In the second part of our series on implicit bias, we explore the relationship between beliefs and behaviors. We also talk with psychologist Mahzarin Banaji about whether research on implicit bias tells us more about groups than it does about individuals.

All of us know what prejudice looks like.

We've seen news stories about swastikas spray-painted on synagogues, or nooses drawn on classroom walls to terrorize black students.

We have heard xenophobic speeches from politicians and watched in horror as ethnic groups around the world have exterminated their enemies.

In the late 1990s, Harvard psychologist Mazarin Banaji and her former PhD advisor, Tony Greenwald of the University of Washington, developed a test of hidden bias called the Implicit Association Test, or IAT.

Unlike the very public spectacle of a burning cross on someone's front lawn, the picture of bias painted by this test was rather subtle.

By measuring the speed of people's associations, the tests showed that large numbers of Americans found it easier to associate white faces with positive concepts than to associate black faces with positive concepts.

Many people were similarly quick to associate men with professional activities, and slow to associate women with such activities.

Lots and lots of Americans appear to have negative associations about the elderly, the overweight, and the disabled.

Crucially, large numbers of the people taking the tests didn't think of themselves as being prejudiced.

Many prided themselves on having egalitarian beliefs.

Their test results often came as a shock.

The feeling was a feeling of dread, I would say.

A feeling of Having had the rug pulled out from under you, you have to start from scratch to now rebuild a view of yourself that will forever be a different view of yourself.

In part 1 of the story, which I strongly recommend you listen to before you listen to this episode, we explored the origins of the psychological test.

Today on the show, we explore what it means that so many people have subtle biases when it comes to their mental associations.

Are these biases benign and just inside people's heads or do they cause people to act in biased ways?

Is there anything we can do to fight our biases?

The surprising connection between our biases and our behavior.

In George Orwell's dystopian novel 1984, there were lots of ways to get in trouble with a totalitarian state.

Protesting in the street was a quick way to get seized by the authorities, but you could also get in trouble for subtler things, like reading the wrong book or having the wrong opinions.

The eyes and ears of the state were everywhere, and subjects were expected to not just do the right things, but to think the right thoughts.

After Harvard University put the Implicit Association test on its website, you can find it at implicit.harvard.edu, interest in the test surged.

Many companies began mandating their employees take the test during diversity training exercises.

As we saw in part 1 of the story, early studies that Mazarin and others conducted suggested a connection between implicit biases and the behavior of individuals taking the test.

In 1 study she conducted with physician Alexander Green, physicians with higher bias scores on the IAT were less likely to prescribe clot-busting treatments to a black patient relative to a white patient.

But soon, other studies started to come out that showed no association between the 2.

People showing higher levels of bias on the test did not act in biased ways.

Critics of the test, including Phil Tetlock at the University of Pennsylvania, started to argue that the test was measuring the equivalent of Orwellian thought crimes, instead of judging people on their words and actions.

I think the IAT is grounded in a reductionist view of human nature.

I conducted this interview with Phil some years ago, in 2017, as we featured him on another episode of Hidden Brain.

It depicts people somewhat as association-driven automatons.

I'm not temperamentally all that comfortable with that form of reductionism, but I wouldn't reject the test on that basis.

I mean, it's a test that is enormously intuitively appealing.

I mean, I've never seen a psychological test take off the way the IAT has and has gripped the popular imagination the way it has, because it just seems on its surface to be measuring something like prejudice.

You've got the differential reaction times, right, between the black

and white stimuli, say, and that this just seems to be a bullseye.

And anybody who denies it is engaging in some kind of scholastic quibbling.

But there is the question of whether or not people who score as prejudiced on the IAT actually act in discriminatory ways toward other human beings in real-world situations.

And if they don't, if there is very close to 0 relationship between those 2 things, what exactly is the IAT measuring?

Meta-analyses, studies tracking a body of research, not just an individual experiment, are 1 way to tell if something is a fad or a fact.

If you have a number of studies linking a medication with positive patient outcomes, for example, you become more certain that the medication actually works.

Meta-analyses of the IAT test data found mixed results.

Some studies showed a connection between implicit bias and real-world discrimination, But plenty of others did not.

For many critics of the test, this was vindication that the test was useless and that the hype over the test was unfounded.

Challenges to the usefulness of Mazurin's research were published in peer-reviewed academic journals.

Soon, Mazurin and her colleagues decided to launch a comprehensive meta-analysis of their own.

I asked her if she found there was a correlation between implicit biases and real-world behavior.

The strength of the correlation is small.

When scientists measure whether 2 things are correlated to each other, they use a scale from 0 to 1.

Taller people tend to be heavier people, so height and weight are correlated.

It doesn't mean every tall person is going to weigh more than every short person.

All it says is that there is a relationship between the 2.

As height goes up, you're likely to weigh more.

But there are lots of things that are not correlated.

Taller people, for example, are not better at math.

The closer a correlation between 2 things is to 0, the less likely the 2 things are correlated.

And when it came to Mazarin's analysis of the IAT?

We published a paper that probably places the correlation at the smallest it has ever been reported to be, around

And that's because we decided to include every single study no matter how poorly it was conducted.

So we decided we will not throw any study out.

There are studies in there that really don't belong.

Somebody did a study looking at whether race bias would predict their degree of smoking.

There should be no correlation, but that study is in there, and it counts as a lack of correlation.

So we have gone overboard in making sure that no matter what the study, we will include it.

And so we show that even when you do that, the correlation is about 0.1.

Within that set, if you begin to use 5 criteria for what is a good study, that any scientist would agree.

You know, the correlation jumps up to 3 times that size.

I first read about Mazarin's work and the IAT around 20 years ago.

It's fair to say I initially thought it would show that individuals with implicit biases always act in biased ways.

I wrote about the test in my 2010 book, The Hidden Brain.

As mixed data started to emerge in the last decade, I found myself having to question my beliefs.

There was no doubt that large numbers of people around the world showed fast or slow associations on the tests.

And the results of the tests generally matched our intuitions about the nature of prejudice.

Across many countries, people more swiftly associated men rather than women with concepts related to science or leadership.

Groups that were in the majority often had negative biases about minorities.

This wasn't about a couple of studies and a few dozen subjects.

There were hundreds of studies and millions of test takers.

But if people's test results were only weakly correlated with real world behavior, what did these results mean?

Did it really make sense for companies around the world to mandate their employees take these tests?

Some were even using test results to figure out who should be involved with HR and hiring or rising to leadership positions.

As critics and supporters used the studies as ammunition for their pet beliefs, Mazarin and her colleagues kept doing research.

The volume of data rolling in meant that they could do something that most social scientists working in other areas could only dream about.

They could start to analyze relationships between test results and real-world behavior, not at the level of individuals or even at the level of companies, but at the level of cities, regions, and nations.

In 2009, they reported preliminary results on an unusual finding.

We had data from a few countries, not as many as we have today, so we are in the process of replicating that to see if with so much more data, because we did this a long time ago, in which We knew that there was a standardized test that was taken across countries in eighth grade, and we could get the data on the gender difference in performance on that mathematics test taken in eighth grade across many countries.

And so for each country, you compute the gender difference.

What is the gender difference in country A?

If boys score a lot higher than the girls, that's a big difference.

There might be countries where boys and girls score about the same.

So you have a lot of countries with a bunch of variation in gender.

And what you're looking to see is whether the countries that showed the largest difference in math performance between girls and boys are also the countries whose people are carrying in their heads a stronger association of boys with math rather than girls with math.

And we reported that there is a robust positive correlation.

Countries with higher gender bias are also the countries where the girls are underperforming compared to boys to a greater extent.

I want to take a moment to sit with this finding.

At the time she did this study, Mazarin didn't think much of it.

But what she was finding was that when you give people an implicit association test, measuring how quickly they associate concepts in mathematics with men rather than with women, they were regional differences, national differences, and test scores.

When you averaged implicit bias scores across entire countries, People in some countries were faster than others when it came to associating men with math and slower in associating women with math.

Now, if you turn to how 8th grade students were doing in the standardized math test, girls in countries with high implicit bias were doing worse than girls in countries with low implicit bias.

The interesting thing is that if you look not at a nation, but at an individual student or an individual school, you might see little to no correlation between implicit bias test results and student performance on the standardized mathematics test.

Or you might find it in 1 school, but not in the next.

It's only when you step back and looked at the big picture that you saw a robust correlation.

It was like 1 of those pixelated images or a painting that uses the style known as pointillism.

It's only when you step back you realize, oh, there's a picture here.

That paper, while I always thought that was an interesting result, I think I wasn't smart enough to realize that the reason we were getting a fairly substantial correlation is because we were wiping out lots of individual level error.

We were collapsing across many different people to come up with a much more stable score of what is going on in the larger system, in the larger environment in which people are sitting.

And when you take the average of that, the average of a whole bunch of people, you're likely to pick up the actual or true correlation in a much better way.

When someone gives you a test, you feel the test is saying something about you.

That is true, but also not completely true when it comes to the implicit association test.

Yes, at 1 level, the tests are telling you about something that is inside your head.

But the test might be telling you something much more important about the culture in which you are living.

When we come back, what happens when we look at the implicit association test not at the level of individuals, but at the level of cities, counties, and nations?

The implicit association test became very popular after psychologist Mazarin Banaji and her colleagues placed the test on the Harvard website.

Millions of people took the test, hoping it would give them a glimpse into their own minds.

Americans are in in this country at this particular point.

Is it because I can't come to say that I'm bad?

And is it just in our nature that there has to be an us and a them?

But as researchers evaluated whether test results revealed real-world behavior, they found mixed results.

Sometimes people who showed high bias on the implicit association test acted in ways that were biased, But at other times, they didn't.

When you looked at all the studies together, the correlation between individual implicit bias test results and individual real-world outcomes was small.

But the torrent of data from IAT test takers around the United States and around the world meant researchers could now start to analyze not just links between test results and individual behavior, but the correlations between average scores in an area and real-world outcomes.

Early on, Mazarin and her colleagues discovered a curious result.

The performance of boys and girls in an eighth-grade standardized math test appeared to be linked to average implicit bias scores in those Nations.

Countries where people were quicker to associate men with science showed a wider gap in test scores.

Girls did worse on the standardized test.

In time, other research along these lines started to emerge.

So Raj Chetty, for example, my colleague in economics, is interested in not only why upward mobility is so slow these days compared to what it was, why the American dream has vanished, but he's also interested in for whom is the American dream more likely and less likely.

And so he took our data and said, well, we can look county by county at IAT.

Instead, identify the county and take the average IAT of all the people in that county and give it just 1 score.

So you collapse across all the people in the county, and you come up with 1 IAT number, and you do that county by county by county.

You line up the counties from the most anti-black to the least anti-black counties.

And you look to see if that predicts upward mobility for black Americans.

In other words, he shows that the higher the race bias, average race bias in a county, the harder it is for black people in that county to be upwardly mobile.

Now that we've understood what the correlation is, I can just rattle off for you.

You know, now I think we're up to about 17 independent studies that have been published that show that higher race bias in a county will predict greater lethal use of force by police against black Americans.

The most recent study shows greater militarization of police departments in those counties, greater threats to maternal health and infant health in the counties that have greater bias, school disciplining differences between white and black kids that are greater in counties that have greater race bias, traffic stops, tickets, et cetera.

And these 17 studies that I'm just mentioning that look at average IAT scores by county or by state or by metropolitan region or by country, they're all averages by region.

They are just predicting up and down the spectrum in a way in which I would never have predicted, but is really exciting to see because these dependent variables are not simple little things.

It's not even how well you do on a math test.

This is about whether you will get disciplined in school and get kicked out.

This is about whether you will live as a baby.

is we would see a stronger correlation at this aggregate level.

So in other words, if you're analyzing my brain and saying, here's your implicit bias, and then you're evaluating me to see, do I hire a black person or a white person?

Why is it that we would see a lower correlation at me and an individual, but when you step out and look at the aggregate, you have a higher correlation.

Any individual score is going to vary based on lots of things jittering around in that moment.

And more importantly, whether you behave in a way that is biased or less biased is going to be multiply determined by little things in the local environment.

For example, my score on race bias may be quite high.

I may be quite anti-black, but it may be that in the moment in which you're testing my behavior, a smiling person appeared in front of me who wiped out my bias, and I responded positively to that person.

Little things like that in the environment can make the behavior move around and not allow the particular measure in which you're interested in to show itself.

So as soon as you aggregate it, for every person like me, somebody else's similar behavior will counter it.

And so all you are doing, the best way to understand it is that when we aggregate, we are removing individual-level noise in the data.

1 analogy to this idea comes from the realm of polling.

In the United States, Lots of polling is done by groups that have either a conservative or a liberal bias.

Unsurprisingly, polls that lean conservative are likely to predict conservative victories.

Liberal polls are likely to predict victories for Democrats.

But something interesting happens when you average out the polls.

The poll that leans too far right gets balanced out by the poll that leans left.

When you average polls, you are likely to get answers that are much more accurate than individual polls.

The same thing happens if you ask people to make estimates of something, say the size of the US economy.

Low estimates and high estimates cancel each other out when you average the answers, leaving you with a better approximation of the correct answer.

This phenomenon is sometimes known as the wisdom of the crowd, meaning the average answer across a group of people is often more accurate than individual answers.

The 2 views are, my view is that as we remove noise, we will see higher and higher levels of correlation.

In other words, as the tests get better, as studies are conducted more carefully, Mazarin is saying she expects the correlations to get better at the individual level, not just at the level of nations.

But she also cites a second possibility, that the IAT is really capturing a reflection in people's minds of something that is in the larger culture.

Somebody I admire greatly and agree with in many ways, and I'm not opposed to this, but this is Keith Payne.

He argues that these things don't operate at the level of the individual.

You will never, no matter how much error variance you remove, it will never get better at the individual level because we become of a certain place when we go into a certain area.

So he did this remarkable study that, if I can just describe really quickly, I will tell you about it.

Keith obtained a map that had been produced by Abraham Lincoln in the 1860s, in which Lincoln had his people plot county by county the proportion of enslaved to free people in every county in the southern states.

And he did this because, obviously, he was this smart guy, a scientist almost, because he thought, if I know that, if I know the proportion of enslaved people in a county, I will be able to make better military predictions about which counties are going to fall faster than other counties.

And the simple idea was, the greater the proportion of enslaved people, the harder they will fight and the more they will resist giving up slavery.

You can look at the map, you can see the counties and little numbers that tell us what the proportions are.

So Keith, in the 21st century, goes back to this map And he says, let's correlate these 2 things, IAT race bias in that county today, and the number of enslaved to free people proportions or ratios in 1860.

And lo and behold, the correlation is quite substantial and high, and he will say, exactly as we've been discussing, he will say, well, how can it be?

Keith Payne, a psychologist at the University of North Carolina, Chapel Hill, looked at implicit bias test results for people in the 21st century.

Why would these results be connected to policies that existed more than 150 years ago?

Everyone who lived in those counties in the 1860s is now very dead.

The proportions of enslaved to free people are not the people whose minds we're measuring.

And it's not even the case that the descendants of those very people live there today.

In the United States, enough migrations have happened that that is not the case.

So Keats' point of view, which is very interesting, is that your mind reflects what is right around you.

And that if you were somebody who lived in Seattle and Microsoft Corporation sent you to some place in Georgia and you arrived there, you become of that place.

If there are many Confederate statues in the town in which you live, your mind will move in the direction of more anti-black bias.

Your children will hear certain things in school, and they'll bring them to your home and as you talk about them you too will start to acquire that.

And these are the things that ultimately create these remarkable correlations, almost unbelievable, and what it tells us is just the long shadow of history and how psychology is able to pick up this incredibly long shadow of history.

That we can look at data from 1860 and we can predict today what that county's race bias is.

Or we can look at the race bias today and we can predict who those people were.

In some ways I think I'm hearing 3 different models.

1 says bias is produced by active animosity and hostility.

People who act in biased ways mean to be biased.

The second says, no, our minds are mirrors and when we go to different places, we are going to reflect what is out there But I think there's also a third model and I might call this the the hypertension model So if you were to measure my blood pressure right now and measure my blood pressure 2 hours from now or 2 days from now, it's going to fluctuate because blood pressure is not super stable.

It depends on what's going on, what's happening to me physically, my mental state.

But if you find that I do have high blood pressure, it doesn't necessarily tell you I'm going to have a heart attack next week or I'm going to have a stroke next month.

So in other words it's a useful measure but at an individual level it's a somewhat crude measure of determining short-term risk.

But if you were to step back and say what's the average hypertension of all the people in California or what's the average hypertension of all the people living in New York and let's say the average hypertension in California was significantly higher than the average hypertension in New York, you could very confidently say you will have many more heart attacks and strokes in California than in New York, even though you can't predict which individuals are going to be affected.

You can say something meaningful about the group, even if you can't be very precise about individuals.

I mean, the same thing goes for smokers and non-smokers.

I might not be able to tell you which individuals are going to develop cancer in any given week, but I know the group of smokers is going to have more cancers than the group of non-smokers.

So, both the mirror model and the hypertension model suggest that if you want to understand how unconscious biases caused biased actions, you need to look beyond the individual mind and look at larger systems and structures.

And I like the example of hypertension for many reasons, 1 of which is that the way we measure hypertension shows that the machine is not terribly reliable for all the reasons that you said.

If my arm is up or down, if I've just eaten, if I've walked, it will vary and sometimes substantially vary.

See all Hidden Brain transcripts on Applepodcasts

Revealing Your Unconscious: Part 2