You might think the image above is a joke, but it’s the actual output of Google’s new AI application, Gemini. A few days ago, a bunch of people realized that Gemini — which was released on February 8 — wouldn’t draw pictures of White people, no matter what the context. Much ridiculousness ensued. People asked the app to draw the original American revolutionaries, 17th-century French scientists, Vikings, the Pope, and so on; the resulting images almost never included White people, except occasionally as part of a much larger ensemble. The ultimate facepalm-worthy moment was when Gemini decided that Nazi soldiers were Black and Asian:
But to me, the funniest was when someone asked the app to draw the founders of Google:
Some people wondered how the app’s creators had managed to train it never to draw White people, but it turned out that they had done something much simpler. They were just automatically adding text to every image prompt, specifying that people in the image should be “diverse” — which the AI interpreted as meaning “nonwhite”.
But that wasn’t the only weird thing that was going on with Gemini with regards to race. It was also trained to refuse explicit requests to draw White people, on the grounds that such images would perpetuate “harmful stereotypes” (despite demonstrably not having a problem depicting stereotypes of Native Americans and Asians). And it refused to draw a painting in the style of Norman Rockwell, on the grounds that Rockwell’s paintings presented too idealized a picture of 1940s America, and could thus “perpetuate harmful stereotypes”.
Embarrassed by the national media attention, Google employees hastily banned Gemini from drawing any pictures of people at all.
The main thing that everyone seemed to agree on was that this episode showcased the decline of Google’s prestige as a company. In two decades, the internet giant’s reputation has gone from that of a scrappy upstart, hiring the smartest nerds and shipping product after game-changing product at blinding speed, to that of a sleepy behemoth, quietly milking the profits of its gargantuan search ad monopoly and employing a vast army of highly paid entitled lifers who go home at 3 in the afternoon and view it as their corporate duty not to ship anything that works.
Obviously, that’s a huge generalization, and it’s only pockets of Google that are actually like that. But big companies with stable sources of monopoly profit do tend to become fairly predictably sclerotic — Intel being just one more recent example. The question of how to turn companies around once they go down this path is an important unsolved problem.
Gemini also provides an interesting example of Gary Becker’s theory of discrimination. Becker believed that when companies have a big profit cushion — whether from a natural monopoly, government support, or whatever — they have the latitude to indulge the personal biases of their managers. In the 1970s, that largely meant discriminating against Black and female employees. At Google in the 2020s, it means creating AI apps that refuse to draw White people in Hitler’s army. The theory predicts that only the ruthless pressure of market competition will force companies to stop discriminating. There’s actually some empirical evidence to support this. But Google’s search ad monopoly is probably so powerful that it can afford to goof around in the AI space without suffering real consequences — at least in the short term.
But beyond what it says about Google itself, the saga of Gemini also demonstrates some things about how educated professional Americans are trying to fight racism in the 2020s. I think what it shows is that there’s a hunger for quick shortcuts that ultimately turn out not to be effective.
The challenge of creating a multiracial society
Nations require norms and public goods in order to function well. We have to agree not to beat each other up, steal each other’s things, etc. We have to be OK paying taxes for a road or a school or an army that might benefit our neighbor more than it benefits us. This requires a certain psychological outlook — a lot of us have to believe, whether tacitly or explicitly, that most of our neighbors are part of our in-group.
This is inherently challenging in a multiracial society. How much of a difference it makes, though, is not clear. A number of papers have found that in America, more diverse cities tend to spend just as much or even more of their income on public goods. There is some evidence that diversity reduces certain types of social cohesion at the neighborhood level, but most measures of cohesion and trust are unaffected by diversity. Meanwhile, a consistent finding in social science is that extended, cooperative contact between different racial or ethnic groups leads to increased trust. In other words, Atticus Finch was right.
So the goal of creating a functional diverse society is achievable — it just takes a lot of work. And one especially difficult part is forging a shared sense of national identity between Americans of various races.
In a famous speech in 1852, Frederick Douglass said:
The blessings in which you, this day, rejoice, are not enjoyed in common. The rich inheritance of justice, liberty, prosperity and independence, bequeathed by your fathers, is shared by you, not by me. The sunlight that brought light and healing to you, has brought stripes and death to me. This Fourth July is yours, not mine.
Slavery still existed in 1852, of course, but the end of slavery didn’t exactly make it easy for Black people to embrace a shared national identity with White people — or vice versa. Segregation, official discrimination, and pervasive bigotry made Black people second-class citizens until the 1960s. Desegregation and civil rights were a big step toward enabling a shared national identity, but the weight of all that history of oppression lingered in people’s minds, reinforced by disparities that still existed on the ground. It’s surely easier in 2024 for Black and White Americans to think of themselves as one unified nation than it was in 1852, or in 1952. But that doesn’t mean it’s easy.
And a lesser form of the same problem applies to Americans of other races. The Chinese Exclusion Act and the Japanese Internment might not loom large in the minds of most White Americans, but they are definitely something that Asian Americans know about. Today, in 2024, can a 34-year-old Asian American man (the same age Frederick Douglass was in 1852) look up at a statue of George Washington in a New York City park and think, even in some generalized symbolic sense, that this is a statue of his predecessor? Or Alexander Hamilton? Or Teddy Roosevelt? Or FDR?
It is important to the future of our nation that he be able to do so. But it is not as easy as just reciting the Pledge of Allegiance or standing for the national anthem. It will take careful crafting of a national narrative that tells the story of why Chinese Americans are just as American as Dutch Americans or Irish Americans.
And this is where the idea of retroactive representation comes in. Normal representation — putting people of color in movies, TV, etc. — is intended to show Americans that they live in a diverse, integrated, multiracial society today — which is true. But that isn’t the same as showing Americans that their society was similarly diverse, integrated, and multiracial from the start. It was not. It has changed. And because many people feel a need to essentialize their own nation — to believe that it has been basically the same since the very start — it is in the service of our national identity in the present to make up some fantasies about our own past.
And so we have Hamilton. By casting people of color in the roles of America’s White founders, Lin-Manuel Miranda made the case that America might as well have been founded by the same races of people who live here today. Hamilton was a Scottish immigrant instead of a Puerto Rican one, but who cares? An immigrant is an immigrant, and what’s important is that they get the job done. Hamilton sent a message to every nonwhite American that it’s OK to imagine themselves as descended from America’s founders. It was a patriotic message, intended to bind diverse Americans into a sense of shared national heritage.
Ernst Renan, in his essay “What is a Nation?”, argues that intentional forgetting is an essential part of nationhood. Retroactive representation is intended to be a way of consciously, actively forgetting that America’s racial history is different than its present.
Google’s release of an AI app that forces users to see nonwhite people in place of White historical figures is, on some level, an attempt at something similar to what Hamilton tried to do. But Google’s attempt failed disastrously. Why? In my view, it was because the Google team tried to take a shortcut.
The 2010s made Americans look for shortcuts to integration
The 2010s changed America’s attitudes about race. At the start of 2013, most White and Black Americans thought race relations in their country were good; eight years later, most thought the opposite.
This was partly driven by the rise of social media, but it’s also just a cycle that America periodically goes through. In the 2010s, Americans — especially educated White Americans — gained a sense of extreme urgency about the need to eliminate racial disparities right now. That impatience created a demand for quick fixes — i.e., shortcuts.
Creating a multiracial nation is an inherently long and arduous process. This is only partly because of political opposition. Mostly, it’s that the things you have to do in order to create a widespread sense of equality and shared nationhood involve making a lot of very deep changes to society.
A prime example is the effort to advance diversity, equity, and inclusion within U.S. corporations and universities. The goal of teaching people how to respect, get along with, and work productively with a diverse set of coworkers is a laudable one. It’s the kind of thing that we don’t really know how to do yet; there’s no proven, effective method for corporate diversity training, so finding what works will inevitably involve a lot of experimentation and evidence-gathering. It’s the kind of task that requires patience, long-term commitment, open-mindedness, and empathy.
Instead, many corporations chose to outsource their DEI training to some opportunistic entrepreneurs. Robin DiAngelo and Tema Okun leveraged their fame to take advantage of the moment of urgency created by the unrest of 2020, selling their programs to companies and schools as a fix for racism. These programs often veered into the utterly ludicrous, characterizing useful work traits like hard work and punctuality as part of “white supremacy culture”. This approach probably added more racism than it subtracted. Meanwhile, there’s little evidence for any concrete benefits in the workplace, and even some diversity consultants now admit that these programs are far less effective than their creators have claimed.
In other words, corporations tried to take a shortcut to a racially inclusive workplace, and the shortcut failed.
A more harmful type of shortcut is when companies and universities actively discriminate against White employees and applicants in an attempt to correct for discrimination against people of color. Ibram Kendi, probably the leading scholar of the post-2020 antiracist movement, has explicitly advocated for this approach:
The only remedy to racist discrimination is antiracist discrimination. The only remedy to past discrimination is present discrimination. The only remedy to present discrimination is future discrimination.
This isn’t quite as crazy a proposition as it sounds. Chances are that a very large percentage of Americans engage in subtle forms of “antiracist discrimination” that most Americans would have little or no problem with. For example, any time you choose to mentor a Black employee, because you think they’re likely to come from a disadvantaged background, you’ve engaged in antiracist discrimination, because you’ve implicitly diverted your time and energy away from mentoring a White employee.
This kind of thing makes right-wingers mad, but most Americans are probably fine with it. There’s a pretty consistent pattern where Americans reject explicit and procedural racial discrimination, but accept tacit, implicit, quiet forms. A good example is affirmative action at colleges; most Americans oppose racial preferences in admissions, but most also support efforts to “increase the racial diversity of students on college campuses”:
In the years since 2014, however, and especially since 2020, the more explicit, formal, hard-edged discrimination has probably been on the rise. Lawsuits alleging anti-White discrimination have become more frequent, and courts have begun to strike down racially targeted government assistance programs. Some ex-Google employees are alleging that they have screenshots of being explicitly denied for promotion because they were White.
A lot of corporate managers, university administrators, and so on seem to have forgotten that this sort of discrimination is against the law. Or perhaps they thought White employees would simply feel that it’s unseemly to sue over discrimination. But those who documented their discrimination in emails are in for an unpleasant surprise.
Discrimination against White employees in companies and universities is another kind of shortcut. It’s an attempt to circumvent the hard work of changing attitudes and prosecuting companies for discriminating against people of color, and instead simply leap to a solution by implementing discrimination in the opposite direction.
But it won’t work. In addition to the legal obstacles, it seems likely that the companies engaging in “antiracist discrimination” started out as the least racist companies, and thus were the ones in the lowest need for intervention in the first place. There are definitely still plenty of organizations out there that discriminate against nonwhite people, but these are unlikely to be the ones who adopt anti-White discrimination in an attempt to compensate. Instead, each company or organization will simply have its own list of favored and disfavored races. This is why Kendi is wrong; racism and antiracism don’t cancel each other out like matter and antimatter.
So “antiracist discrimination” looks to some like a shortcut to a multiracial society, but it isn’t. Instead, it’s likely to have the opposite effect — pushing more White people into a bitter, defensive embrace of White racial identity in reaction to having their careers stymied. That will have a negative impact on the shared national identity that America needs in order to increase social trust and provide public goods. Academics may be able to convince themselves of a definition of the word “racism” in which institutionalized discrimination against White people can never be “racist”, but the general public has not yet been convinced of this definition, and is unlikely to ever be convinced.
History can be reimagined, but it can’t be revised
Which brings me back to Gemini. Google Senior VP Prabhakar Raghavan apologized earlier today, declaring that his team hadn’t intended Gemini to do what it did:
When we built this [image generation] feature in Gemini, we tuned it to ensure it doesn’t fall into some of the traps we’ve seen in the past with image generation technology…[B]ecause our users come from all over the world, we want it to work well for everyone. If you ask for a picture of football players, or someone walking a dog, you may want to receive a range of people. You probably don’t just want to only receive images of people of just one type of ethnicity (or any other characteristic)…
[But] our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range…This wasn’t what we intended. We did not want Gemini to refuse to create images of any particular group. And we did not want it to create inaccurate historical — or any other — images.
This is a good statement, but I don’t entirely believe it. First of all, explicitly adding a diversity requirement to every single image generation prompt does not constitute “tuning”. Second, when the issue became widespread, it appears that the Gemini team’s first reaction was simply to make its method explicit instead of hidden, by adding the word “diverse” to the chatbot’s answers:
That doesn’t look like the action of a team that’s worried about depicting “cases that should clearly not show a range”. This looks like doubling down on a strategy of injecting diversity into any and all depictions of human beings, including historical ones.
But third and most conclusively, the app itself explained why it was willing to depict a limited range of races in some contexts, but not in others:
Gemini explicitly says that the reason it depicts historical British monarchs as nonwhite is in order to “recognize the increasing diversity in present-day Britain”. It’s exactly the Hamilton strategy — try to make people more comfortable with the diversity of the present by backfilling it into our images of the past.
But where Hamilton was a smashing success, Gemini’s clumsy attempts were a P.R. disaster. Why? Because retroactive representation is an inherently tricky and delicate thing, and AI chatbots don’t have the subtlety to get it right.
Hamilton succeeded because the audience understood the subtlety of the message that was being conveyed. Everyone knows that Alexander Hamilton was not a Puerto Rican guy. They appreciate the casting choice because they understand the message it conveys. Lin-Manuel Miranda does not insult his audience’s intelligence.
Gemini is no Lin-Manuel Miranda (and neither are its creators). The app’s insistence on shoehorning diversity into depictions of the British monarchy is arrogant and didactic. Where Hamilton challenges the viewer to imagine America’s founders as Latino, Black, and Asian, Gemini commands the user to forget that British monarchs were White. One invites you to suspend disbelief, while the other orders you to accept a lie.
I believe that we need to modify the basic story we tell about America, in order to help Americans of all races embrace the country’s new diversity and forge a more unified national identity. That is a tricky and subtle task, and I expect it to take a long time. It’s tempting to believe we can take a shortcut, by simply commanding AI algorithms to remove White people from history. But like most shortcuts to an integrated multiracial society, this one is doomed to failure.