutoff. That cutoff creates a clean comparison between families just above and just below the line. The authors surveyed 3,354 households in 2009.

The vouchers worked as intended: home computer ownership rose by 54.6 percentage points and computer use by 3.4 hours a week.

The academic results:

“Children who won a voucher to purchase a computer had significantly lower school grades.”

And now the complication that makes this more interesting than a simple cautionary tale. The same children scored higher on cognitive and computer tests:

• Raven’s Progressive Matrices: +0.327 SD (95% CI 0.059 to 0.595)
• Computer skills test: +0.268 SD (95% CI 0.040 to 0.496)

The time-use data explain the trade. Reporting at least an hour a day of homework fell by 9.6 percentage points; at least an hour of reading fell by 9.3 points; at least an hour of television fell by 7.0 points. The authors note that the additional computer time “was mostly spent playing games.” Only 9% of households had educational software, despite government provision.

One further finding is genuinely actionable, and it is the only thing in either study that is: parental rules about homework roughly halved the negative effect on maths grades. Rules about computer use, by contrast, successfully cut computer time by 3.2 hours a week without improving grades at all. Limiting the device did nothing; protecting the homework did something.

Note that Peru found no Raven’s gain while Romania found one. These are different populations, ages, devices and durations, and we are not going to reconcile them by picking a favourite. The consistent finding across both is the one that matters: digital skill improved, academic attainment did not.

What the programme got wrong

Not the hardware. The machine shipped, it survived rural conditions well enough to be used, and children learned to operate it.

What failed was the theory of change. OLPC’s premise was that access was the binding constraint - that children were held back by not having a computer, and that removing that obstacle would let learning follow. Both experiments put the computer in the child’s hands and both found that access was not the constraint. In Peru the teaching did not reorganise around the device. In Romania the device competed successfully with homework and reading and won.

This is a different question from whether laptops help university students taking notes in lectures, which is its own literature with its own answer, and a different question again from whether removing phones from schools improves learning, which is measured separately and is also less impressive than advertised. What OLPC tested was the distribution of hardware as an educational intervention in its own right, and it is the cleanest test of that idea anyone has run.

Limitations and disputes

• The Peru and Romania studies disagree about cognitive skills. Romania found a Raven’s gain of about a third of a standard deviation; the Peru ten-year follow-up found none. Both are reported here; neither is discounted.
• Malamud and Pop-Eleches acknowledge recall bias in retrospective time-use reports, a 52% survey response rate, and that their estimates are intention-to-treat rather than effects of actual usage intensity.
• The Peru follow-up relies on teacher and student self-reports for the classroom-adoption explanation.
• These are studies of specific programmes in specific places. Neither shows that computers cannot help children learn. They show that these distributions, in these systems, did not raise attainment.
• The unit cost of the XO laptop is not reported here because the available sources conflict and none is primary.
• Nothing here evaluates the OLPC software or the constructionist pedagogy behind it, only the outcomes the trials measured.

Play a 400 Hz tone into one ear and a 416 Hz tone into the other, and you will hear a third thing that is not there: a 16 Hz pulse, generated inside your head, because the two signals only meet after the ear.

That much is uncontroversial. It is what happens next that the entire industry rests on.

The claim sold with these recordings is that the illusory beat pulls your brain’s own electrical rhythms into step with it - 16 Hz for beta, “alert focus”; 40 Hz for gamma; 8 Hz for alpha, “relaxed concentration”. Tune the beat, tune the brain. It is called brainwave entrainment, and it is a testable mechanism, which is unusual and welcome.

Three separate bodies of evidence bear on it and they do not agree.

The pooled evidence is positive

Garcia-Argibay, Santed and Reales (2019), in Psychological Research, meta-analysed 22 studies yielding 35 effect sizes across memory, attention, anxiety and analgesia. Their headline: an overall effect of g = 0.45, 95% CI [0.28, 0.62] - medium, significant, and by their assessment consistent.

They also tested for publication bias and did not find it. Egger’s regression was non-significant (z = −0.019, p = .984); trim-and-fill suggested no missing studies; a p-curve analysis indicated genuine effects with estimated mean power of 86%. Heterogeneity was moderate (I2 = 61.3%).

Their moderator findings are specific enough to be useful. Masking the beats with white or pink noise made no difference to effectiveness. Exposure before the task, or before and during, outperformed exposure during the task only. Longer exposure worked better.

Basu and Banerjee (2023), in the same journal, narrowed the question to memory and attention alone: 15 studies, 31 effect sizes, g = 0.40. Also medium, also significant. But their systematic review alongside the meta-analysis found “conflicting results, especially concerning theta and beta’s efficacy on memory… and attention-related tasks,” and they closed by calling for “more research with robust study designs.”

There is one number in the 2019 analysis that should temper how you read both. The mean sample size was about 20 participants per group. These are small studies pooled into a medium effect.

The experiments that tested attention directly found nothing

If binaural beats improve sustained attention, the place to see it is a long, dull vigilance task - exactly the setting where attention reliably decays.

Robison, Obulasetty, Blais, Wingert and Brewer (2022), in Psychological Research, ran three experiments with beta-frequency (16 Hz) stimulation during sustained attention tasks, funded by the US National Science Foundation. Their own summary: “Overall, we did not observe any consistent evidence that binaural beat stimulation can augment sustained attention or its subjective and physiological correlates.”

The detail is worth having, because it shows how a positive-looking result dissolves. Experiment 1 did produce faster reaction times under stimulation - but the effect was “modest in magnitude, and model comparisons using Bayes Factors were indiscriminate.” Experiment 2 found a shallower decline over time but no change in what participants reported thinking about. Experiment 3 found nothing in pupil measures of arousal.

Melnichuk, Cooper and Hawk (2025), in Scientific Reports, went after the mechanism and the behaviour in the same study. 80 undergraduates were assigned across a 2×2×2×2 design - beta or gamma beats, 340 or 400 Hz carrier tone, onset before or with the task, masking noise present or absent - and performed a 33-minute, 1,200-trial continuous performance task.

The EEG worked. Beta beats raised power at 16 Hz (p = 0.001); gamma beats raised power at 40 Hz (p < 0.001). Entrainment, in the narrow electrical sense, happened. White noise attenuated it, particularly for gamma.

The attention did not follow. Binaural beats produced no reduction in the decline over time: F(1,53) = 0.1, p = 0.751. Two isolated cells did show better overall performance - gamma with the 340 Hz carrier (p = 0.002), and beats with white noise (p = 0.033) - but this is a sixteen-condition design with roughly eight participants per cell, and the authors list small cell sizes first among their limitations. We are reporting those two results as what they are: single cells in a wide factorial, not findings to build on.

The authors’ reading is measured: binaural beats “may modulate other cognitive aspects rather than sustained focus.”

The mechanism everybody assumes is not established

This is the part that reframes the whole question.

Ingendoh, Posny and Heine (2023), in PLOS ONE, systematically reviewed every study that measured what binaural beats do to brain oscillations with EEG in healthy adults, under controlled conditions. Fourteen studies survived their quality screening — three more were excluded during quality assessment for methodological inadequacy.

The tally:

Their assessment of the field’s condition is blunt. Twelve of sixteen studies lacked a control group design. Only two took multiple pre- and post-measurements. Stimulation ran anywhere from 1 to 30 minutes, carrier frequencies from 100 to 900 Hz, and control conditions ranged from silence to pink noise to monaural beats. Researchers could not even agree what counts as evidence of entrainment, using absolute power, relative power, normalized power, steady-state responses, frequency-following responses, source localization and connectivity measures more or less interchangeably.

One pattern in their results sits awkwardly beside the 2019 meta-analysis: “all studies embedding binaural beats in pink noise reported no effects in line with the brainwave entrainment hypothesis.” Garcia-Argibay and colleagues concluded masking made no difference to behavioral outcomes; Ingendoh and colleagues found it abolished the electrical ones. Those are different measurements, so this is not a contradiction - but it is a reason to doubt that the behavioral effect, where it appears, is running through entrainment at all.

Their conclusion: the question “cannot [be] settled at this point,” and existing studies reporting psychological or physiological changes attributed to entrainment “should be considered with caution.”

The review declared no funding and no competing interests.

So what should you do with a focus playlist

Here is the honest position, which is less satisfying than either the marketing or the debunking.

Something may be happening. Two independent meta-analyses find a small-to-medium pooled effect that survives their own bias checks. That is not nothing, and it would be overreach to call binaural beats a placebo on this evidence.

But the specific claim - that a 16 Hz beat produces focus by synchronizing your brainwaves - has an EEG literature that contradicts it more often than it supports it, and the two best-designed tests of sustained attention specifically both came back empty. The 2025 study is the sharpest version: the entrainment was measurable, and it still did not slow the decline in attention.

If a track helps you settle into work, that is a real experience and there is no reason to stop. Just notice what the evidence licenses. It does not license the frequency claims printed on the thumbnails, and it does not distinguish binaural beats from any other steady, unobtrusive sound you happen to find conducive.

Which is worth saying because the neighboring questions have clearer answers. Whether unstructured background noise helps depends on who you are and how loud it is - that is measured. Whether music helps turns substantially on lyrics, and two meta-analyses disagree for a reason that is now understood. Neither of those literatures depends on a mechanism nobody has demonstrated.

Limitations and disputes

• The two meta-analyses pool small studies. Mean group size in the 2019 analysis was about 20. Both author teams call for stronger designs.

• The positive and negative findings measure different things. The meta-analyses pool across memory, attention, anxiety and pain; the null experiments tested sustained attention specifically. It is possible binaural beats do something real that is not sustained focus.

• We did not obtain per-domain effect sizes for attention alone from either meta-analysis; the accessible records reported the overall pooled figures and subgroup descriptions but not attention-specific confidence intervals. Those are not printed here.

• The 2025 factorial design has roughly eight participants per cell. Its two significant condition effects are reported as provisional and should not be read as showing which frequency to choose.

• No study reviewed here tested the commercial products people actually use, which mix beats with music, and none of the evidence transfers cleanly to those.

• Funding was clean in the studies that found nothing - the 2022 experiments were NSF-funded and the 2023 review declared no funding or competing interests. We did not obtain funding declarations for every pooled study inside the meta-analyses.

Attention on a long monitoring task really does fade, as radar research showed in 1948. The popular ten-minute lecture limit has no primary data behind it.

Two claims get muddled together whenever anyone talks about attention running out.

The first is that human attention capacity has shrunk over the last two decades. That is a claim about history, it is weakly supported, and it is handled elsewhere on this site - the eight-second figure and the switching-rate measurements are traced in full on the attention span page, and whether social media specifically shortened anything is a separate question with its own answer. Neither is repeated here.

The second claim is much older, much better evidenced, and almost never distinguished from the first: within a single continuous task, attention measurably decays, and it starts doing so quickly. That is the vigilance decrement, and it has been studied since the Second World War.

This page is about the second claim, and about what happened when somebody attached a number to it and applied it to lecture theatres.

The founding experiment came out of radar

The decrement was discovered because it was operationally urgent. Radar operators in the Second World War were missing contacts late in a watch, and nobody could say by how much.

Norman Mackworth (1948), in the Quarterly Journal of Experimental Psychology, built an apparatus to measure it. The Clock Test presented a pointer moving in short jumps, roughly one per second, around a blank dial. Occasionally the pointer made a double jump. The observer’s only job was to press a button when that happened, for a long, featureless watch.

The apparatus survives as a standard task. What it showed survives too.

Klein and Feltmate (2025), reviewing the seventy-five years since, describe Mackworth’s result as “a decline in the probability of detecting brief targets as the time on task increases,” robust “after the first half hour on the task.” The performance curves the review reproduces run from roughly 67–85% detection at the start down to roughly 50–75% by 80 to 120 minutes in.

We have not printed Mackworth’s sample size or his exact miss rates. The 1948 paper itself could not be retrieved, and the figures above are as reported in the 2025 review, which is a secondary description of a primary source. The direction and the timing are firm; the precise numbers should be read from the original by anyone who needs them.

One detail from Mackworth’s work matters more than the numbers, and Klein and Feltmate single it out. A salient interruption, or feedback on performance, could reduce or eliminate the decrement. Urging the observer to try harder had no effect at all.

That is the finding people skip, and it is the practical one.

Nobody agrees on why it happens, and that is not a dodge

For thirty years the field argued about whether the decrement is a genuine loss of perceptual sensitivity - you actually become worse at seeing the signal - or a shift in decision criterion - you become more reluctant to call something a signal.

The argument opened badly. Klein and Feltmate note that in the two earliest papers to apply signal detection theory to the problem, “there was immediately a dramatic conflict”: Broadbent and Gregory (1963) found criterion shifts, while Mackworth and Taylor (1963) found reduced detectability. In the 1970s the consensus settled on criterion; then See, Howe, Warm and Dember (1995) meta-analysed the literature and found that genuine sensitivity declines were, in Klein and Feltmate’s summary, “much more prevalent than previously believed.” Then in 2016 Thomson and colleagues pointed out that a conservative criterion shift “can masquerade as a loss in sensitivity” when false alarms are already near zero.

The 2025 review’s verdict is unusually honest for a research summary: “There will not be one single explanation for the vigilance decrement. There is an element of truth to each one, and they will often be operating simultaneously.”

We have not printed the pooled effect size from See et al. (1995); the paper could not be retrieved directly and we are not willing to quote a meta-analytic number at second hand.

The old explanation was wrong in an interesting way

The intuitive account of a long dull watch is that it is understimulating - you drift off because nothing is happening. If that were right, the fix would be stimulation.

Warm, Parasuraman and Matthews (2008), in Human Factors, argued the opposite and assembled the evidence for it. Their abstract states the traditional view plainly and then dismantles it: vigilance tasks “have typically been viewed as undemanding assignments requiring little mental effort,” and the decrement “considered to result from a decline in arousal brought about by understimulation.”

What they found instead, across four lines of evidence - task-type comparisons, subjective workload ratings, transcranial Doppler measurements of cerebral blood flow, and stress measures - was that “vigilance requires hard mental work and is stressful.” Subjective workload during vigilance is high and rises with processing demands. Cerebral blood flow changes track the performance decline. Task engagement falls and distress rises, and both scale with difficulty.

Monitoring something for nothing is not boring in the sense of being easy. It is expensive.

Where the ten minutes came from

This is the part that should change how you read teaching advice.

Somewhere in the transfer from the vigilance laboratory to the seminar room, the decrement acquired a number: student attention drops after ten to fifteen minutes, so break the lecture into chunks.

Wilson and Korn (2007), in Teaching of Psychology, went looking for the study behind it. They reviewed the note-taking studies, the observational studies, the retention studies, the self-report studies and the one physiological measurement that the claim is built on, and concluded that “the research on which this estimate is based provides little support for the belief that students’ attention declines after 10 to 15 min.” Several of the foundational sources, they found, “were not empirical studies, but secondary sources or personal observations.” One influential source was a personal observation supported by colleagues’ anecdotes. Another reported judges’ ratings without reliability estimates.

Bradbury (2016), in Advances in Physiology Education, followed the citation chain further and found where it terminates. Multiple standard teaching texts assert the ten-to-fifteen-minute limit; Bradbury’s finding is that “all of the above reports do not provide any primary data on attention but are content to all cite the same single initial report as the basis for the 10- to 15-min attention span assertion.”

That report is Hartley and Davies (1978), Note taking: A critical review.

Two things are wrong with it as a foundation. First, as Bradbury puts it, “attention span is not actually the subject of the article; rather, the subject of the manuscript is in fact ‘note taking.’ ” Second, the data point in the wrong direction: the decline in note-taking occurred during the last ten minutes of the lecture, not the first, and reflected the lecture running out of content rather than the students running out of attention. Hartley and Davies themselves conceded that note-taking is not a good proxy for attention.

Bradbury’s conclusion: “the available primary data do not support the concept of a 10- to 15-min attention limit.”

So the number that shaped a generation of teaching guidance came from a review of note-taking, measured the wrong thing, at the wrong end of the lecture.

What survives

Strip out the invented number and something real is still standing, and it is more useful than the number was.

Sustained attention to an unchanging monitoring task declines within tens of minutes. That is measured, replicated since 1948, and not in dispute. What is in dispute is the mechanism, and what has been refuted is the specific claim that human attention has a fixed ten-to-fifteen-minute ceiling that applies to lectures.

Three things follow that the evidence does support:

Effort does not fix it. Mackworth found that instructing observers to be more attentive did nothing, while interruptions and feedback did something. If your response to fading concentration is to try harder, you are using the one lever that was tested and found not to move.

It is not underload. Warm and colleagues’ central finding is that monitoring is resource-hungry and stressful. Treating a long watch as an easy task that people fail through laziness misreads what is happening to them.

Individual variation is large. Wilson and Korn’s specific criticism of the ten-minute literature is that it “failed to take into account individual differences, relying solely on group differences.” Any single number for how long anyone can concentrate is doing something the data do not license.

The interruption cost of switching away from a task is a different measurement with a different literature, and where the working time actually goes is set out separately.

Limitations and disputes

• Mackworth 1948 was not read in the original. Its description here rests on Klein and Feltmate’s 2025 review, and no sample size or exact detection rate is printed.
• See et al. (1995) was not retrievable. Its conclusion is reported as characterised in the 2025 review; no pooled effect size appears here.
• The mechanism is genuinely unresolved. Sensitivity loss, criterion shift, resource depletion and mind-wandering accounts all have support, and the reviewers say so.
• Laboratory vigilance tasks are not lectures, meetings or driving. The decrement is measured on deliberately monotonous target-detection tasks. Generalising the timing to rich, self-paced activities is exactly the error the ten-minute claim made.
• The debunking of the ten-minute figure is not a claim that lectures hold attention indefinitely. It is a claim that no primary data established that particular limit.

There is a question underneath this one that most coffee articles skip, and it is the interesting part.

If you drink coffee every day, then by the time you reach for the first cup you have been without caffeine for eight or ten hours. You are not a neutral baseline. You are in mild withdrawal. So when the coffee makes you feel better, there are two possible readings: the drug improved you, or the drug returned you to where you already were.

Researchers have been arguing about which it is for more than twenty years, and the argument is not settled. What follows is what each side actually measured.

The largest test of the withdrawal question

Rogers, Heatherley, Mullings and Smith (2013), in Psychopharmacology, ran the study designed to separate the two readings. It is large by the standards of this field: 369 adults, aged 18 to 62, split into 157 non-low consumers and 212 medium-high consumers, all required to abstain from caffeine from 7pm the previous evening. Double-blind, placebo-controlled. The caffeine group received 100 mg at 11:15am and a further 150 mg at 12:45pm.

The design matters. If caffeine is a genuine enhancer, it should lift the people who rarely drink it at least as much as the habitual drinkers. If it is withdrawal reversal, the habitual drinkers should start below the light drinkers and be restored to their level - but not above it.

What they found:

• Caffeine did improve physical speed in both groups: “faster tapping speed and faster simple and choice reaction times.”
• The medium-high consumers arrived at the lab already impaired - lower mental alertness and greater sleepiness than the light consumers, before anything was given.
• Caffeine restored those consumers to the light consumers’ placebo level. It did not lift them past it.
• In the light consumers, caffeine reduced sleepiness but “had little effect” on mental alertness, because it also raised anxiety and jitteriness, which cancelled the gain.
• And the part the title refers to: caffeine “did not affect mental alertness, or the number of errors made on the recognition memory and choice reaction time tasks in non-low consumers.”

The authors’ summary of what the habitual drinker’s morning lift actually is: “a return to the normal state of affairs… rather than enhancement to above the normal state.”

Faster, not smarter. That is where the paper’s title comes from.

This is the mature version of an argument the same group set out eight years earlier. James and Rogers (2005), also in Psychopharmacology, reviewed the experimental literature and concluded that the apparent benefits “are almost wholly attributable to reversal of adverse withdrawal effects associated with short periods of abstinence from the drug.” Their methodological point is the one that still bites: an ordinary placebo-controlled caffeine study, run on habitual drinkers who abstained overnight, cannot tell the two explanations apart. Only long-term withdrawal designs can.

The other side, and it is not a fringe position

Einöther and Giesbrecht (2013), reviewing double-blind placebo-controlled attention studies in the same journal, reached the opposite conclusion: caffeine “improves performance on simple and complex attention tasks, and affects the alerting, and executive control networks,” and “caffeine’s effects cannot be attributed to withdrawal reversal.”

One thing a reader should know when weighing that review: both authors were employed by Unilever R&D Vlaardingen at the time of publication. Unilever is a large tea business. That is an author affiliation, not a finding of misconduct, and the review is a serious piece of work - but the disclosure belongs next to the conclusion, and the paper’s accessible record carries no competing-interests declaration.

The most recent quantitative answer comes from a source with no such adjacency. Kløve and Petersen (2025), in Psychopharmacology, meta-analysed 31 randomised, double-blind, placebo-controlled trials of pure caffeine in 1,455 healthy, rested adults. Their pooled results:

| Outcome | | Effect (Hedges’ g) |

| Reaction time | | 0.28 |

| Accuracy | | 0.27 |

Doses of 200 mg and above produced larger effects than smaller doses. Reaction time improved roughly linearly with dose; accuracy improved and then declined, which is to say more is not simply better.

And the finding that speaks directly to Rogers: the authors report that “the effect of caffeine was not related to differences in habitual caffeine consumption.” If the benefit were purely withdrawal reversal, heavy consumers should show the biggest gains. In this pooled analysis they did not.

So the honest state of the evidence is this. There is a real, small-to-moderate acute effect on both speed and accuracy across 31 trials. There is also a large, well-designed single study showing that for habitual drinkers the alertness part is restoration rather than enhancement. Those two things can both be true - the meta-analysis pools rested participants across many consumption histories, while Rogers built his study specifically to expose the consumer-group difference - and nobody has yet run the design that would settle it.

Two figures we could not verify and are therefore not printing: the heterogeneity statistics and publication-bias assessment for the 2025 meta-analysis, neither of which appeared in the accessible record.

The cost that arrives eight hours later

There is one effect nobody disputes, and it is the one people notice least.

Drake and colleagues (2013), in the Journal of Clinical Sleep Medicine, gave healthy sleepers 400 mg of caffeine - roughly two to three cups of coffee - at six hours, three hours and zero hours before bed, with a placebo day for comparison. Taken six hours before bedtime, caffeine reduced objectively measured total sleep time by more than one hour.

The part worth sitting with: participants did not notice. The publishing association’s own record of the study states that “subjective reports suggest that participants were unaware of this sleep disturbance.”

Sample size: twelve people. That is small, and we are reporting it as a signal rather than a settled quantity, partly for that reason and partly because the full paper could not be retrieved - the journal blocks automated access and the archive copy sat behind a challenge page. The figures above come from the American Academy of Sleep Medicine’s own record of the study it published. We have not printed the minute-level numbers.

Still, the shape of it is the thing. A 3pm coffee costing an hour of sleep you cannot feel yourself losing is exactly the loop that makes tomorrow’s coffee feel necessary.

What this means if you drink coffee to concentrate

Not much changes, which is itself worth saying. Caffeine is not a fraud. It makes you faster, and on the pooled evidence it makes you somewhat more accurate too.

What is not supported is the stronger belief people carry into a working day — that caffeine adds a capacity you would not otherwise have. For someone who drinks it daily, at least part of what the first cup does is undo what its absence did. Rogers found that the medium-high consumers, at baseline, were measurably worse off than people who barely drink the stuff.

And the effect is dose-shaped, not dose-unlimited. Accuracy in the 2025 analysis improved up to a point and then fell away.

This page deliberately stays inside one question. Whether unstructured background noise helps is a separate literature with a different answer. Whether music does is another one again, and it turns on lyrics. The claim that a particular kind of sound can tune your brainwaves into a focused state is a third and separate mechanism claim, tested differently, and it is not covered here.

Limitations and disputes

• The central question is open. Two credible sources reach opposite conclusions about withdrawal reversal. We have reported both rather than choosing.
• Rogers et al. (2013) report their own contamination problem: 42% of the non-low consumer group had detectable salivary caffeine or paraxanthine at baseline, which would understate the withdrawal effect they were measuring.
• The 2013 study used self-selected consumer groups. The authors argue this reflects exposure history rather than pre-existing differences, but it is not random assignment to a caffeine habit, and no ethical study could be.
• The sleep study had twelve participants and could not be read in the original.
• This is about healthy, rested adults doing laboratory attention tasks. It is not about sleep-deprived shift workers, for whom the applied literature is different, and it is not medical advice about how much caffeine anyone should consume.
• Nothing here is a claim about caffeine and health. That is a separate body of evidence and this page does not touch it.

Two claims travel together through every article about writing for the web. People scan pages in an F shape. And they read only 20 to 28% of the words.

The first is a real observation that its own authors have since narrowed considerably. The second was never measured by anyone.

Where the F-pattern comes from

Jakob Nielsen (2006), in a Nielsen Norman Group article, reported that “we recorded how 232 users looked at thousands of Web pages” and described a dominant F-shaped scanning pattern.

That is the entire method statement. No statistics, no significance tests, no peer review. The document is published by a usability consultancy.

Kara Pernice (2017), updating it for the same organisation, reports that the famous heatmaps came from “more than 45 people” and that a 2017 follow-up used 47 participants. So the canonical N of 232 is not the N behind the canonical image.

The 2017 update does not retract, and it is more interesting than the original. It narrows the claim to “the content area” and calls the F-pattern “the default pattern when there are no strong cues.” It names five other patterns - layer-cake, spotted, marking, bypassing and commitment. And it reframes the F-pattern as a symptom of bad formatting rather than a law of reading: “Good web formatting reduces the impact of F-scanning.”

No prevalence figures are given for any of the six patterns, in either the 2006 or the 2017 document. Nobody has published how often the F actually occurs.

The statistic that was never measured

This is the part worth carrying away.

Nielsen (2008) wrote: “On the average Web page, users have time to read at most 28% of the words during an average visit; 20% is more likely.”

He attributes it to Weinreich, Obendorf, Herder and Mayer (2008), “Not Quite the Average: An Empirical Study of Web Use,” in ACM Transactions on the Web - a browser-instrumentation study of 25 users, retaining 45,237 page views after cleaning, with data collected in 2005.

That study did not measure words read. It logged dwell time.

The 20–28% figure is an arithmetic conversion of time-on-page into words-read, using an assumed reading speed. A behavioural log with no eye-tracking has been turned into a precise-sounding statistic about reading, and then cited for eighteen years as though someone had counted.

We should be clear about what we did and did not verify. We could not open the Weinreich paper - two routes failed. So we are reporting the figures as Nielsen describes them, and our claim is about the inference, which is visible in his own account of it: he describes a dwell-time study and produces a words-read percentage.

What the peer-reviewed work found

The academic eye-tracking literature does not reproduce a single dominant pattern.

Djamasbi, Siegel and Tullis (2011), at HCI International, found the opposite of an F: “users’ viewing pattern in our study was more scattered than those reported in previous studies, which used pages with a relatively less complex visual hierarchy.”

That is a specific and useful boundary condition. The F appears on simple pages. Complex visual hierarchies produce scattered viewing - which is exactly what the 2017 update independently concluded when it called the F a symptom of poor formatting.

Strzelecki (2020), in Information, systematically reviewed 56 peer-reviewed empirical eye-tracking papers from 2004 to 2019. It reports no single scanning path - users adopt breadth-first or depth-first strategies.

It also delivers a methodological verdict on the field that should temper everything above: roughly half the reviewed studies fell below the recommended 30-participant minimum, samples were predominantly university students, and “few studies replicated prior research.”

That review covers search engines rather than content pages. As far as we could establish, no systematic review of general web-page reading patterns exists. That absence is itself worth stating, given how confidently the F-pattern is taught.

What this leaves you with

Something real is being described. People do scan, they do favour the top and left of a content block in left-to-right scripts, and attention does fall off down the page. Nobody disputes the phenomenon.

But it is not a law, and it is not a target. The organisation that named it now calls it what happens when formatting gives the eye nothing to grab. Designing for the F is designing for failure.

Stop quoting the 20–28% figure. It is a conversion of dwell time, not a measurement of reading, and the study it cites did not report it.

Be sceptical of any web-reading statistic without an N and a method. In this literature the foundational documents are consultancy publications with participant counts and nothing else, and the peer-reviewed studies that do exist are mostly small, mostly students and mostly unreplicated.

The practical implication survives the debunking. If scanning patterns are a symptom of formatting, then headings, short paragraphs and front-loaded sentences are worth using - not because they defeat an F-shaped eye movement, but because they give attention somewhere to land. That is the same conclusion, held for a better reason.

For why interfaces teach people to discard whole regions of a page by shape, see banner blindness. For what happens to comprehension when links interrupt reading, see hyperlinks and comprehension. For reading comprehension by surface, see screen reading vs paper.

Limitations of this article

The founding sources are not peer-reviewed. Both the 2006 and 2017 documents are consultancy publications that report participant counts and no statistics. We have said so rather than treating them as research, and we have not printed any effect from them because none is given.

We could not read the Weinreich paper. Two routes failed. Every figure attributed to it here is read in Nielsen’s secondary account, not in the primary. Our argument concerns the inference he describes, not numbers we have checked. Anyone building on this should obtain the primary text and confirm that no words-read percentage appears in it.

The N discrepancy is unresolved. The 2006 article says 232 users; the 2017 article attributes the famous heatmaps to “more than 45 people.” We report both and cannot reconcile them.

We located only one peer-reviewed study directly contradicting the F-pattern, and it is a conference paper. This article would be stronger with more, and we did not find more.

The systematic review we cite covers search-engine results pages, not content reading. We use it for its methodological findings about the field, not as a review of the F-pattern, which it does not address.

No prevalence data exists for any scanning pattern, so nothing in this article can tell you how often the F occurs.

Most countdown timers in a large crawl could be shown to be fake; most low-stock counters could not. The CMA case, the scarcity meta-analysis and where these cues backfire.

There are two confident positions on scarcity messages, and the evidence supports neither.

One says the counters are all lies. One says they are just information. The measured answer is that most low-stock counters could not be shown to be fabricated, most countdown timers could, and that the real problem is something else entirely: you have no way to tell which is which.

That is why a competition regulator got involved rather than a fraud squad.

What the regulator actually did

On 6 February 2019, the UK’s Competition and Markets Authority obtained signed undertakings from Agoda, Booking.com, Expedia (with ebookers and Hotels.com), and trivago. The investigation had opened on 27 October 2017 and enforcement action began on 28 June 2018. The compliance deadline was 1 September 2019. On 13 September 2019, a further 25 online booking sites signed up to the same sector-wide principles.

The CMA identified four practice areas. Two matter here.

Pressure selling. The CMA named “strategically placing sold out hotels within search results to put pressure on people to book more quickly” - and, critically, misleading popularity indicators where other customers were viewing different dates from the ones you were looking at.

That last detail is the sharpest fact in this article. “12 people are looking at this right now” was, in identified cases, true of a hotel and false of your trip.

Search results ranking. Commission paid by hotels affected the order results appeared in, without that being disclosed.

The other two areas were discount claims - such as comparing a weekend rate to a weekday rate, or a luxury suite to a standard room - and hidden charges, with the CMA requiring all compulsory taxes and fees in the headline price.

A parallel European track reached similar ground. Under the EU’s Consumer Protection Cooperation Network, Booking.com proposed EU-wide commitments that were accepted on 18 December 2019, following findings that offers, discounts and prices were presented in ways that impeded informed comparison.

How often are they fake?

Mathur and colleagues (2019), at CSCW, crawled roughly 11,000 shopping websites and 53,000 product pages and counted.

They also checked which were fabricated, and the answer splits sharply.

157 of 393 countdown timers on 140 sites either reset after expiry with the offer still valid, or expired with the offer still valid - about 40%.

29 activity notifications on 20 sites were falsely generated or made misleading statements.

17 of 632 low-stock messages on 17 sites - about 2.7% - were confirmed to use a scheduled deterministic pattern or randomly generated stock values.

So the honest claim is not “the counters are fake.” Verified fabrication is concentrated in timers and popups. For low-stock counters, only a small fraction could be proved fabricated - and the rest are simply unverifiable from outside.

The study also found that 22 third-party services enabling these patterns were embedded in 1,066 sites, some openly advertising features that enable deceptive messages. The cues are frequently not built by the retailer at all; they are bought.

Do they work?

Barton, Zlatevska and Oppewal (2022), in the Journal of Retailing, meta-analysed 131 studies and 416 effect sizes, N = 22,655 for purchase intentions.

So the cues work, at a small-to-moderate effect.

But note the moderator test, because it kills a tempting conclusion: the difference between scarcity types was not significant, Q(2) = 1.76, p = .414. Despite the gap between 0.39 and 0.28, “only 2 left beats 12 people viewing” is not an established finding.

And where they fail

Fenko, Keizer and Pruyn (2017), at ICORIA, tested 268 people drawn from a national opera and ballet ticket database - 94.8% of whom had previously bought tickets online - in a two-by-two design crossing social proof with scarcity.

Scarcity backfired. Main effect on purchase intention F(1) = 4.85, p < .05, with the scarcity condition lower: M = 2.90 (SD .90) against M = 3.14 (SD .88). The direction was contrary to the authors’ hypothesis.

Social proof did nothing: “The main effects of social proof… on purchase intention were non-significant (all p’s > .05).”

Two caveats belong with it. This is conference proceedings, not a peer-reviewed journal, and the context is opera tickets rather than hotels. The social-proof manipulation check was only marginal (t = 1.91, p = .057), so that null is partly a manipulation-strength problem. It is suggestive, not decisive.

Still, it is a real result showing these cues can reverse in an audience that recognises them.

There is a context effect too. Song, Noone and Han (2019), in the International Journal of Hospitality Management, found the effectiveness of unit scarcity against option scarcity depends on booking lead time: with a long lead time, unit scarcity messages worked better “as they are perceived as more credible, and induce higher perceptions of a sellout risk”; with a short lead time, both worked equally. We could not obtain that study’s sample size or statistics and are reporting its stated finding only.

What this leaves you with

The cues do move behaviour, modestly. Pooled δ = 0.31 across 131 studies.

Most low-stock counters were not shown to be fabricated. If you have been telling people the numbers are all invented, the best available measurement does not support that.

Most countdown timers were. Around 40% either reset or expired with the offer intact.

The counter you cannot check is the problem, not the counter that lies. A regulator’s actual finding was subtler than fabrication: popularity indicators counted people looking at different dates. That is technically true and practically misleading, and no amount of scepticism on your part could detect it.

Recognising the tactic can reverse it. In an experienced audience, scarcity messaging reduced purchase intention.

Nothing here is consumer or purchasing advice, and nothing here is legal advice. This article reports what a regulator found and what the research measured.

For friction deliberately built into leaving, see the roach motel dark pattern. For decline options framed to shame you, see confirmshaming. For visible counts on content rather than commerce, see what visible metrics do. For the wider pattern, see the persuasive design guide.

Limitations of this article

We read the CMA’s practice findings in its press release, not in the five signed undertakings themselves. The dates, signatories and deadlines come from the case record; the characterisation of the practices comes from the release.

We could not obtain the sample size or statistics for Song, Noone and Han (2019), and have printed neither.

The backfire study is conference proceedings, in an opera-ticket context, with a marginal manipulation check on one of its two factors.

We could not obtain compliance percentages from the European enforcement track -how many sites were checked and how many were non-compliant - and have therefore reported only the dates and outcomes.

We did not print any conversion-rate figures from agency case studies or vendor marketing, of which there are many and none with a method.

The prevalence crawl is from 2019. The proportions may have changed, particularly given the enforcement action described above.

One boundary note. This article covers commercial urgency and scarcity cues in a transaction, adjudicated by regulators. Visible engagement counts on content - likes, view counts - are a different subject and are covered at position 14.

The honest answer is uncomfortable and it is on the record, published by the company that runs most of them.

In December 2014, Google’s own security blog announced that “today’s Artificial Intelligence technology can solve even the most difficult variant of distorted text at 99.8% accuracy. Thus distorted text, on its own, is no longer a dependable test.”

That was twelve years ago. The puzzle stopped separating humans from machines, and it kept being shown to humans.

The task was always designed to take something from you

This is not a conspiracy theory. It is in the founding paper.

Von Ahn, Blum, Hopper and Langford (2003), at Eurocrypt, wrote in their abstract:

“Much like research in cryptography has had a positive impact on algorithms for factoring and discrete log, we hope that the use of hard AI problems for security purposes allows us to advance the field of Artificial Intelligence.”

And: “captchas based on these AI problem families, then, imply a win-win situation: either the problems remain unsolved and there is a way to differentiate humans from computers, or the problems are solved and there is a way to communicate covertly on some channels.”

Five years later the same group made the harvesting explicit. Von Ahn, Maurer, McMillen, Abraham and Blum (2008), in Science, described reCAPTCHA feeding unrecognised words from book scans into CAPTCHA challenges:

• reCAPTCHA transcription accuracy: 99.1% at word level - 216 errors in 24,080 words
• Standard OCR on the same words: 83.5% - 3,976 errors
• Professional human “key and verify” transcription: more than 99%

So unpaid users matched professional transcribers. In one year, the paper reports, people solved more than 1.2 billion CAPTCHAs, deciphering over 440 million words - “equivalent to over 17,600 books manually transcribed.”

Google bought reCAPTCHA in September 2009 and said why on its own blog: to “increase fraud and spam protection for Google products but also to improve our books and newspaper scanning process.”

The labour harvesting was designed, disclosed and celebrated. That part is documented fact.

What it costs you, measured

Searles, Prapty and Tsudik (2025), at the Symposium on Usable Security and Privacy, ran a 13-month live field study on a real account-creation service. Participants were unwitting - going about ordinary business, not told they were in a study.

Over 3,600 distinct users; 9,141 valid records; over 9,000 reCAPTCHAv2 challenges.

About 20% of sessions escalated to an image challenge. That 10.3 seconds is the number that matches your experience of the thing; the 3.53-second average is dominated by the 80% of cases where you just clicked a box.

The very large number, and why we are not printing it

That paper is also the source of a figure now everywhere: 819 million hours spent on reCAPTCHA, worth $6.1 billion in free wages, and a separate claim of $888 billion in cookie value to Google.

We traced how they were built, and we are not printing them as findings.

The hours figure assumes 100 million reCAPTCHAs per day, every day, from 2009 to 2023 - a rate taken from an archived 2010 Google FAQ and held constant for fourteen years. It also uses a US federal minimum wage of $7.5; the actual figure has been $7.25 since July 2009.

The arithmetic is internally consistent. The inputs are not defensible. And note the tension with the Science paper’s own 2008 figure of 1.2 billion CAPTCHAs in a year - about 3.3 million a day, not 100 million.

The $888 billion figure fails harder. It multiplies sessions by a per-cookie value drawn from a study measuring cookies over an average 279-day lifetime, so one user generates many sessions but one cookie. It uses that study’s mean of €2.522 when the median is €0.022 - a skew of roughly 115 times. And the value in question is publisher ad-impression revenue on a European ad exchange, not revenue accruing to Google from reCAPTCHA.

We are describing these numbers rather than repeating them, because they are the most widely circulated claims about CAPTCHAs and they do not hold up.

One more correction, in fairness to Google. We could find no current Google document stating that reCAPTCHA v2 image answers train its machine-learning models. Its current documentation says the opposite: “Data gathered is used only for reCAPTCHA’s operation and security, not for personalized advertising by Google.” The documented harvesting belongs to the text and book-scanning era. The image-era claim is an inference, and we are labelling it as one.

The machines already win

Searles and colleagues (2023), at the 32nd USENIX Security Symposium, had 1,400 participants solve 14,000 CAPTCHAs across modern schemes.

Human solving times ran from 3.1–4.9 seconds for a reCAPTCHA checkbox to 18.1–42.8 seconds for the harder puzzle schemes, with accuracy between 50% and 85% depending on scheme.

Against published bot performance: bots solved reCAPTCHA click in 1.4 seconds at 100% against humans at 3.1–4.9 seconds and 71–85%. On distorted text, bots managed under a second at 99.8% against humans at 9–15.3 seconds and 50–84%.

Important caveat: the bot figures in that comparison are taken from prior published attacks, not run by these authors. It is a literature comparison, not a head-to-head experiment.

Plesner, Vontobel and Wattenhofer (2024), accepted at COMPSAC, report: “we can solve 100% of the captchas, while previous work only solved 68–71%.” They also found “no significant difference in the number of challenges humans and bots must solve to pass.”

Their own conclusion notes the experiments covered 50 to 100 runs and recommends future work “possibly ranging in the hundreds or thousands.” A 100% rate over roughly a hundred runs is not a 100% rate in general.

The system also leans on something other than the puzzle. With cookies and browsing history present, the median number of challenges required dropped from 5.00 to 2.00. Which is the tell: the current version, reCAPTCHA v3, per Google’s own developer documentation, “will never interrupt your users” and instead “returns a score (1.0 is very likely a good interaction, 0.0 is very likely a bot),” with a default threshold of 0.5.

The puzzle was never the security. It is increasingly not even the mechanism.

Who pays most

Bigham and Cavender (2009), at CHI, tested audio CAPTCHAs with 162 participants - 89 blind, 73 sighted.

A blind user takes roughly five times as long and succeeds at about half the rate. An optimised non-visual interface raised blind first-try success to 68.5%, which shows the gap is a design failure rather than a law of nature.

The W3C reached the same place. Its document on the inaccessibility of CAPTCHA states: “All interactive approaches require users to perform a task believed to be relatively easy for humans but difficult for robots. Unfortunately the very nature of the interactive task inherently excludes many people with disabilities, resulting in a denial of service to these users.”

It names blind users, low-vision users, deaf and hard-of-hearing users - audio alternatives are distorted to the point of unintelligibility - users with cognitive disabilities, and non-English speakers.

And you can just buy your way past

Motoyama and colleagues (2010), at USENIX Security, measured the CAPTCHA-solving service market directly, submitting tens of thousands of challenges to eight services over months.

Verbatim: “the customer cost to have 1,000 CAPTCHAs solved is now commonly $2/1,000 and can be as low as $1/1,000.” Accuracy across services ran 80.45% to 98.62%, with a median response of 14 seconds.

On the worker side: “in 2007, CAPTCHA solving routinely commanded wages as high as $10/1,000, but by mid-2008 a typical offer had sunk to $1.5/1,000, $1/1,000 by mid-2009, and today $0.75/1,000 is common.” The average monthly payout to the top 100 workers at one service was $106.31, falling to $47.32 a month later.

That data is from 2009–2010 and we are not presenting it as current pricing. We could locate no peer-reviewed measurement of CAPTCHA-farm prices since. Every current figure we found was vendor marketing.

But the structural point survives: a determined attacker pays a fraction of a cent. You pay ten seconds.

What this leaves you with

They are hard because they have to be. Once machines read distorted text at 99.8%, the only way to keep a puzzle machine-proof is to make it human-hostile too. The difficulty you experience is the residue of a lost arms race.

They are not primarily testing the puzzle any more. Risk scoring, cookies and browsing history do much of the work; the challenge is what happens when the score is ambiguous.

The cost is unevenly distributed. Blind users pay roughly five times the time for half the success rate.

Be sceptical of the enormous dollar figures, including the ones in the peer-reviewed paper. They rest on a 2010 FAQ, a wage that does not exist, and a mean 115 times its own median.

And a real gap worth knowing: we could find no study measuring whether CAPTCHAs actually reduce abuse. The literature measures human cost, machine solvability and solver prices. What happens to spam when you remove the CAPTCHA appears to be essentially unstudied.

For another interface that spends your attention while you wait, see the psychology of waiting. For consent interfaces that ask you to agree to things you cannot read, see do people read terms and conditions and should I accept cookies.

Limitations of this article

We could not verify the “100 million per day” figure that drives the widely quoted hours claim. It is sourced to an archived 2010 Google FAQ we could not open. That is why we describe the extrapolation rather than repeat its output.

The bot-versus-human comparison mixes sources. The human data is the authors’; the bot data comes from prior published attacks.

The 100% solve rate rests on roughly 50–100 runs, by the authors’ own account, and required a VPN, simulated mouse movement and cookies.

The CAPTCHA-farm pricing is sixteen years old.

We could not confirm exact page numbers for the accessibility study, and we report the W3C document as the Group Draft Note it currently is, dated 16 December 2021 - not as a finished Working Group Note.

Two disclosures. The 2023 bot-comparison paper includes authors at Microsoft and a US national laboratory - one a Google competitor. The 2008 Science paper was written by reCAPTCHA’s founders about their own system, twelve months before Google bought it. And the paper making the $888 billion claim carries no funding statement or competing-interests declaration at all, which is itself notable in a paper naming a company.

For the adjacent problem of repeated account-entry tasks, RoboForm provides password storage and form autofill; it does not solve CAPTCHAs.

Three claims are made for dark mode: it is easier on your eyes, it helps you sleep, and it saves your battery.

The first has no controlled evidence behind it, and the one crossover trial that tested it found no difference. The second tests the wrong variable. The third is true only at a brightness almost nobody uses.

What the research does show is the opposite of the folk belief, with a genuinely interesting exception at the end.

The measured finding runs the other way

The technical terms are positive polarity - dark text on a light background, what you would call light mode - and negative polarity, light text on dark.

Piepenbrock, Mayr, Mund and Buchner (2013), in Ergonomics, tested 169 people (85 older, mean age 69.8; 84 younger, mean age 22.6) in a two-by-two between-subjects design.

Visual acuity: polarity F(1,163) = 69.31, p < .01, η2 = 0.30. Younger adults t(82) = 9.93, p < .01, d = 2.17. Older adults t(81) = 2.53, p = .01, d = 0.58.

Proofreading: polarity F(1,165) = 9.92, p < .01, η2 = 0.06, with no interaction by age.

Positive polarity - light mode - won on both, for both age groups.

Piepenbrock, Mayr and Buchner (2014), in the same journal, found the mechanism with 35 participants in a within-subjects design:

A brighter screen constricts the pupil, which sharpens the retinal image. That is the whole effect, and it is why the finding goes the way it does.

The condition everyone omits

Dobres, Chahine and Reimer (2017), in Applied Ergonomics, tested 34 participants across ambient light levels - roughly 0 lux against 4,750 lux - in a repeated-measures design.

In the dark: F(1,33) = 49.60, p < .001. At 3 mm text, 122.3 ms to read under negative polarity against 84.1 ms under positive.

In bright light: F(1,33) = 0.19, p = .665. No polarity difference at all.

The light-mode advantage is conditional on dim ambient light and vanishes in bright conditions. If you are outdoors or in a well-lit room, the research has nothing to say about which mode you should use.

There is a further wrinkle that points at the real mechanism. A 2009 paper by Buchner, Mayr and Baumgartner is titled, in effect, that the positive-polarity advantage is due to high display luminance - that it is a brightness effect rather than a polarity effect as such. We could not read that paper and are reporting only its title claim, which the ambient-light interaction independently supports.

And a contrary result

While and Sarvghad (2024), in an IEEE VIS short paper, tested 104 participants (52 under 60, 52 over) on data-visualisation tasks - bar, line and scatterplots - rather than text.

Neither mode was consistently superior. For older adults, 48% performed better under positive polarity against 39% negative. For younger adults it reversed: 31% positive against 56% negative. The authors report that “the effects of contrast polarity on visual analysis performance do not noticeably change with age.”

They also found that stated preference frequently mismatched the empirically better mode.

So the light-mode advantage is a finding about reading text in dim light, not a general law of displays. Reading a scatterplot is not reading a paragraph.

The eye-strain claim

This is the claim in the search query, and it is the one with the least behind it.

Sengsoon and Intaruk (2025), in the International Journal of Environmental Research and Public Health, ran an experimental crossover with 30 participants, one hour per mode, split between gaming and reading.

Asthenopia scores rose substantially in both conditions: light mode 11.47 ± 1.68 to 18.37 ± 6.96 (p < .001); dark mode 12.40 ± 2.87 to 18.87 ± 7.01 (p < .001). Between modes: p > .05 - not significant.

Their stated conclusion: “No statistically significant difference in visual fatigue was observed between the two modes.”

Dry eye increased in both. An hour on a screen produced measurable eye strain either way.

We searched for controlled evidence that dark mode reduces digital eye strain and found none. Every confident assertion we encountered was on a commercial blog. That is the honest answer to the question this article is named after.

We should also say what we could not find: no systematic review or meta-analysis of display polarity exists that we could locate. For a question this widely asked, that is a real gap.

The sleep claim tests the wrong thing

Dark mode changes polarity. Night-shift modes change colour temperature. These are different variables and the popular coverage collapses them.

The relevant experiment is on colour temperature. Nagare, Plitnick and Figueiro, in Lighting Research & Technology, had 12 young adults view iPads from 23:00 to 01:00 across four nights, including Night Shift at two warmth settings.

All three lighting conditions significantly suppressed melatonin, and there was no significant difference between the two Night Shift settings. The authors’ conclusion: “changing screen color alone is insufficient for limiting the impact of PEDs on melatonin levels in the evening, and that screen brightness should also be reduced.”

Intensity dominates spectrum. And note what this does not test: we could find no study testing polarity on melatonin at matched luminance. The sleep claim for dark mode specifically is untested, not disproven.

The battery claim, with the number attached

Dash and Hu (2021), at MobiSys, profiled OLED display power across four Google and Motorola phones and six apps.

• At 30–50% brightness - typical real-world use - dark mode saves 3–9% of power on average.
• At 100% brightness, it saves 39–47%.

The dramatic figure is real and applies at a brightness setting most people do not use. At normal brightness the saving is below the threshold of noticing.

Any battery figure quoted without a brightness level is not traceable to the measurement.

Where dark mode genuinely wins

This is the exception, and it is the one nobody markets.

Legge (2016), in Visible Language, reports that for normally sighted readers “contrast polarity has little or no effect on reading speed, and only a weak effect on reading acuity.”

But for low vision: “a subset of people with low vision read 10% to 50% faster with bright letters on a black background.”

The mechanism is specific: “This contrast-polarity effect is often due to abnormal light scatter in the eye, which reduces the effective retinal-image contrast; for example, an eye with cataract” - because a white background scatters more light inside the eye than a dark one.

That is vision science, not medical advice, and we are reporting it as such. It is also the only place in this literature where light-on-dark has a large, mechanistically explained advantage.

Note that Legge’s characterisation of the normal-vision effect as “weak” sits in tension with Piepenbrock’s η2 = 0.30 on acuity. The defensible synthesis is that positive polarity reliably wins on acuity and fine-detail tasks in dim light, and that its advantage for ordinary reading speed in normal vision is small.

What this leaves you with

Pick whichever you prefer. For most people, in most lighting, the measured differences are small, conditional and sometimes reversed. Stated preference did not predict performance in the one study that checked.

If you are reading small text in a dim room, light mode has the better evidence. That is the specific condition where the effect is large.

In bright light, the research says nothing. The polarity effect disappears.

Do not choose dark mode for eye strain. No controlled study has found it helps, and the one crossover trial found both modes produced substantial strain after an hour.

Do not choose it for sleep. The variable that matters is brightness, and the closest experiment found a colour shift did not significantly reduce melatonin suppression.

Choose it for battery only if you run your screen at full brightness. Otherwise it is 3–9%.

If you have low vision, the calculus may be genuinely different - and that is a question for an eye-care professional, not for this article.

For a product sold against screen light whose trials came back null, see do blue light glasses work. For what evening screen use actually does, see why you can’t stop scrolling in bed at night. For a typography intervention that did not survive replication, see Sans Forgetica.

Limitations of this article

We could not read the 2009 paper that argues the polarity advantage is really a luminance effect. We report only its title claim, which the ambient-light interaction independently supports, and we print none of its statistics.

We read the 2014 pupil study in an in-press manuscript, which lacks final volume and page numbers.

We read the melatonin study in institutional summaries, not the paper. Its sample is 12 people and we have printed no statistics from it.

We read the battery figures in a university press release, not the conference paper. Note that the researchers’ institution has applied for a patent on the power-modelling technology used.

The eye-strain crossover is 30 participants, all female, in a narrow age band.

The visualisation study excluded users of accessibility tools, so it says nothing about that population.

We found no meta-analysis of display polarity and no study testing polarity against melatonin at matched luminance. Both are gaps, not findings.

One funding note: the ambient-light study was supported by a US Department of Transportation university transportation centre and a Toyota settlement research programme, with data collection facilitated by a commercial type foundry and a car manufacturer; one author was a type director at that foundry. Disclosed, and worth knowing.