PISA 2025 Results (Volume I)

Annex A1. Trends in reading scores over three PISA cycles (2018-2025)

Mean reading performance in PISA assessments declined significantly between 2018 and 2025 in a vast majority of countries and economies; a trend that is more often negative and steeper than what is observed in mathematics and science. For example, on average across 35 OECD countries with comparable data, 15-year-old students’ mean performance in reading dropped by 11 score points between 2018 and 2022, and by a further 14 points between 2022 and 2025 (25 points in total between 2018 and 2025). When considering this seven-year period and all PISA participants, about three out of four countries and economies (56 out of 74 with comparable data) show significant declines in mean reading performance, and only five countries/economies – all of them still performing significantly below the OECD average – improved (Table I.B1.2a.50). These sobering results warrant a closer investigation, to confirm that results reflect what students know and can do, and to identify possible causes underlying these trends.

This annex presents further evidence to corroborate the reality of the trends reported in Chapter 2 of this report, and to further dissect and contextualise these trends. It will start with a short summary of the main findings from Chapter 2. The focus on reading and on the 2018-2025 period allows to compare not only overall test scores, but also to examine changes in how students engaged with different reading tasks at a granular level.

Trends in reading scores from 2018-2025: A period of widespread decline

Between 2018 and 2025, mean reading scores declined by 25 score points on average across 35 OECD countries (Table I.B1.2a.37). Using the estimates of the yearly learning gain of students as a benchmark (Box I.2.4), this implies that 15-year-old students in 2025 were performing in reading at the level that could be expected of 14-year-olds just seven years earlier; in other words, their development of reading competence was lagging behind by one year, at a minimum.

Over a longer period, and although not all trajectories for individual countries follow this pattern, the average trajectory followed by OECD countries is one of accelerating decline, with small negative changes before 2018 (starting around 2012), followed by large drops between 2018 and 2022, and then again between 2022 and 2025 (Figure I.1.5).

When performance trends for different subjects are compared, countries’ and economies’ relative standing is remarkably consistent; even though reading trends are more often negative than mathematics or science trends. Countries and economies that had the most severe drops in reading also tended to experience larger drops in mathematics and science; in turn, countries and economies that improved in science often had non-significant changes (or less dramatic drops) in reading performance. The strength of the association (as measured by the linear correlation coefficient) between countries’/economies’ reading and science trends over the 2018-25 period is .79; and .82 between reading and mathematics trends (Tables I.B1.2a.36, I.B1.2a.37 and I.B1.2a.38). This suggests that common factors influence student learning in all subjects; the highly correlated trends also reflect the central place of reading in the other PISA assessment domains. In mathematics, for example, PISA does not assess only procedural knowledge, but emphasises students’ capacity to apply mathematical knowledge and reasoning to real‑world situations, which are presented to them in the form of written texts.

Differences in reading trends within countries and economies

In many countries, performance gaps between the highest-performing and the lowest-performing students widened between 2018 and 2025, as they did on average across OECD countries (Table I.B1.2a.43). However, the spread of performance increased only modestly in comparison to the changes in levels of performance: performance levels moved almost parallel to the average trend both at the top of the distribution (-21 points at the 90th percentile) and at the bottom (-28 points at the 10th percentile), on average across 35 OECD countries (Table I.B1.2a.40).

More surprisingly, perhaps, are the trends observed across different quarters of socio-economic status. Indeed, over the 2018-2025 period, the largest drops were observed among the most advantaged quarter of socio-economic status, and the socio-economic gap with the most disadvantaged students reduced somewhat (by 11 points, on average across 35 OECD countries: from 88 score points in 2018 to 77 points in 2025). Between 2018 and 2025, the socio-economic gap in reading performance narrowed in 36 countries and economies, and widened only in two (Croatia and Zambia) (Table I.B1.2b.23).

Taken together, these two findings suggest a new phenomenon wherein a group of students from relatively affluent families, but who struggle academically, are contributing most to the overall decline in mean scores.

The negative trends in reading, however, are so pervasive that no single group of students appears to be immune from them; and, at the same time, no single group of students can be claimed to be solely responsible for the decline. Understanding differences in the extent to which different groups of students were affected by the declines can facilitate the search for the root causes; this should not be an end in itself, in an attempt at assigning blame. More specifically, any hypothesis about the causes – whether global or national – must be tested against the observed patterns to see whether it plausibly explains them or conflicts with how those causes would be expected to affect different groups.

With these notes of caution in mind, the next paragraphs describe differences across genders and across students with different immigration backgrounds.

Trends were similar for both boys and girls, on average across 35 OECD countries, and when considering the full 2018-2025 period (Table I.B1.2c.28).

The gap between immigrant students and non-immigrant students also did not change significantly – suggesting parallel trends for both immigrant and non-immigrant students – in the vast majority of countries with a substantial population of students with an immigrant background (more than 5% of students in each cycle, starting with 2018; see Table I.B1.2d.1), and where the overall trend in mean reading performance is negative (Table I.B1.2d.8). This pattern is observed on average across 29 OECD countries with significant proportions of immigrant students; and, individually, in 22 of the 31 countries/economies with available data that meet the two conditions (high immigrant population size and negative overall trends).

A clear pattern of widening gaps between immigrant and non-immigrant students, driven by more negative trends among immigrants (over the 2018-2025 period), is only observed in seven countries and economies with an overall negative trend in reading: Argentina, Croatia, Israel, Latvia, Lebanon, North Macedonia and Serbia. At the same time, even in the remaining countries/economies, the overall trend is often more pronounced than the trend among immigrants or among non-immigrants, considered separately. In some cases, such as in the Netherlands*, the overall trend is significantly negative, but trends for each sub-population separately are not flagged as significant (Table I.B1.2d.8). This is because, in addition to the within-group trends, the overall trend is also influenced by the changing composition of the population. In many OECD countries, the proportion of students with an immigrant background increased significantly between 2018 and 2025, with a marked increase in the number of first-generation immigrant students whose families arrived in the country after their birth and, often, after they had already started school in another country (Table I.B1.2d.1). When this is observed, it can be concluded that the overall trend would have been somewhat more positive, had the proportions of immigrant and non-immigrant students stayed constant.

In particular, among countries where immigrant students scored significantly below non‑immigrants in PISA 2025, and where they represented at least 5% of the total population throughout the 2018-2025 period, the proportion of students with an immigrant background increased by more than 3 percentage points in Portugal, Germany, Norway*, Ireland, Slovenia, Iceland, Austria, Switzerland and France (in descending order of that increase) (Tables I.B1.2d.1 and I.B1.2d.8). While the performance gap with non‑immigrant students remained stable in many of these countries, the size of this gap, combined with the growing share of students from the lower-performing group, can account for a small portion of the negative trend in reading performance. In other words, the trend would have been less pronounced, had there not been an increase in the proportion of students with an immigrant background in these countries/economies.

The change in mean scores that can be explained merely by the changing composition of the student body, however, never exceeds a few points. Such demographic changes are never the main driving force behind the large trends observed. For example, on average across 29 OECD countries, the proportion of students with an immigrant background increased by 3.8 percentage points between 2018 and 2025 (Table I.B1.2d.1). During this period, the gap between immigrant students and non-immigrant students remained around 40 score points, and reading performance declined by 25 score points (Table I.B1.2d.8). A simple multiplication suggests that out of these 25 points, about 1.5 points (0.038 × 40) can be attributed merely to the changing composition of the student body. A similar calculation for Portugal, for example, would suggest that about 3 points (0.113 × 32) out of 30, can be explained by the growing portion of immigrant children in the student population.

Are PISA trends consistent with trends observed in other international assessments of reading?

PISA is the only international assessment of reading proficiency of students around the age of 15. Many countries and economies that participated in PISA 2025, however, also participate in the International Association for the Evaluation of Educational Achievement's Progress in International Reading Literacy Study (PIRLS), which assesses primary-school students towards the end of Grade 4. PIRLS is repeated every five years, and its most recently published results are based on data collected in 2016 and 2021 (Mullis et al., 2023[1]).

Although there are significant differences in how the two assessments define their target populations (as a grade-based population in PIRLS, and as an age-based population in PISA) and what they assess in reading (reflecting the younger age of students assessed in PIRLS),1 an approximate correspondence can be established between the cohorts that participated in PIRLS and those that participated in PISA. For example, students who were at the end of Grade 4 in 2016 mostly turned 10 in that year; and would therefore become eligible for participation in PISA some six years later, in 2022.

PISA 2025 mean scores for countries and economies are highly correlated with PIRLS 2021 mean scores across 27 participants who took part in both assessments (r = 0.74). The largest deviation in relative results is observed in Bulgaria, which scored much lower in PISA than what could be expected based on its students’ results in PIRLS.

Differences between PIRLS results in 2016 and 2021 often tend to anticipate differences that are observed, in older students, between 2022 and 2025 in PISA (r = 0.34, n = 23). Even though the correlation among trends is much weaker than for mean scores, most of the countries/economies that observed negative trends in PIRLS between 2016 and 2021 also observed negative trends in PISA between 2022 and 2025; and on average across 23 countries/economies that participated in all four assessments, results declined both in PIRLS and in PISA. A few countries/economies saw declining results in PISA in spite of stable (Australia, France and Hong Kong (China)) or even improving (Singapore) results in PIRLS. Still, for the majority of countries and economies, the most recent results in PISA were foreshadowed by the results of similar cohorts at a younger age, when they were assessed in PIRLS. This suggests that the root cause of the recent declines observed in PISA could, in these cases, be plausibly found in the earlier childhood years of today’s 15-year-olds, and do not necessarily reflect post-2022 phenomena.

Where did students’ performance decline in the PISA test?

The PISA 2025 reading test is based on the PISA 2018 reading test, whose development was guided by the framework revision undertaken between 2015 and 2017. In particular, the reading tasks that students had to solve in 2025 were already used in the two previous assessments (if not for longer), and the overall assessment experience of students was very similar to those of 2018 and 2022. This offers an opportunity to examine changes in student performance on a large number of indicators (i.e. through multiple numerical measures, each providing a somewhat different insight), and in a highly controlled setting, not confounded by differences in test design and task content.

To dissect the overall reading trend into smaller components and identify the driving forces behind each country or economy’s trend, these indicators are derived first at the level of individual items, then combined over items that share common characteristics, in order to identify patterns. The main item-level indicator of performance is the weighted percent-correct score, which corresponds to the estimated percentage of the 15-year-old population that successfully answered the item. Several indicators are derived from this through simple functional transformations (in particular, success odds and their logarithm). Because of the adaptive nature of the test, harder tasks were more frequently given to higher-performing students; and conversely, easier tasks were more frequently given to lower-performing students. Annex A3 explains how percent-correct scores for a representative population of 15-year-old students can be computed from the PISA reading test for all items, in spite of its adaptive nature.

Average success rates (percent-correct scores, hereafter, P+) on reading items are consistent with the overall trends reflected in mean scores on the PISA reading scale. In countries with strong negative trends for mean scores, success rates declined the most; in countries with non-significant or even positive trends in mean scores, they declined the least or even increased. The correlation between differences in average P+ rates and differences in mean scores is 0.91 for 2022-2025 comparisons; and 0.90 for 2018‑2025 comparisons (Table I.A1.7).

Student performance throughout the reading test

Average success rates can be similarly computed over smaller sets of items. In particular, it is possible to compute, and compare, success rates over items that were administered in the first, non‑adaptive section of the reading test separately from items that were administered in the adaptive sections (Stage 1 and Stage 2) of the reading test (Table I.A1.7). Each of these subsets shows a pattern of change in success rates, over the years, that is broadly consistent with the overall trend in mean performance. This implies that a shorter reading test would likely have led to similar results: results did not decline because students disengaged with the test after a while, but because students performed below their peers from three or seven years earlier from the very beginning. Nor did the adaptive nature of the test – whereby good students progressively encounter harder questions – influence the direction of trends. Indeed, the vast majority of declines are clearly visible already in the first, non-adaptive section of the test.

Student performance across different types of reading tasks

Declining average success rates imply that, on average, items became more difficult for the student population assessed in PISA. By tracking success rates on single items over recent cycles, it is possible to identify common characteristics of the items whose difficulty for students increased more than average (students’ ability to solve the tasks successfully declined the most); and of those whose difficulty changed the least. In this section, these patterns are presented through odds ratios, and using the 2018 odds of successfully responding to an item as the baseline. An odds ratio above 1 indicates that the likelihood of success increased; conversely, an odds ratio below 1 indicates that the likelihood of success decreased.

In particular, items were grouped along the following text and task characteristics:

• Position in the test: Core (items from units that were administered in the initial, non-adaptive section of the test); Stage 1; Stage 2.

• Position within the unit: First item in the unit (in which new texts are introduced for reading); later items in the same unit, asking questions about the same source texts already introduced in the first item.

• Text length: Items based on shorter texts (fewer than 200 words in the English source version); on medium-length texts (between 200 and 399 words); or on longer texts (400 words or more).

• Response format: Open response (human- or computer-scored); simple multiple choice; complex selection (e.g. drag-and-drop).

• Text source: Single (the answer requires reading text from a single source); multiple (the answer requires reading text from multiple sources).

• Cognitive process: Locating information (scanning a single source to locate the answer to the question, or searching, across several sources, for the desired information); understanding (including both literal comprehension and connecting information to establish an overall meaning); evaluating and reflecting (involving higher-level processes where readers must go beyond the literal or inferred meaning of the text to assess the quality and validity of text content and form).

Some of these dimensions seem to be associated with meaningfully different trends over time. In other words, some aspects of the texts and tasks included in the PISA reading test are associated with even stronger-than-average declines in students’ ability to successfully engage with them, on average across OECD countries, while others appeared less affected by the decline.

Clear differences in relative success rates between 2018 and 2022 emerge in particular by text length, response format and cognitive process; smaller differences are also observed by cognitive process (Table I.A1.8):

• Performance dropped more strongly on items from units that used longer texts as stimuli, on average across OECD countries.

• Performance also dropped more strongly on items that used a simple multiple-choice response format (typically, in PISA, with four response options, of which one is correct).

• Finally, not all processes appear to be equally affected by the decline in reading performance between 2018 and 2022. When considering the average changes observed across OECD countries, least-affected were items that demanded from students to locate information, which students could successfully solve even with only a superficial, yet targeted, engagement with the text. Conversely, the most-affected items relate to higher-level processes (evaluating and reflecting), which require careful reading, and often required students to remember and integrate multiple pieces of information, either from a single source or from multiple sources.

Over the longer period between 2018 and 2025, differences by response format were even stronger, as students were somewhat more successful in 2025 on open-response items than three years earlier, but performance continued to decline on the remaining items (Table I.A1.10); and differences by text length continued to be visible, even though they did not widen. In contrast, differences by the targeted cognitive process narrowed somewhat: between 2018 and 2025, performance declined similarly, irrespective of the targeted process (Table I.A1.9).

Finally, the small differences observed between 2018 and 2022 by position of the unit within the test were compounded by similarly small differences in the next period, to produce a more significant pattern when considering the longer period (2018-2025) (Table I.A1.9). On average across OECD countries, performance dropped throughout the test, but it dropped more severely in the final adaptive stages than on items that were presented to students at the beginning of the test. In other words, performance declined over the years and, in addition, performance declined during the test, in relative terms, to produce larger declines in the latter sections of the test compared to the declines observed in earlier sections.

Trends are more parallel on other dimensions; in particular, PISA data do not indicate, on average across countries, greater difficulties (or reluctance) of students to respond to items that require searching or integrating meaning and information across multiple texts (sources); compared to items that relied on a single source of text. Nor do they indicate large differences in success rates depending on the position of the item within the unit, as the first item, or as a further item based on the same text(s).

Results for individual countries and economies may sometimes differ from the average patterns across OECD countries presented above and are, in general, affected by wider margins of uncertainty than average results. They are available in the underlying tables (Tables I.A1.8, I.A1.9, I.A1.10).

What kind of mistakes became more common between 2018 and 2025?

If success rates on PISA reading tasks declined, incorrect (or partially correct) responses must have increased. But what kind of mistakes became more common?

In this section, the analysis focuses on two types of incorrect responses that can be identified on most, if not all, items: skipped responses (item responses left blank, in-between non-missing responses); and hasty responses (item responses that were not scored as correct, and on which students spent less time than the fastest, yet accurate respondents). More precisely, a “minimum threshold” of time effort required for each item was defined as the 1st percentile of the (unweighted) distribution of response times among all respondents who solved the item correctly, across all participating countries and economies.2 This means that only 1% of all full-credit respondents were able to respond to the item in less time, and 99% of full-credit respondents spent longer than this threshold. The minimum threshold of time required was increased to 10 seconds in all cases where, mostly because of correct guess answers, more than 1% of full-credit respondents had provided the correct response in less than 10 seconds. Respondents who did not obtain a full-credit score on the item, but who spent less time than this threshold, are considered to have given a hasty, speeded response (as defined, a “skipped” response may also be a “hasty” response).

Table I.A1.12 shows how the proportion of skipped and speeded responses evolved between 2018 and 2025. A clear trend emerges: on average across OECD countries, the average proportion of hasty (incorrect and fast) responses increased significantly between 2018 and 2025. The average proportion of hasty responses increased by over 4 percentage points over the period, reaching a total of about 9%. The average increase across items of hasty responses exceeded 10 percentage points in three countries/economies (Kosovo, Dominican Republic and Bulgaria, in descending order of that increase). Similar analyses for the mathematics and science test are presented in Tables I.A1.13 and I.A1.11; analyses for math, in particular, also show an increase in the proportion of hasty responses between 2022 and 2025, with the largest increases observed in Serbia, Malta and Israel.

In the PISA test, sequences of items that are left blank at the end of the test are not considered when scoring students’ answers. Similarly, in these analyses, the corresponding students and (missing) answers are excluded from the percentages of “skipped” or “hasty” respondents. Indeed, in PISA, these items are considered as if they were not administered to students, who, most likely, did not reach them. Such non-reached items are also ignored when estimating item difficulty and student performance. However, because, over time, students became more likely to skip or to provide a fast (if incorrect) response rather than to continue working on a task that they perceived as difficult, it is interesting to note that the proportion of non-reached items declined in reading, by about 2 percentage points (Table I.A1.12), on average across OECD countries.

Reading-fluency trends

Just like in 2018 and 2022, the PISA 2025 reading test started with a three-minute-long assessment of students’ reading fluency. Students were asked to evaluate the sensibility of 21 or 22 sentences (i.e. two out of six possible sentence blocks). The same question was asked about each sentence: “Does the sentence make sense? Yes/No”. Each set of sentences included both sensible sentences (e.g. “The student read the book last night”) and nonsensical sentences (e.g. “Airplanes are made of dogs”). The instructions given to students, including a short practice test, can be found online (https://doi.org/10.1787/098bab1a-en).

Students who completed the test quicker – while still taking the time required to make sound, sensible judgements – are more fluent readers. When comparing results over time, longer response times (for a given level of accuracy) would suggest declining levels of reading fluency; while lower levels of accuracy in the sensibility judgements could indicate a lack of foundational skills in reading, but also a lack of attention and engagement.

A simple comparison of the percentage of correct responses across years immediately shows, in most countries and economies, a declining proportion of students who made accurate assessments of the proposed sentences. In most countries and economies, the proportion of students who correctly classified each sentence as either “sensible” or “not-sensible” declined, and the direction of trends in percent-correct scores on reading-fluency items mirrors the direction observed for success rates on the remaining, main section of the reading test (Table I.A1.7). Nevertheless, a few countries stand in contrast to this trend; in particular, in Georgia, Japan, Kazakhstan, Montenegro, Chinese Taipei and the United Arab Emirates, success rates on reading-fluency items declined between 2018 and 2025, but higher success rates were observed in the longer reading comprehension tasks (in all these countries/economies, the overall trend in reading performance is reported as not significantly different from 0; Table I.B1.2a.37).

A more nuanced understanding of this trend requires considering, for each of the six blocks of reading sentences that were presented to students, both the number of correct sensibility judgements and the time taken to form these judgements.3 It is then possible to classify students who took the PISA fluency test into four categories:

• A first distinction is between those who engaged with the fluency test as expected and those who, instead, tried to rapidly guess the correct answers, before they could even read the full text. The latter, “hasty” test-takers, made more than one mistake in their sensibility judgements, and completed the assessment faster than 99% of those test-takers in their country and language group who made one mistake at most. Their performance is hard to interpret, as it might reflect a lack of foundational skills in reading, but also a lack of attention and engagement.4

• A second distinction is among the remaining, and more engaged test-takers. Among these, language-specific norms have been derived for response time on reading-fluency tasks to distinguish between “fluent” readers, “slow” readers and “inaccurate” readers. The norms correspond to the pace at which at least 50% of Level-3 (which corresponds to the average proficiency of 15-year-old students across OECD countries) students read, excluding those few Level‑3 students who made more than one mistake in their sensibility judgements. Students who read at this level can take many features into account when comparing, contrasting or categorising information, even if the required information is not prominent or there may be a considerable amount of competing information:

  • “Accurate” readers made, at most, one mistake in their sensibility judgements and took, at most, twice as long as the norm defined for their language group.

  • “Slow” readers are those who took more than twice as long to complete the reading-fluency test as the norm defined for their country and language group (regardless of whether the answers were accurate or not).

  • “Inaccurate” test-takers made, just like “speeded” readers, more than one mistake, but completed the test neither too fast nor too slowly.

Table I.A1.15 shows that, between 2018 and 2025, the proportion of accurate (and fluent) readers declined by 7 percentage points across 35 OECD countries on average; meanwhile, the proportion of “slow” increased only by about 1 percentage point. Concurrently, the proportion of “hasty” readers increased by about 5 percentage points and almost doubled between 2018 (6.6%) and 2025 (11.4%). Most of this increase was over the most recent period (2022-2025).

However, not all countries and economies experienced similar trends. While the proportion of accurate, fluent readers declined in most countries and economies between 2018 and 2025, and this decline often coincided with an increase in the proportion of speeded test-takers, there are exceptions. In Morocco, the proportion of “slow” readers increased the most, possibly reflecting the expansion of secondary education in the country, to students who would have previously dropped out of school. In Kazakhstan, the proportion of “inaccurate” readers increased, but the proportion of “speeded” test-takers did not change significantly. And in a few countries – most notably, in the Philippines and Brunei Darussalam – the proportion of fluent readers increased between 2018 and 2025. In these two countries, the increase in the proportion of fluent readers is just one of several positive trends found in PISA.

Can differences in student motivation and test-context effects explain the observed reading trends over the 2018-2025 period?

Performance on the PISA tests reflects what students know and can do; but also how motivated they are to do well on the test, on the day of testing.

For most students, PISA is a low-stakes assessment: they can refuse to participate in the test with no negative consequences and do not receive any feedback on their performance. There is a risk, therefore, that students do not do their best on the test; and that they behave differently from what they would do in other circumstances.

How did students’ motivation to do well on the test change over time? This section summarises several indicators of student engagement using PISA 2025 data, and compares them, where appropriate, to PISA 2018 and 2022 data.

A number of approaches have been developed to assess differences in students’ motivation in low-stakes tests (Buchholz, Cignetti and Piacentini, 2022[2]) between individuals or groups (e.g. across countries and economies). Some approaches rely on test-takers’ own perceptions and reports about their effort and dispositions; others rely on the observation of student behaviour during the test. Among the latter, one can further distinguish between invasive approaches, which require dedicated resources such as human proctors, eye-tracking devices or the administration of bespoke test modules, and non-invasive approaches, which rely only on students’ interactions with the test and questionnaire forms. This section uses self-reports and non-invasive behavioural indicators.

Self-reported effort invested in the PISA test and student motivation for learning

In most countries/economies, students reported making less effort on the test in 2025 than in 2022 and in 2018: on average across 35 OECD countries, the difference corresponds to -0.5 points on the 10‑point scale (compared to 2018; and -0.3 points compared to 2022) (Table I.A1.1). Meanwhile, the proportion of students who indicated that they invested very little effort in the PISA test increased from 3.5% to 5.4%. Reports about the effort students would have made had the test counted towards their grades were also lower (by 0.2 points on average across 35 OECD countries) but the decline was more marked for reports about the actual effort students made.

There is a moderate positive association between country/economy-level changes in effort reported by students and reading trends; but also between the effort students reported they would have made on a regular school test and reading trends. This seems to imply that changes in student dispositions towards PISA could reflect wider changes in student dispositions towards assessments of learning, and perhaps towards learning more generally. Moreover, much of these associations appear driven by Georgia and the Philippines, among the few countries where self-report effort improved between 2018 and 2025 (as did reading performance, in the Philippines; in Georgia, the observed change was positive, but not significant). In contrast, among countries and economies where average self-report effort declined by more than 0.5 points on the 10-point scale, the direction and magnitude of performance trends vary significantly, from no decline at all (e.g. in Italy and Chinese Taipei) to very large drops (e.g. in Iceland and Norway*).

Still, the above analyses of reading trends have shown that the declining scores also, and perhaps primarily, reflect a decline in focus and attention; in particular, the kind of focus and attention that is a constituent part of the reading construct in PISA. Interestingly, the questionnaire answers of the students who took the PISA test often indicate a negative shift in general dispositions towards schooling and towards effortful learning, which may have played a role in the trends observed. In particular, between 2022 and 2025, changes in the proportion of students who agreed, or strongly agreed, with the statement “I am curious about many different things” declined, and more so, on average, in countries and economies where reading results declined the most. Meanwhile, changes in the share of students who disagreed, or strongly disagreed, with the statement “School has been a waste of time” also correlate with the direction and strength of reading trends: even though only few countries and economies saw an increase in this share, relatively large increases were recorded in countries, such as Germany, Latvia and the Netherlands*, where mean reading scores were, in 2025, much below the corresponding scores in 2022 (see Chapter 3; Tables I.B1.3.2 and I.B1.3.71).

Such correlations do not constitute proof of causality, nor are they a reliable guide for assigning priorities among competing hypotheses; still, they can suggest a wider context for the performance trends, one that is equally in need of interpretation and explanation.

PISA data can document shifts in attitudes towards schooling and learning among 15-year-olds, but cannot show the extent to which these shifts reflect differences across successive cohorts or, rather, broader time trends that affect all cohorts (including, for example, the current 15-year-olds’ parents) to a similar extent.

Successive PISA datasets provide a series of snapshots of student learning at age 15; but children’s performance on tests is influenced by the cumulative effect of the school and family environment over all previous years. For example, many countries have witnessed, after the first COVID pandemic (i.e. starting around 2020), persistently high levels of student absences at all levels of education (OECD, 2026[3]). International data on fourth-grade students, i.e. at an age where student absences are more likely to reflect parental attitudes than student attitudes, indicate that the proportion of students who are absent from school at least once every other week rose from 10.7% in 2019 to 14.8% in 2023, on average across OECD countries (OECD, 2026, pp. 29-30, based on TIMMS data[3]). Even if student absences were, by 2025, slowly returning to lower levels from a peak level during the pandemic years, from the point of view of the students who were assessed in 2025 the lower levels of school attendance experienced over five years likely had a bigger impact than the level of absenteeism at age 15, which, as shown also by PISA data, tends to be lower than in 2022 (Table I.B1.4.161).

Behavioural measures of test engagement and fatigue

Some behavioural measures of test engagement, which allow for trend comparisons and can therefore help contextualise the observed reading trends, can be derived from other domains assessed in PISA, and even from the questionnaire.

Given the earlier findings that indicated somewhat stronger declines in the latter part of the reading test, compared to the initial section, one may wonder whether this reflects more generally a drop in the capacity to sustain effort; and whether a two-hour computer-based test may not exceed the endurance of 15‑year‑olds. To what extent are the negative trends in reading reflective of shorter concentration spans?

Differences in performance by hour of testing and domain order

A first angle to investigate student fatigue consists in comparing the performance of students who sat the reading test in the first hour and in the second hour of testing. A similar comparison can be conducted for all subjects; and if the differences in performance are indicative of the students’ capacity to sustain a cognitive effort, in general, results should be relatively consistent regardless of the subject. Indeed, by virtue of the random allocation of test booklets to students, the groups of students that are being compared in each case are strictly equivalent.

Tables I.A1.3, I.A1.4 and I.A1.5 present these comparisons for science, mathematics and reading. Performance differences between the first and second hour are often negative, and sometimes relatively large (more than 10 score points in reading, on average across 35 OECD countries, regardless of the year). However, the between-country variation in these differences is not strongly correlated across subjects, and the differences themselves do not seem to have widened over time: more often than not, the negative trend in reading between 2018 and 2025 was even stronger for students who took reading as the first subject (Table I.A1.4). Only a few countries/economies stand out for showing a more negative reading trend for students who took the reading test in the second hour (Belgium, Kosovo and Lithuania) than for students who took the reading test in the first hour.

An alternative interpretation of these differences emphasises the role of interactions across subjects. In this interpretation, performance is not only influenced by a domain’s position in the test, but by the precise nature of what students did in the hour before. For example, doing a test that is perceived as difficult, by most students, in the first hour (e.g. mathematics) may have a negative effect on students’ reading performance in the second hour; while doing an easier test in the first hour may have a positive effect. On average across OECD countries domain-order effects are not pronounced (Table I.A1.6); but this average does not necessarily reflect the absence of domain-order effect: it may result from the large variety of forms that domain-order effects take across countries and economies. For example, in Chinese Taipei, students who took the reading test after a mathematics test achieved higher scores than students who took it after the assessment of “learning in a digital world”; while a difference in the opposite direction is observed for students in Finland.

Differences in performance between the beginning and the end of the science test

A second angle to investigate the role of fatigue, and how it may have impacted performance trends (and reading trends in particular), exploits the science test, and in particular its linear version, which was used for 25% of test-takers. The item bank in science was divided in three item sets (A, B and C), from which sequences of about 11 to 13 items were extracted (these sequences are referred to as “testlets”). Each linear test consisted of three such testlets, and the order in which these testlets were administered was rotated across students, such that, out of these 25% of student, one third started with item set A, one third with item set B, and one third with item set C; and similarly, for the second and third testlets. The change in the percentage of correct responses over the three possible positions of an item set in the linear test therefore provides an indirect measure of test fatigue.

Table I.A1.2 shows that in most countries, results declined modestly over the course of the science test; only Korea and Kenya exhibit the opposite pattern of better results on the same items when they are located at the end of the test, compared to the beginning. The largest declines, with differences in P+ exceeding 4 percentage points, were observed in Bulgaria, Israel and Thailand (in descending order of the magnitude of the decline).

Overall, student fatigue during the PISA test, in general, does not appear to play a significant role in the observed trends; countries where fatigue effects are more pronounced in 2025 did show similar declines, compared to 3 or 7 years earlier, as countries in which fatigue effects are less pronounced.

Patterned response behaviour (straightlining)

A third indicator of fatigue, or perhaps more generally of disengagement in the PISA test, is based on mixed scales in the student questionnaires. Mixed scales are item batteries that measure a single construct using both positively worded and negatively worded items. An example – and the only mixed scale that can be compared over three cycles of PISA, from PISA 2018 to PISA 2025 – is the sense-of-belonging scale, which contains apparently contradictory statements such as “I feel lonely at school” and “I make friends easily at school”. In both cases, students are asked to indicate the extent to which they agree with the sentence provided.

Students who do not pay attention to the sentence, or to the response scale, may sometimes respond with the same response option (e.g. “strongly agree”) regardless of the item; thus contradicting themselves half of the time (the sense-of-belonging scale contains exactly three positively worded items and three negatively worded items). Students are said to “straightline” because their responses to these items are all perfectly aligned, on the response sheet – and most likely misaligned with how they truly feel. Straightlining is not the only possible pattern of disengagement, but is perhaps the easiest to detect.

On average across 23 OECD countries, the percentage of students who “straightlined” in the sense-of-belonging item battery was somewhat lower in 2025 than in both 2022 and 2018. This suggests that students were not less engaged with the PISA test and questionnaire than in previous years. In fact, there are only limited exceptions where straightlining increased significantly, and only when considering the most extreme version of it (when students always select the “strongly agree” or “strongly disagree” option), which remains very rare. The largest increase, from 0.9% to 2.2% (+1.3 percentage points), was observed in Latvia.

Conclusion

Declining trends of reading performance are a robust, if sobering, finding from the PISA 2025 data. They are unlikely to be a measurement artefact, an aberrant observation that emerged only in the specific context of the PISA assessment; the signal is strong, stronger than any caveat that may result from the limitations of current educational measurement tools.

At the same time, a closer look at the trend assigns a leading role, for the declining success rates, to the declining capacity of students to engage with reading tasks which require careful reading and sustained attention. Over time, students became more and more likely to provide a hasty, superficial and incorrect response, and less likely to spend the time and effort required by a correct response. If test engagement appears, therefore, to play a role in the observed declines, it is because the capacity to direct and, often, sustain attention for a relatively long period is part of what the “reading” test in PISA assesses.

Comparisons of PISA 2022 and 2018 response patterns, moreover, suggest that students struggled more in 2022, in a majority of countries/economies and compared to some four years earlier, when asked to respond to questions about longer texts, or when asked to hold in their memory multiple pieces of information and to combine them, e.g. in order to assess the quality or trustworthiness of a written text; whereas performance held more steady when they were asked to find a single piece of information in the text. Over the longer period (2018 to 2025), differences by cognitive process were less pronounced, and declines were more uniform across all types of tasks; in contrast, difficulties related to text length remained a strong driver of the overall decline.

Earlier analyses of PISA data provide interesting pointers to how schools, teachers and parents can help create opportunities for students to acquire strong reading skills. In particular, the 21st‑Century Readers (OECD, 2021[4]) report pointed to the importance, for teachers, of explicitly teaching navigation strategies and, for students and their parents, to the continued relevance of print reading as a way of building up the “muscles” and agility required in the largely disintermediated and flooded information landscape of the early 21st Century. There is no doubt that PISA will revisit these questions, and support similar analyses, as it engages in a revision of the framework for assessing reading and as reading returns as a focal domain in PISA 2029.

Whether the declining results in reading reflect a less central role of reading skills in an age where large language models can instantly summarise long documents, and where audio and video streams often replace written text, cannot be determined by PISA data. The debate about the consequences of falling reading scores transcends the power of PISA data, and must be informed by values and worldviews too. Nevertheless, the importance of reading in the PISA test, for the moment, reflects a consensus among the countries that are represented in its Governing Board and among the experts who contribute to the assessment frameworks5: even if reading practice is declining at work, in civic contexts or for leisure, and may continue to decline in the future, reading competence remains a required foundation for developing a wide range of human capacities that contribute to individual and societal success – from reasoning, to empathy, to critical thinking – and therefore remains a central goal for education.

Table I.A1.1. Trends in reading scores over three PISA cycles (2018-2025) – Tables for Annex A1

Table I.A1.1	Effort invested in the PISA test, 2018 through 2025
Table I.A1.2	Endurance in the PISA 2025 science test
Table I.A1.3	Performance differences by testing hour in science, 2018 through 2025
Table I.A1.4	Performance differences by testing hour in reading, 2018 through 2025
Table I.A1.5	Performance differences by testing hour in mathematics, 2018 through 2025
Table I.A1.6	Performance differences in the second hour, as a function of the test taken in the first hour
Table I.A1.7	Average percentage of correct responses to trend items in reading, 2018 through 2025
Table I.A1.8	Change in the likelihood of correct responses in 2022, compared to 2018, by item characteristics
Table I.A1.9	Change in the likelihood of correct responses in 2025, compared to 2018, by item characteristics
Table I.A1.10	Change in the likelihood of correct responses in 2025, compared to 2022, by item characteristics
Table I.A1.11	Types of missing and incorrect responses to science items (PISA 2025)
Table I.A1.12	Types of missing and incorrect responses to trend items in reading, 2018 through 2025
Table I.A1.13	Types of missing and incorrect responses to trend items in mathematics, 2022 through 2025
Table I.A1.14	Response patterns and time in the PISA reading-fluency test, 2018 through 2025
Table I.A1.15	Reading fluency, 2018 through 2025
Table I.A1.16	Reading fluency, by proficiency levels in reading
Table I.A1.17	Straight-lining on sense-of-belonging questions, 2018 through 2025

Statlink

References

[2] Buchholz, J., M. Cignetti and M. Piacentini (2022), “Developing measures of engagement in PISA”, OECD Education Working Papers, No. 279, OECD Publishing, Paris, https://doi.org/10.1787/2d9a73ca-en.

[5] Mullis, I. and M. Martin (eds.) (2019), PIRLS 2021 Assessment Frameworks, TIMSS & PIRLS International Study Center, Lynch School of Education, Boston College and International Association for the Evaluation of Educational Achievement (IEA), https://pirls2021.org/wp-content/uploads/sites/2/2019/04/P21_Frameworks.pdf (accessed on 26 June 2026).

[1] Mullis, I. et al. (2023), PIRLS 2021 International Results in Reading, TIMSS & PIRLS International Study Center, https://doi.org/10.6017/lse.tpisc.tr2103.kb5342.

[3] OECD (2026), Every Day Counts: Understanding, Preventing and Responding to School Attendance Problems, OECD Publishing, Paris, https://doi.org/10.1787/7c6f6c3e-en.

[4] OECD (2021), 21st-Century Readers: Developing Literacy Skills in a Digital World, PISA, OECD Publishing, Paris, https://doi.org/10.1787/a83d84cb-en.

[6] OECD (2019), PISA 2018 Assessment and Analytical Framework, PISA, OECD Publishing, Paris, https://doi.org/10.1787/b25efab8-en.

Notes

← 1.  PIRLS defines reading literacy as “the ability to understand and use those written language forms required by society and/or valued by the individual” (Mullis and Martin, 2019[5]). PISA similarly defines reading literacy as “understanding, using, evaluating, reflecting on and engaging with texts in order to achieve one’s goals, to develop one’s knowledge and potential and to participate in society” (OECD, 2019[6]). While conceptually related, the actual assessments are tailored to the specific target age group (Grade 4 for PIRLS, 15-year-olds for PISA). For example, reading passages used in PIRLS are “typical of those read by [Grade 4] students in their everyday experiences”; PISA, in contrast, also includes some tasks based on work-related texts, because its respondents are older and because it aims to assess their ability to meet real-life challenges that they will encounter after leaving compulsory schooling.

← 2.  Due to minor differences in screen layout, and to a change in the test-taking platform in 2025, which might have affected the measurement of time on task, the “minimum time threshold” is defined separately for each year.

← 3.  Due to minor differences in screen layout and to a change in the test-taking platform in 2025, which might have affected the measurement of time on task, timing data are converted into categories of “fast”, “regular” and “slow”, based on theresholds defined separately by country and year.

← 4.  The operational definition of “hasty test-takers” in reading fluency is similar to how hasty responses are defined on the main reading tasks. However, to account for the greater influence of linguistic aspects (script and orthography, text length, etc.) on the time effort required for reading fluency tasks, the thresholds are set separately by country and (major) language. This is possible also thanks to the larger number of accurate responses on which such thresholds can be computed for reading-fluency tasks, which count among the easiest in the reading test.

← 5.  At the time of writing, the most recent draft version of the PISA 2029 Reading Framework is available online at EDU/PISA/GB(2026)6 (accessed on 12 June 2026).