What Does Nick Bostrom Think?

Predicting unstated beliefs is often harder than you think

The recent controversy surrounding Nick Bostrom inspired this post, but its lessons can be applied widely. For example, it is relevant for understanding the arguments in this conversation between Coleman Hughes and Charles Murray. The point is very simple: you probably don’t know everything Nick believes about race just because you know he believes Whites are more intelligent than Blacks. Let’s get started by looking at this contingency table

This table works with counts or proportions. Say you have two variables, X and Y, one for each directional combination of the binary outcomes High and Low. If the two are positively associated, you will have an excess of correspondences in the boxes A and D. Say you have a correlation of r = 0.20. The corresponding contingency table has concordant pairs adding up to 60%:

This observation was popularized by one of the fathers of meta-analysis, Bob Rosenthal. He dubbed this the Binomial Effect Size Display (BESD) and it amounted to the observation that correlations can be thought of like flipping an increasingly rigged coin. If you were selecting someone to hire to work a job and you had a correlation of 0.50 between holding a degree and being a good employee, you know that picking a person based on possessing a degree will get you a good worker 75% of the time, which is what the corresponding pairs in the table will add to. If the correlation is only 0.30, you’ll get a good worker 65% of the time. The rigging of the coin is given by the neutral coin’s odds of 50% plus half the correlation multiplied by 100.

A clever reader might realize that this discovery by Rosenthal shouldn’t be used to describe r, but instead to describe another association statistic, Kendall’s τ (that’s tau)¹. τ is a measure of concordance calculated when we take the number of high/high and low/low estimates, call them concordant and contrast them with the number of high/low and low/high estimates we dub discordant, before we subtract the number of concordant pairs from the number of discordant pairs, and then divide that quotient by the total number of pairs to get our τ of 0.2. Ta-da!

The basics are easy to understand. With a correlation of 1 between two attitude measures, everyone’s attitudes on one perfectly proxies their attitudes on another and they go in the opposite direction with a correlation of -1. With a correlation of 0, knowing a person’s attitude towards one thing tells you nothing about their attitude towards some other thing. For correlations between these values, knowing one attitude will tend to tell you about the attitudes it is correlated with to varying degrees. If linearity does not hold, however, our predictive powers may not hold. The question we really want to ask is not how associated two things are, but instead, given one thing, how well can we predict something else? I might have made these seem similar, but nonlinearity breaks them apart, as we will see.

---

In Bostrom’s dramatic case, he expressed the view that Blacks are less intelligent than Whites. Bostrom probably holds that belief because he is an avid reader and he is not the sort of person to get upset by things that are contentious but ultimately empirical. The idea that he holds this view or any other one because he’s a hateful person who is simply biased against Blacks (or X, Y, and Z) is based on nothing for a few reasons.

Firstly, there is no reason holding this or any other serious belief about group differences is necessarily racist or in any way evidence of prejudice. Prejudice is an operative word here, since taking a view because it has turned out to be empirically confirmed is not a pre-judgment. Moreover, adopting a view that is distasteful because it is true is also not prejudicial or an indicator of bias. Secondly, honest people will have to admit that intelligence is general, both etiologically and phenotypically, and that its measurement in the racial/ethnic groups “Black”, “White”, “Asian”, and “Hispanic” in the US today is done without bias. To quote a recent review on the subject of test bias,

Single-group or differential validity has not been found and likely is not an existing phenomenon with regard to well-constructed standardized psychological tests. [This refers to the predictive validity of tests.]

and

No consistent evidence of bias in construct measurement has been found in the many prominent standardized tests investigated. [This refers to violations of measurement invariance.]

and

Content bias in well-prepared standardized tests is irregular in its occurrence, and no common characteristics of items that are found to be biased can be ascertained by expert judges. [This refers to the common observation that psychometrically biased test items are not intuitively knowable. In fact, they often make little sense whatsoever as they occur in items with all sorts irrespective of their content.]

The claim that tests are biased has simply not been substantiated by more than a half-century of research on the topic and the lack of bias is important, because it means that the measured differences between these groups are real. Hence, the repeated discovery that representative samples of Blacks and Whites differ in measured IQs (or better, levels of g) by about one d means Whites do, in fact, tend to be more intelligent. There is obviously overlap in these group’s distributions, but that is such a trite note that it’s a waste of space to say it in reasonable company. An almost up-to-date summary of this gap is given in the below plot.

Since this plot was made, a few more studies have come out, new years of SAT and ACT data have dropped, norming sample data has been published about or otherwise become available, and so on and so forth, and those results have brought the mean gap up slightly to about 1.08 d, but for simplicity, let’s stick to 1.

This gap was the same size when Bostrom wrote his email to the Extropians listserv in 1996, just two short years before the publication of Arthur Jensen’s The g Factor, a text that provided ample discussion of the gap from every angle. Bostrom probably knew about the gap, as the subject used to be less taboo and it is relevant for a person interested in human enhancement like he is. In the not-so-distant past the email originated in, you could get away more easily with publishing about the gap, talking about it, and even broaching the possibility that it had some policy relevance somewhere, somehow. This fact is virtually unavoidable if you have any interest in intelligence research, as Bostrom did and does.

Bostrom is likely still aware that intelligence differs between racial groups, and that has nothing to do with hate, it is a simply a matter of following the research. And it is good research!

Don’t just take it from me.

It would be better for all of us if we acknowledged the gap, since doing so may motivate strenuous efforts to understand and alleviate it, or even to figure out how to allow people to overcome intelligence deficits. It could figure as a motivation behind nootropic research and the promotion of reproductive technologies that can help, like embryo selection or the use of genetic information for early detection of cognitive deficits that can be managed. This will be returned to.

---

Alright, so as a scientist with an interest in intelligence, Bostrom probably acknowledged the Black-White gap because of his awareness of the mountain of research dealing with group differences. It’s true he could have thought negatively about Blacks and sought out intelligence research to justify whatever his beliefs may be, but assuming racism is extreme and wanting for evidence. What about regular people who likely haven’t ever done more than skim a scientific paper? We do have some data.

For this test, I will be using the General Social Survey (GSS). The GSS is a rich dataset administered over many years, replete with questions about every topic under the sun, from a vocabulary test (the wordsum) to questions regarding contemporary political issues. Here are my most important variables:

• The first is race, which I will use for subsetting to the White sample for the analyses minus correlating the next two variables.

• The next two are intlwhts and

• intlblks, which are answers to the questions “Do people in this group tend to be unintelligent or tend to be intelligent?” for Whites and Blacks, respectively (1 - 7).

• The next variables are about explicit or possible racism towards Blacks. For explicit racism, we have marblk, “Would you oppose a close relative or family member marrying a Black person?”, and

• racpres, “If your party nominated a Negro/Black/African-American for President, would you vote for him if he were qualified for the job?”.

• For possible racism, we have natrace, “Are we spending too much, too little, or about the right amount on improving the conditions of Blacks?” (1 - 3),

• helpblk, “Some people think that Blacks/Negroes/African-Americans have been discriminated against for so long that the government has a special obligation to improve their living standards. Others believe that the government should not be giving special treatment to Blacks/Negroes/African-Americans. Where would you place yourself on this scale, or haven’t you made up your mind on this?” (1 - 5),

• racpush, “Negroes/Blacks/African-Americans shouldn’t push themselves where they’re not wanted”, (1 - 4).

We can use intlwhts and intlblks to compute scores for whether a person thinks Blacks, Whites, or neither are more intelligent by taking their difference. But first, the distribution of intelligence estimates for either race in the White subsample of the GSS looks like this

Clearly most people seem to be biased towards the center of the scale since it probably feels like a neutral point. This is even clearer when the data is presented in a response matrix.

In the Black sample, 34.2% of people responded with the scale’s center for their ratings of Blacks and Whites, and in the White sample, 42.4% did. Curiously, in the White sample, the correlation between intlwhts and intlblks was r = 0.316 and in the Black sample, it was 0.485². Variables were dichotomized from this point forward.

Looking at the first explicit racism variable, there was a small but temporally consistent correlation of around 0.23 between the belief that Whites are more intelligent and disapproval of a relative engaging in interracial marriage with a Black person.

Data for the other explicit racism question, about whether someone would vote for a qualified Black presidential candidate from their party, was only asked in 1996, 1998, 2000, and 2002, and it correlated at around 0.19 with the belief that Whites are more intelligent. One of the possible racism measures, about whether Blacks have pushed “themselves where they’re not wanted” was also only asked in a few years. It correlated at 0.20 with the belief that Whites are more intelligent. The results for the other two possible racism measures were less interesting:

These correlated, respectively, at 0.09 and 0.10 with the belief that Whites were more intelligent, but results were often not significant. Given these correlations, we should have the following contingency tables under linearity.

Left to right, top to bottom: marblk, racpres, natrace, helpblk, racpush. Theoretical contingency tables.

But we, in fact, do not.

Left to right, top to bottom: marblk, racpres, natrace, helpblk, racpush. Empirical contingency tables.

Why did the correlations end up so wrong? Linearity failed! The odds ratios for comparisons of people who believe Whites are more intelligent and those who do not are considerable, but knowing a person believes Whites are more intelligent than Blacks doesn’t help you at all for the two explicitly racist beliefs, it doesn’t help you very much for the first two possibly racist beliefs, and it may actually help you for the final possibly racist belief.

In order, if you know someone believes Whites are more intelligent and nothing else, you have a coin flip. Given that 85% of the people who believe Whites are more intelligent would vote for a qualified Black person running for president with their party of choice, so with them, you have a very rigged coin flip, with odds on “not racist” coming up.

For the more dubiously racist beliefs, 27% of the people who believed Whites are more intelligent also believed we give too much money to improving the conditions of Blacks, which compares to 36% of those who don’t believe Whites are more intelligent. A mere 10% of the people who believed Whites are more intelligent than Blacks think we should give Blacks special treatment, which compares to 18% of the people who don’t. 44% of the people who believed Whites are more intelligent than Blacks don’t think Blacks shouldn’t push themselves where they’re not wanted, compared to 68% of the people who don’t.

For the two things that are hard to explain as anything but racism, you can either not predict people’s views on that topic by knowing their views on intelligence differences, or they probably give the non-racist response regardless of their belief about intelligence differences. For two out of the three maybe racist things, the result is slightly worse than the rest of population, and for the last, it’s quite a bit worse, but these measures are only dubiously about racism in the first place.³

---

With general population facts out of the way, we still don’t know why the belief that Whites are more intelligent than Blacks is related to these things in these magnitudes, but we can at least now say that there’s not a particularly substantial statistical risk presently associated with the belief White are more intelligent than Blacks.⁴ For comparison with everything else up there, the intelligence gap between Blacks and Whites is one d, or a correlation of 0.45.

To get to the how, I asked non-researchers in several school Discord group chats and subreddits to take am anonymous survey.⁵ I simultaneously asked 35 intelligence researchers from the latest completed issues of Intelligence and The Journal of Intelligence to take a questionnaire with a few unique questions.

I ended up with 236 non-researcher responses and 21 intelligence researcher responses. An additional five intelligence researchers opened but did not take the survey. The group of non-researchers was winnowed by purging people who failed the attention check, people who completed the questionnaire in less than two minutes, people who were not White, people who said they had personally researched the topic of intelligence or achievement differences between groups before, people who repeated the survey, people who used a nonexistent or fake email address, people who replied that they lied or did not take the survey seriously, people who failed to complete all the questions, and people who identified themselves as researchers in some capacity. There were too few members of the non-researcher group that had done prior research on this topic (n = 19) to really make a difference given how poorly I believe people’s definitions of research tends to be when they’re not really researchers.

For the non-researchers, I asked the questions from the GSS in addition to the question “Do you dislike Black people?” as another way to get at explicit racism. I then provided them with the following blurbs

An average or mean is a number that summarizes a set of values. It is the sum of the values in the set divided by the number of values included in the set. For example, the average of the set {1, 2, 8, 9, 3} is 4.6 and the average of the set {6, 8, 2, 4, 9} is 5.8.

A standard deviation is a number that summarizes the variation in a set of values relative to the mean. It is the square-root of the squared sum of each value in a set minus the mean divided by the size of the set. For example, the standard deviation of the set {1, 2, 8, 9, 3} is 3.65 and the standard deviation of the set {6, 8, 2, 4, 9} is 2.86.

They were then given the opportunity to enter numbers separated by commas in an applet that output the mean and standard deviation of that set and they could move on whenever they wanted. They were then told

Researchers have discovered that the measured average intelligence level of Black Americans in terms of IQ points is 85 with a standard deviation of 15. The average intelligence level of White Americans has been measured at 100 with a standard deviation of 15 (Roth et al., 2006; Frisby & Beaujean, 2015; Fuerst, Hu & Connor, 2021). This means that 15.9% of Black Americans are as intelligent as the average White American.

And asked

Do you consider this information important to know when you think about Black Americans as a group?

And

Has this information changed any of your views about Black Americans?

And finally, the racism questions were asked a second time to assess change.

For the intelligence researchers, I asked if they believed in intelligence differences between Blacks and Whites, whether they were familiar with the research on group differences in intelligence, and whether they believed there was a non-zero genetic contribution to the differences. They were then asked to predict how people’s views would change in response to being made aware of the intelligence differences between White and Black Americans. Here are the results.

Firstly, very few non-researchers said they disliked Blacks, disapproved of a relative marrying a Black person, or said they wouldn’t vote for a qualified Black presidential candidate from their party. In every response category, there were nonsignificant shifts towards the racist belief but the only significant change was a movement towards the belief that Whites are smarter than Blacks. To interpret the chart labels understand that more people said Blacks shouldn’t push where they don’t belong, more people believed we spend enough or too much on Blacks, and more people said we shouldn’t offer Blacks special treatment, but in no case were these changes significant. Only small portions of the sample believed the intelligence difference information was important and, oddly enough given the nonsignificant changes in their other responses, a handful said this information changed their views towards Black Americans. Only a handful of people typed what that meant to them below that question.

The results for researchers may have been more interesting.⁶

None disliked Blacks, none disapproved of a relative marrying a Black person, and none would avoid voting for a qualified Black presidential candidate from their party. Very few said Blacks shouldn’t push where they’re not wanted, about 40% said we spend enough or too much, and a few supported special treatment for Blacks. For the researcher-specific questions, a narrow majority thought information about group differences would lead people to change their views in a deleterious way.

Most (18) of the researchers claimed to be familiar with the literature on group differences in intelligence and most of those claimed to believe in the differences (17), while most of them believed Whites had been shown to be more intelligent (16), and most of them believed there was a nonzero genetic contribution to the difference (15).

Whether comparing individuals who ought to have been knowledgeable about group differences or people who had just learned of them, there was little evidence that knowledge about intelligence differences between Blacks and Whites was related to much racist belief. For the researcher sample, this might be explained by selection, and for both groups, it could be explained by bad sampling or lying. We can only take for granted that most people do not lie about their views on anonymous surveys.

---

What of Bostrom? He’s a person whose research has undoubtedly exposed him to the literature on group differences and intelligence more broadly. I believe his views are probably like other researchers in that domain: based on large, meaningful, and replicable group differences, and unrelated to racial prejudice. There’s certainly no reason to believe he’s a racially prejudiced person if we acknowledge that a belief in group differences in intelligence fails to manifest them.

What about Asians and Jews? Intelligence research supports substantially higher IQs for Asians and Jews than for Whites. I have never seen someone worried that the result of mass acceptance of intelligence research would result in widespread Asian or Jewish supremacism.

The reality is that intelligence research has nothing to do with racism and prescientific concepts of race. Figures like Hitler had narrative-driven, almost spiritual reasons for race hate coupled with the rejection of intelligence research when it failed to conform to their narratives.⁷ The common claim that some of America’s early 20th century anti-immigration activists relied on IQ research isn’t even supported. The reason is obvious, because as Terman once noted, the Chinese would embarrass any attempts to incorporate IQ into immigration policy.⁸

A more important question than these is what happens if we don’t recognize intelligence differences? We would clearly have to look elsewhere to explain the numerous real-life consequences of intelligence differences. Given their importance, the only imaginable place one could turn to may end up being something crazy, like systemic, unprovable, and apparently unalterable racism. The end result is blood libel and its consequences against those accused of being its perpetrators.

The things differences in intelligence explains really are incredible. They include economic mobility, the income gap, and even the sentencing gap. Bostrom is interested in genetic enhancement, and if we can enhance away these gaps, why shouldn’t we? We certainly won’t fix them if we reject them, since society is fixed and biology - genetics included - is mutable.

At the end of the day, Bostrom has shown that he believes in a better world that can only be reached with honesty and courage without prejudice. You should too.

1. In the case of dichotomous variables, the correlation is known as the Phi coefficient, the Mean Square Contingency coefficient, Matthew’s correlation coefficient, or the absolute value of Cramér’s V. τ, however, is a related, intuitively understandable coefficient that aligns with the logic behind the binomial effect size display.

2. These significantly differed.

3. For various reasons, people might be interested in the political activity of those who believe Whites are more intelligent than Blacks. The GSS has enough on this to render a simple conclusion: they care less about politics and they’re less involved with it. A few quick facts follow:

• These people are less likely to be in political clubs (r = -0.07, d = -0.140; subsequent effect sizes in d; mempolit)

• They are no less interested in politics (-0.002; polint and polint1) and they discuss it about as much as other people (-0.060; discpol), but they are less likely to express their political views online (-0.221; polinter) or try to change other people’s views (-0.100; chngeoth). Despite this, they believe about equally that it is important to be active in political associations (0.002; actassoc).

• They are more likely to believe political parties encourage people to become active in politics (0.100; polactve),

• but they are less likely to participate in political organizations or associations (-0.262; partpart), marginally less likely to volunteer for political activities in the last year (-0.080; volwkpol. But by grppol, 0.080), to donate to political activities (-0.140; polfunds), or to attend political meetings or rallies (-0.120; attrally).

• They are less likely to belong to a political party at all (-0.181; grpparty).

• They are less likely to be contacted by or appeared in the media to express their views (-0.221; usemedia).

• They are less likely to contact politicians or civil servants to express their views (-0.140; cntctgov).

• They are less likely to sign petitions (-0.100; signdpet) or participate in political product boycotts (-0.366; avoidbuy).

• They are less likely to “do something” if an unjust law came into effect (-0.140; actlaw).

For the most part, these people seem like a more apathetic demographic. They are also slightly less intelligent (-0.260, or -3.90 IQ points) as measured by the wordsum. The default assumption with respect to these people is that their beliefs will probably not achieve any ends in the political arena because, mutatis mutandis, they’re not even in it.

4. Other research also frequently supports this point. For example, Williams et al. (1999) found no relationship between the belief that Whites are more able than Blacks and support for government intervention to help them and then they found a positive and inconsistently significant relationship with support for affirmative action, quite contrary to the expectation that the belief is solely associated with things considered bad with respect to race. Zach Goldberg has also documented several instances in which the belief that Whites have more in-born ability than Blacks is associated with more acceptable opinions. A final related finding is that White liberals - who tend more than White conservatives to endorse racial equality and equity - tend to use simpler language around Blacks.

5. I want to particularly thank those who still responded to the survey when their chat’s moderators banned me right away.

6. Here are some previous surveys of researchers.

7. There are many amusing examples of contemporary commentary to the effect that Hitler was an opportunist with respect to his use of the concept of race. For example, this book begins with the lines “For instance, we had the great dictators, Hitler and Mussolini, openly saying that ‘race’ is a fraud but cynically using it to unite their gullible followers. At the same time, the two great democratic nations, Britain and America, while condemning Fascist and Nazi doctrines, and proclaiming equal justice for all were using ‘race’ as a fetish to keep their own citizens and subjects divided. White people were set against black peoples by the bogey of miscegenation.”

For more information regarding Nazi views on intelligence research, see Rindermann’s (2018) Cognitive Capitalism, section 3.3.2.

8. Thus embarrassing racists.