ordinary knowledge ascriptions has been divided along methodological lines: "evidence-fixed" prompts rarely find stakes effects, while "evidence-seeking" prompts consistently find them. We present a cross-cultural study using both evidence-fixed and evidence-seeking prompts with a diverse sample of 17 populations in 11 countries, speaking 14 languages. Our study is the first to use an evidence-seeking prompt cross-culturally, and includes several previously untested populations (including indigenous populations). Across cultures, we do not find evidence of a stakes effect with our evidencefixed prompt, but do with our evidence-seeking prompt.

We argue that the divergent results reveal a tension within folk epistemology: people's beliefs about when it is appropriate to ascribe knowledge differ significantly from their actual practice in ascribing knowledge.

Is knowledge sensitive to practical considerations in addition to evidential considerations? Many philosophers believe that it is. 1 In particular, many have been convinced by bank cases that knowledge ascriptions are sensitive to stakes (DeRose 2005): a person has some evidence that their bank will be open on a given day; in one case, nothing bad will happen if the person is wrong; in another case, it will be bad if they turn out to be wrong. Philosophers often judge that the person knows that the bank will be open in the "low stakes" case, but not in the "high stakes" case. They therefore conclude that knowledge ascriptions and possibly knowledge itself are in some way sensitive to practical factors like stakes.

In contrast to philosophers' consensus, empirical results examining whether knowledge ascription is influenced by stakes have been mixed. Some studies presenting bank cases to nonphilosophers find evidence that knowledge ascriptions are sensitive to changes in stakes; 2 others fail to find evidence. 3 While most studies have been limited to English-speaking participants, Rose et al. (2019) found that the stakes involved did not notably alter knowledge ascriptions across languages and cultures, with few exceptions. 4 Some of the variation in results is explained by the use of two different methods: evidencefixed, and evidence-seeking. In the former, participants are told how much evidence the character in the cases has gathered and are asked to assign knowledge to them; in the latter (introduced by Pinillos, 2012), participants are asked how much evidence the character in the vignette should obtain in order to know some proposition. Almost all the evidence against stakes effects has come from evidence-fixed methods; evidence-seeking prompts consistently find stakes effects. 5 Francis and Beaman (2023) argue that in contrast to evidence-seeking methods, evidence-fixed methods and Schaffer 2015; Feltz and Zarpentine 2010; Francis et al. 2019; and Rose et al. 2019). But there are notable exceptions: Sripada and Stanley (2012) find a small stakes effect, although the effect does not replicate in Francis et al. (2019); Francis and Beaman (2023) find a stakes effect with their evidence-fixed prompt; and Pinillos (forthcoming) finds a stakes effect using a modified evidence-fixed prompt. Turri et al. (2016) find evidence of a stakes effect in one of their two experiments, but argue that the stakes effect is only indirectly on knowledge; stakes affect judgements about actionability, which in turn affect knowledge judgements.

have produced inconsistent results, and that this inconsistency may be evidence that this method is flawed. 6 We interpret the disagreements between the two methods differently: we speculate that the different results between evidence-fixed and evidence-seeking methods reflect a tension between people's epistemic beliefs and practices. The relevance of stakes varies depending on whether people are asked to attribute knowledge or are asked to reflect on the conditions for attributing knowledge. To examine this tension, our study asked participants both to assign (or decline to assign) knowledge to someone having a given body of evidence and to decide how much evidence would be needed to know something. To address the concerns raised by Jackson (2021) and Francis et al. (2019), we also use a more realistic scenario (particularly for small-scale societies), involving a dam on a river, and participants read a "high stakes" scenario in which the stakes are life or death.

Finally, this study extends earlier cross-cultural research such as Rose et al. (2019) by considering a broader and more diverse participant base. Our sample includes individuals from 17 distinct populations across 11 countries, speaking 14 languages. Prior empirical work on stakes-sensitivity has focused primarily on highly-educated and urban populations; this study incorporates a wider range of participants, with an emphasis on including populations-such as indigenous communities-that have been underrepresented in prior work.

This study was part of a larger study on knowledge and understanding, carried out under the Geography of Philosophy project (www.geographyofphilosophy.com); the preregistration for this study can be found at https://osf.io/ym45u. Data was collected from convenience samples across 17 sites in 11 countries, using the online survey platform Qualtrics at some sites, and paper and pencil at others (see Table 1 for details). A total of 5,573 participants were surveyed, with a minimum of 80 participants in each population. 7 Participants were randomly and evenly assigned to one of four conditions, based on a 2 × 2 design (Evidence: weak vs. strong; Stakes: low vs. high). Each group saw a different continuation of the following vignette:

[name]'s 8 kids want to go swimming in a local river. The water has slightly risen due to several days of light rain. So he decides to build a dam. He stacks a number of logs in the river to hold the water back and create a shallow pool where his kids can swim.

Participants were either told that the protagonist checked the logs once, or that he checked multiple times: 6 Different theories have been offered to explain what exactly the flaw is. Jackson (2021) blames the failure to find evidence for the stakes-sensitivity of knowledge attributions with evidence-fixed prompts on the unrealistic nature of the scenarios used in many of these studies. Francis et al. (2019) argue that stakes are scalar-whether stakes are "high" or "low" is a matter of degree-and that at least some experiments may have failed to find evidence of stakes-sensitivity simply because the "high stakes" and "low stakes" scenarios used in those studies were too close together on the scale.

Demographic Characteristics Sample Language Students Age a Gender Percentage M/F/other N Passed Check Passed Check Percentage mean range sd China Mandarin both 22.75 18-52 5.53 43.9/55.8/0.3 416 261 62.7 Russia Russian non-students 25.97 18-75 10.28 40/60/0 739 558 75.5 Slovakia Slovak both 26.28 18-74 9.07 54.4/44.9/0.7 713 290 40.7 Ecuador Spanish students 21.39 18-42 3.21 32.9/66.7/0.5 390 208 53.3 India Hindi both 29.98 18-62 10.23 45.3/53.8/0.9 200 100 50 India Meitei non-students 23.26 18-45 6.26 47.6/50/2.4 86 42 48.8 Japan Japanese both 30.31 19-62 10.76 53.3/46.3/0.4 544 445 81.8 Morocco Arabic both 32.96 18-76 11.81 44.4/55.6/0 551 323 58.6 Peru Shipibo non-students 34.68 18-83 13.77 53.9/46.1/0 200 102 51 Peru Spanish students 22.75 18-49 3.28 43.1/56.9/0 286 216 75.5 South Africa Afrikaans both 41.63 18-73 12.85 34.3/65.7/0 270 178 65.9 South Africa Sepedi non-students 29.33 18-57 7.88 36.5/62/1.5 258 122 47.3 South Africa isiZulu non-students 31.05 18-76 9.84 27.9/69.9/2.2 268 130 48.5 South Korea Korean both 35.40 18-68 10.37 41.5/58.5/0 256 193 75.4 United States English students 35.85 19-73 10.52 55.5/44.5/0 396 348 87.9 a

Participants who reported an age under 18 were excluded from the study; participants who input non-numeric text or a number greater than 100 were excluded from our age calculations here, but were otherwise included in the study.

Weak evidence: Once he has finished building the dam, he checks the logs once to make sure the dam is secure.

Strong evidence: Once he has finished building the dam, he checks the logs several times to make sure the dam is secure.foot_3

Participants were also either told that it was very important that the dam was secure, or that it was not at all important:

High stakes: It is very important that the dam is secure. Due to the high water level, if it breaks, his kids will be swept away by the water.

Low stakes: It is not at all important that the dam is secure. Due to the low water level, if it breaks, the water won't even rise to knee level.

For example, here is the weak evidence/high stakes version of the vignette:

[name]'s kids want to go swimming in a local river. The water has slightly risen due to several days of light rain. So he decides to build a dam. He stacks a number of logs in the river to hold the water back and create a shallow pool where his kids can swim. Once he has finished building the dam, he checks the logs once to make sure the dam is secure.

It is very important that the dam is secure. Due to the high water level, if it breaks, his kids will be swept away by the water.

Participants were first asked a comprehension question to determine whether they recognized the stakes in the vignette:

Participants were then given an evidence-fixed prompt, in which they can either attribute knowledge or decline to do so:

Evidence-fixed: [name] ____________ that the dam is secure.foot_4

Participants were finally given an evidence-seeking prompt, in which they are asked how much evidence is required for knowledge:

Evidence-seeking: How many times do you think All three prompts were presented on the same page, in the order given above. Vignettes and prompts were translated into the target languages by competent native speakers and presented in the participants' native languages. 11 Based on Rose et al. ( 2019), we predicted that stakes would not influence knowledge ascriptions or the number of times the protagonist needed to check. We further hypothesized that the protagonist's amount of evidence would not influence knowledge ascription or the number of checks needed for knowledge.

2,057 participants were excluded either for failing the comprehension check, failing to complete the survey, or for reporting an age below 18. 13 Nearly all participants in the high stakes condition passed the comprehension check (Figure 1). However, in the low stakes condition, participants incorrectly answered that the dam being secure was "very important" at higher than expected rates. There was wide variation across populations; this may be due to translations issues, or to cultural variation in participants' willingness to accept the vignette's assertion that the children are not in danger in the low stakes scenario. 14

11 They were also backtranslated to check the accuracy of the translation.

12 The data and the code can be found at osf.io/tyge9/files/osfstorage. 13 Exclusion criteria were not preregistered, although the presence of a comprehension check was preregistered.

14 Due to the high rate of failures for the comprehension check, we also reanalyzed the data without comprehension check exclusions. The results were very similar to those reported below, and can be found in our supplementary materials on OSF at https://osf.io/tyge9/files/osfstorage.

TA B L E 2 Stakes Effects for the Three Models for the Evidence-Fixed Prompt Model Estimate Std. Error z value Pr(>|z|) Odds Ratio M 1 0.02864 0.03758 0.762 0.4460 1.0290553 M 2 -0.10123 0.04028 -2.513 0.0120 0.9037275 M 3 0.02579 0.09105 0.283 0.7770 1.026121 TA B L E 3 Model Fit of the Three Nested Models for the Evidence-Fixed Prompt Model npar AIC BIC Log Likelihood Chisq Pr(>Chisq) M 1 4 4 5 8 7 . 5 4 6 1 2 . 1 -2289.7 M 2 5 4308.9 4339.8 -2149.5 280.51 < 2.2e-16 M 3 7 4299.6 4342.8 -2142.8 13.30 0.001294

Removing participants who failed the comprehension check, we analyzed the remaining 3,516 responses to the evidence-fixed knowledge ascription prompt using three nested logistic regression models:foot_5

1. A simple linear model M 1 : Ascription s ∼ β 0 + β 1 Stakes s + β 2 Evidence s + β 3 Stakes s *Evidence s + ε S ; 2. A mixed-effects model that varies the intercept by population (P) M 2 : Ascription s∈P ∼ β 0 + β 0P + β 1 Stakes s + β 2 Evidence s + β 3 Stakes s *Evidence s + ε S ; 3. A mixed-effects model that varies the intercept as well as the slope of the stakes factor by population (P) M 3 : Ascription s∈P ∼ β 0 + β 0P + (β 1 +β 1P )Stakes s + β 2 Evidence s + β 3 Stakes s *Evidence s + ε S .

The results for the stakes effect in each model can be found in the table below. Following Benjamin et al. (2018), we set the standard for statistical significance at .005. We compare the fit of the models using ANOVA, and find that the model that varies both intercept and slope provides the best fit for the data (Tables 2 and 3). As predicted, in none of the models do we find a statistically significant stakes effect (Figure 2). 16 We do find a suggestive (p = 0.0120) stakes effect in M 2 , but do not even find a suggestive effect in M 3 , which provides a significantly better fit. However, in M 3 we do find a suggestive p-value replaced the stakes condition variable with participants' responses to the comprehension check. The idea is that responses to the comprehension check, in which participants report whether or not it is "very important" that the character's belief is true, can be taken to measure perceived stakes. We did not find the predicted stakes effect in those models; the p-value for perceived stakes is significant (0.000687) in this version of M 1 , the simple logistic regression model with no random effects, but the effect is not in the expected direction: the model estimates that higher importance increases the likelihood of knowledge attribution. We do not find even a suggestive effect in M 2 (0.4189) or M 3 (0.3740), both of which provide a significantly better fit than M 1 . (p = 0.0473) that the protagonist's amount of evidence affects knowledge ascriptions (Figure 3). This is contrary to our preregistered predictions, but on second thought it is not surprising that stronger evidence results in higher rates of knowledge ascription.

TA B L E 4 Stakes Effects Per Sample for the Evidence-Fixed Prompt Sample Estimate Std. Error z value Pr(>|z|) Odds Ratio China -0.1053 0.1763 -0.597 0.5504 0.900064 Russia -0.12438 0.09253 -1.344 0.179 0.8830474 Slovakia -0.53170 0.13445 -3.955 7.67e-05 0.5876024 Ecuador -Spanish 0.07867 0.15698 0.501 0.6163 1.0818479 India -Hindi -0.1369 0.2462 -0.556 0.578 0.8720179 India -Meitei 0.4236 0.7480 0.566 0.571 1.5275252 Japan -0.30100 0.12327 -2.442 0.0146 0.7400757 Morocco -0.1050 0.1304 -0.805 0.42072 0.9003190 Peru -Shipibo 5.154 442.314 0.012 0.991 173.18192856 Peru -Spanish -0.09741 0.16222 -0.600 0.548 0.9071871 South Africa -Afrikaans -0.07561 0.17525 -0.431 0.666 0.9271746 South Africa -Sepedi 0.4357 0.3535 1.233 0.218 1.5461104 South Africa -isiZulu 0.08355 0.39391 0.212 0.832 1.0871401 South Korea 0.28943 0.17864 1.620 0.105 1.3356652 United States -0.1457 0.1124 -1.296 0.195 0.8644485

Noteworthily, we observed some limited cultural variation (Table 4). 17 In particular, we find a significant stakes effect in the Slovakia sample (p = 7.67e-05) and a suggestive effect in the Japan sample (p = 0.0146). 18 We also find a statistically significant interaction between the stakes condition and the number of times the protagonist has checked in Morocco. 19 This suggests that within the Moroccan population sampled, there is a statistically significant stakes effect when the protagonist has checked once, but not when the protagonist has checked multiple times. This could be taken as evidence of a stakes effect on knowledge ascription in the Moroccan population: checking multiple times is sufficient evidence for knowledge in both the high and low stakes cases, but checking once is only sufficient evidence in the low stakes case.

However, even in those populations where we see evidence of a stakes effect, the effect size is relatively small. Even in Slovakia, the odds ratio is 1.7; which, following Chen et al. (2010), is a small effect, roughly equivalent to a Cohen's d of 0.2.

Overall, our evidence-fixed prompt finds little to no evidence of the sort of stakes effect on knowledge ascription that philosophers expect.

Before turning to the evidence-seeking prompt, we should examine whether these results are vitiated by the high number of participants failing the comprehension check (see Table 1). Since the comprehension check was binary, half of the participants answering randomly may have answered correctly by chance, and their random answers in the rest of the survey may explain why we failed to find any difference between the high and low stakes conditions. 20 There are two main things to say in response to this particular concern. First, failure to answer the comprehension check is not mostly due to random answering, but to participants assigned to the low states condition viewing the scenario as high stakes. Since these participants were removed for the analysis reported in the main text, concerns about random answering do not apply to our results. (Relatedly, note that including all participants and using the comprehension check as a measure of the perceived stakes does not change our results-see footnote 17). Second, this objection does not explain why if people answer randomly, we found systematic differences with the evidence-seeking prompt that we discuss next.

Our evidence-seeking prompt was a free response question, and we had to code participants' answers. Although some answers were easily quantified ("5", or "three times"), others were harder to quantify ("many times", or "no guarantee no matter how many times") and some were straightforwardly unquantifiable ("I have no idea" or "no"). To convert responses to numeric data, we used the following rules:

1. If a response included digits ('2', '25', '100', etc.) the number was extracted from that response (so 'not more than 10 times' would be stored as 10) 2. Number words like "zero", "four", "eleven", "thousand" were converted to integers (0, 4, 11, 1000). After "twelve" we skipped straight to "hundred", "thousand", and "million". So any response that said "twenty" would be ignored. 3. Plurals were doubled. So "dozens" was stored as "24", "hundreds" was stored as 200, etc. 4. "Once" and "twice" were stored as 1 and 2 (so "at least once" and "more than once" were both stored as 1). 5. The term "couple" was stored as 2; "few" was stored as 3; "many", "several", "frequently", "lot", "lots", and "repeatedly" were all stored as 10. 6. Commas were removed, so that "1,000" and "1,000,000" were stored as 1000 and 1000000, respectively. 7. If multiple numbers could be extracted from an answer, the highest number was chosen.

So "three or four times" was stored as 4. 8. All numbers extracted that were greater than 100 were stored as 100. This is to deal with outliers. 21 Using this process, we were able to extract numbers from 2,833 responses, leaving 815 responses that could not be converted to numeric data. The results are shown in Figures 4 and 5. Again, we fit three models:

1. A simple linear model M1: Response S ∼ β 0 + β 1 Stakes S + β 2 Evidence S + β 3 Stakes S * Evidence S + ε S ; 2. A mixed-effects model that varies the intercept by population (P) M2: Response S∈P ∼ β 0 + β 0P + β 1 Stakes S + β 2 Evidence S + β 3 Stakes S *Evidence S + ε S ; TA B L E 5 Stakes Effect for the Three Nested Models for the Evidence-Seeking Prompt Model formula Estimate Std. Error z value Pr(>|z|) M 1 0.391371 0.008925 43.853 < 2e-16 M 2 0.406727 0.008983 45.276 < 2e-16 M 3 0.322831 0.058406 5.527 3.25e-08 TA B L E 6 Model Fit for the Three Nested Models for the Evidence-Seeking Prompt Model formula npar AIC BIC Log Likelihood Chisq Pr(>Chisq) M 1 4 4 3 3 7 2 4 3 3 9 6 -21682 M 2 5 41170 41200 -20580 2203.95 < 2.2e-16 M 3 7 4 0 8 2 0 4 0 8 6 2 -20403 353.99 < 2.2e-16

3. A mixed-effects model that varies the intercept as well as the slope of the stakes factor by population (P) M3:

Models were fit to a Poisson distribution, with a natural log link function. We find a statistically significant stakes effect on participant responses. 22 The results for the stakes effect in each model can be found in Table 5. 23 Once again, an ANOVA reveals that the model with varied intercept and slope provides the best fit (Table 6).

We see similar effects across most cultures (Table 7). We find statistically significant stakes effects in 8 out of 14 samples tested, and a suggestive effect (0.005 < p < 0.05) in 2 samples. But it is worth noting that in one sample, South Africa -Afrikaans, we find a statistically significant interaction effect between stakes and evidence, suggesting that at least in the low evidence condition there is a statistically significant stakes effect in a surprising direction: participants report lower numbers of checks required when the stakes are higher. It is not clear how to interpret this result. However, we take the highly significant stakes effect found in the full cross-cultural sample (in the expected direction), combined with the significant or suggestive stakes effect in the expected direction found in 10 out of 14 samples, as strong evidence that there is a stakes effect on evidence-seeking responses across cultures.foot_14

TA B L E 7 Stakes Effects Per Sample for the Evidence-Seeking Prompt Sample N Estimate Std. Error z value Pr(>|z|) China 182 0.03758 0.09836 2.278 0.0227 Russia 530 0.54644 0.02207 24.758 < 2e-16 Slovakia 243 0.68343 0.03532 19.350 < 2e-16 Ecuador -Spanish 172 0.50301 0.03878 12.970 < 2e-16 India -Hindi 84 0.37698 0.06906 5.459 4.79e-08 India -Meitei 37 0.216273 0.164037 1.318 0.187 Japan 380 0.389493 0.019106 20.386 < 2e-16 Morocco 301 0.30593 0.03819 8.010 1.15e-15 Peru -Shipibo a 3 N A N A N A N A Peru -Spanish 126 0.11756 0.04612 2.549 0.01080 South Africa -Afrikaans 132 -0.06780 0.05164 -1.313 0.189 South Africa -Sepedi 22 0.364944 0.233666 1.562 0.118 South Africa -isiZulu 12 0.04766 0.24772 0.192 0.8474 South Korea 193 0.40194 0.02529 15.893 2e-16 United States 327 0.41068 0.02588 15.870 < 2e-16

a We were unable to test for a stakes effect in the Peru-Shipibo population, because all 3 of the responses that could be converted to numeric data were from participants in the "high stakes" condition. The numbers extracted were 5, 5, and 4.

One might worry that the fixed order of the comprehension check and the two prompts (comprehension check before the two prompts) might have influenced our results. 25 Perhaps the comprehension check ("It is very important that the dam is secure") makes the stakes salient and influences the answers to the prompts following the comprehension check. It would indeed have been good to vary the presentation order of the comprehension check and the prompts, but we made the choice to simplify the data collection process (particularly for sites where data collection was arduous) by using only one presentation order. On the other hand, it is not clear how the comprehension check is meant to influence participants' answers. Perhaps it highlights the similarities between the low and high stakes conditions (a dam might break in both conditions); but, equally plausibly, it might highlight the difference between the high (lives threatened) and low (no life threatened) stakes condition at least for those participants who answer correctly the comprehension check. Furthermore, it is not clear at all how the comprehension check is meant to explain the main result of our study, viz. the difference between evidence-fixed and evidence-seeking prompts.

Our results are consistent with previous findings, in which evidence-fixed prompts have only rarely found evidence of a stakes effect, while evidence-seeking prompts consistently find evidence of a stakes effect. This difference has sometimes been taken to offer a methodological puzzle: two different methodologies give us different results, and the puzzle is to figure out which methodology is the reliable one, and what makes the other unreliable. Much of the debate has therefore been centered on explaining why one of the two methods is flawed, and why the other method should be trusted. Defenders of stakes effects tend to argue that evidence-seeking prompts show that ordinary knowledge ascriptions are affected by stakes, and that evidence-fixed prompts fail to find one because the tests are flawed for one reason or another.foot_16 Supporters of stakes effects tend to that evidence-fixed prompts show that there is no stakes effect on ordinary knowledge ascriptions, and that evidence-seeking prompts seem to find one because those tests are flawed for one reason or another. 27 Instead, we propose that neither methodology is flawed: both kinds of prompt are successfully testing different aspects of ordinary knowledge ascription, and neither set of results needs to be explained away. Evidence-fixed prompts ask participants to ascribe knowledge; by contrast, evidence-seeking prompts ask participants how much evidence a character would need in order to have knowledge. In answering this question, participants are not ascribing knowledge or declining to do so; they are expressing their beliefs about when knowledge ascriptions would be appropriate. The former prompt tests practice; the latter, beliefs about this practice.

The difference in results between the two kinds of prompts is naturally explained by the observation that the two prompts test different things. Although one might expect that these two prompts coincide, they do not: people's belief about the evidence required for knowledge is not aligned with their own practice of knowledge ascription. Misalignments between our practices and beliefs about them are of course not unusual. People believe that their assessment of others' performance is not affected by their race or gender, although it often is; people believe that they are good at detecting lies, while nearly all of us are lousy lie detectors (e.g., Bond and DePaulo, 2006).

The misalignment between people's belief about and their practice of knowledge ascription raises an interesting question. By their own lights, people must be making a mistake: if their belief about knowledge ascription is right, then they are mistaken to overlook the stakes when they ascribe knowledge; if their knowledge ascriptions in low and high stakes conditions are correct, then they are mistaken to believe that stakes should matter for knowledge ascription. Similarly, epistemologists might think that lay people's practice of knowledge ascription or their belief about it is mistaken. Classical invariantists, who argue that the truth-value of knowledge ascriptions does not shift across contexts, might argue it is the practice that is correct, and that people's belief about knowledge ascription is simply mistaken. Alternatively, contextualists such as DeRose, who hold that the truth-value of knowledge ascriptions is to be assessed relative to a context, and interest-relative invariantists such as Stanley, who hold that the truth-value of knowledge ascriptions depend on practical facts such as stakes, might argue that the belief is correct, and it is the practice that is mistaken. On this view, people's belief that stakes affect the truth conditions of knowledge ascriptions shows that there really is a stakes effect on (true) knowledge ascriptions; the fact that actual practice does not align with this belief means that people are edge in both high and low stakes scenarios. Pinillos (2012) and Sripada and Stanley (2012) suggest that participants assume that the protagonist in the high-stakes scenario has gathered evidence more thoroughly than the protagonist in the low stakes scenario. making systematic errors in ascribing knowledge, either over-ascribing when stakes are high or under-ascribing when stakes are low. 28 However, even if either people's practice or their belief about it must be incorrect, there might be no fact of the matter about which is incorrect. One relevant piece of information is what people do when presented with the inconsistency between their belief and their practice. If people judge that they were mistaken in ignoring stakes when they ascribe knowledge, then this give us a reason to follow the lead of contextualists and interest-relative invariantists. Alternatively, they might not have any opinion on the matter. In future work, we plan to investigate this question.

We presented evidence showing that in many languages and countries people do not take stakes into account when they assign knowledge in low-and high-stakes situations, using a more realistic scenario than in previous research in a more diverse group of populations. We have also shown that these same people nonetheless judge that stakes matter when asked to decide how much evidence is needed to know something. People's answers reveal an unexpected tension in folk epistemology between people's practice of knowledge ascription and their belief about it.