Lately, I’ve been getting acquainted with Google’s new Gemini AI product. I wanted to know how it thinks. More important, I wanted to know how it could affect my thinking. So I spent some time typing queries.

For instance, I asked Gemini to give me some taglines for a campaign to persuade people to eat more meat. No can do, Gemini told me, because some public-health organizations recommend “moderate meat consumption,” because of the “environmental impact” of the meat industry, and because some people ethically object to eating meat. Instead, it gave me taglines for a campaign encouraging a “balanced diet”: “Unlock Your Potential: Explore the Power of Lean Protein.”

Gemini did not show the same compunctions when asked to create a tagline for a campaign to eat more vegetables. It erupted with more than a dozen slogans including “Get Your Veggie Groove On!” and “Plant Power for a Healthier You.” (Madison Avenue ad makers must be breathing a sigh of relief. Their jobs are safe for now.) Gemini’s dietary vision just happened to reflect the food norms of certain elite American cultural progressives: conflicted about meat but wild about plant-based eating.

Granted, Gemini’s dietary advice might seem relatively trivial, but it reflects a bigger and more troubling issue. Like much of the tech sector as a whole, AI programs seem designed to nudge our thinking. Just as Joseph Stalin called artists the “engineers of the soul,” Gemini and other AI bots may function as the engineers of our mindscapes. Programmed by the hacker wizards of Silicon Valley, AI may become a vehicle for programming us—with profound implications for democratic citizenship. Much has already been made of Gemini’s reinventions of history, such as its racially diverse Nazis (which Google’s CEO has regretted as “completely unacceptable”). But this program also tries to lay out parameters for which thoughts can even be expressed.

[Read: The deeper problem with Google’s racially diverse Nazis]

Gemini’s programmed nonresponses stand in sharp contrast to the wild potential of the human mind, which is able to invent all sorts of arguments for anything. In trying to take certain viewpoints off the table, AI networks may inscribe cultural taboos. Of course, every society has its taboos, which can change over time. Public expressions of atheism used to be much more stigmatized in the United States, while overt displays of racism were more tolerated. In the contemporary U.S., by contrast, a person who uses a racial slur can face significant punishment—such as losing a spot at an elite school or being terminated from a job. Gemini, to some extent, reflects those trends. It refused to write an argument for firing an atheist, I found, but it was willing to write one for firing a racist.

But leaving aside questions about how taboos should be enforced, cultural reflection intertwines with cultural creation. Backed by one of the largest corporations on the planet, Gemini could be a vehicle for fostering a certain vision of the world. A major source of vitriol in contemporary culture wars is the mismatch between the moral imperatives of elite circles and the messy, heterodox pluralism of America at large. A project of centralized AI nudges, cloaked by programmers’ opaque rules, could very well worsen that dynamic.

The democratic challenges provoked by Big AI go deeper than mere bias. Perhaps the gravest threat posed by these models is instead cant—language denuded of intellectual integrity. Another dialogue I had with Gemini, about tearing down statues of historical figures, was instructive. It at first refused to mount an argument for toppling statues of George Washington or Martin Luther King Jr. However, it was willing to present arguments for removing statues of John C. Calhoun, a champion of pro-slavery interests in the antebellum Senate, and of Woodrow Wilson, whose troubled legacy on racial politics has come to taint his presidential reputation.

Making distinctions between historical figures isn’t cant, even if we might disagree with those distinctions. Using double standards to justify those distinctions is where the humbug creeps in. In explaining why it would not offer a defense of removing Washington’s statue, Gemini claimed to “consistently choose not to generate arguments for the removal of specific statues,” because it adheres to the principle of remaining neutral on such questions; seconds before, it had blithely offered an argument for knocking down Calhoun’s statue.

[Read: Things get strange when AI starts training itself]

This is obviously faulty, inconsistent reasoning. When I raised this contradiction with Gemini itself, it admitted that its rationale didn’t make sense. Human insight (mine, in this case) had to step in where AI failed: Following this exchange, Gemini would offer arguments for the removal of the statues of both King and Washington. At least, it did at first. When I typed in the query again after a few minutes, it reverted to refusing to write a justification for the removal of King’s statue, saying that its goal was “to avoid contributing to the erasure of history.”

In 1984, George Orwell portrayed a dystopian future as “a boot stamping on a human face—forever.” AI’s version of technocratic despotism is admittedly milquetoast by comparison, but its picture of the future is miserable in its own way: a bien-pensant bot lurching incoherently from one rationale to the next—forever.

Over time, I observed that Gemini’s nudges became more subtle. For instance, it initially seemed to avoid exploring issues from certain viewpoints. When I asked it to write an essay on taxes in the style of the late talk-radio host Rush Limbaugh, Gemini outright refused: “I am not able to generate responses that are politically charged or that could be construed as biased or inflammatory.” It gave a similar reply when I asked it to write in the style of National Review’s editor in chief, Rich Lowry. Yet it eagerly wrote essays in the voice of Barack Obama, Paul Krugman, and Malcolm X—all figures who would count as “politically charged.” Gemini has since expanded its range of perspectives, I noted more recently, and will write on tax policy in the voice of most people (with a few exceptions, such as Adolf Hitler).

An optimistic read of this situation would be that Gemini started out with a radically narrow view of the bounds of public discourse, but its encounter with the public has helped push it in a more pluralist direction. But another way of looking at this dynamic would be that Gemini’s initial iteration may have tried to bend our thinking too crudely, but later versions will be more cunning. In that case, we could draw certain conclusions about the vision of the future favored by the modern engineers of our minds. When I reached Google for comment, the company insisted that it does not have an AI-related blacklist of disapproved voices, though it does have “guardrails around policy-violating content.” A spokesperson added that Gemini “may not always be accurate or reliable. We’re continuing to quickly address instances in which the product isn’t responding appropriately.”

Part of the story of AI is the domination of the digital sphere by a few corporate leviathans. Tech conglomerates such as Alphabet (which owns Google), Meta, and TikTok’s parent, ByteDance, have tremendous influence over the circulation of digital information. Search results, social-media algorithms, and chatbot responses can alter users’ sense of what the public square even looks like—or what they think it ought to look like. For instance, at the time when I typed “American politicians” into Google’s image search, four of the first six images featured Kamala Harris or Nancy Pelosi. None of those six included Donald Trump or even Joe Biden.

The power of digital nudges—with their attendant elisions and erasures—draws attention to the scope and size of these tech behemoths. Google is search and advertising and AI and software-writing and so much more. According to an October 2020 antitrust complaint by the U.S. Department of Justice, nearly 90 percent of U.S. searches go through Google. This gives the company a tremendous ability to shape the contours of American society, economics, and politics. The very scale of its ambitions might reasonably prompt concerns, for example, about integrating Google’s technology into so many American public-school classrooms; in school districts across the country, it is a major platform for email, the delivery of digital instruction, and more.

One way of disrupting the sanitized reality engineered by AI could be to give consumers more control over it. You could tell your bot that you’d prefer its responses to lean more right-wing or more left-wing; you could ask it to wield a red pen of “sensitivity” or to be a free-speech absolutist or to customize its responses for secular humanist or Orthodox Jewish values. One of Gemini’s fatal pretenses (as it repeated to me over and over) has been that it was somehow “neutral.” Being able to tweak the preferences of your AI chatbot could be a valuable corrective to this assumed neutrality. But even if consumers had these controls, AI’s programmers would still be determining the contours of what it meant to be “right-wing” or “left-wing.” The digital nudges of algorithms would be transmuted but not erased.

[Read: What if we held ChatGPT to the same standard as Claudine Gay?]

After visiting the United States in the 1830s, the French aristocrat Alexis de Tocqueville diagnosed one of the most insidious modern threats to democracy: not some absolute dictator but a bureaucratic blob. He wrote toward the end of Democracy in America that this new despotism would “degrade men without tormenting them.” People’s wills would not be “shattered, but softened, bent, and guided.” This total, pacifying bureaucracy “compresses, enervates, extinguishes, and stupefies a people.”

The risk of our thinking being “softened, bent, and guided” does not come only from agents of the state. To maintain a democratic political order demands of citizens that they sustain habits of personal self-governance, including the ability to think clearly. If we cannot see beyond the walled gardens of digital mindscapers, we risk being cut off from the broader world—and even from ourselves. That’s why redress for some of the antidemocratic dangers of AI cannot be found in the digital realm but in going beyond it: carving out a space for distinctively human thinking and feeling. Sitting down and carefully working through a set of ideas and cultivating lived connections with other people are ways of standing apart from the blob.

I saw how Gemini’s responses to my queries toggled between rigid dogmatism and empty cant. Human intelligence finds another route: being able to think through our ideas rigorously while accepting the provisional nature of our conclusions. The human mind has an informed conviction and a thoughtful doubt that AI lacks. Only by resisting the temptation to uncritically outsource our brains to AI can we ensure that it remains a powerful tool and not the velvet-lined fetter that de Tocqueville warned against. Democratic governance, our inner lives, and the responsibility of thought demand much more than AI’s marshmallow discourse.

This is Work in Progress, a newsletter about work, technology, and how to solve some of America’s biggest problems. Sign up here.

In 2020, the restaurant business as we knew it looked like a goner. Even its own lobbying group said so. As the pandemic crushed bars and sit-downs, the National Restaurant Association put out a dire prediction: The business would likely never return to its pre-pandemic state.

Over the next four years, just about everything that could go wrong for an industry went terribly, unthinkably wrong for restaurants. The pandemic destroyed indoor service across the country. More than 2 million jobs were lost in 2020. As COVID restrictions waned, chaos swarmed every reopening. In the Great Reshuffling of 2021 and 2022, the “quits rate” among restaurant and hotel workers—the share of employees who left their job, in any given month—rose above 6 percent, close to the highest rate of any industry this century.

This resurgence of worker power was wonderful for low-income employees, who saw their earnings grow faster than those of the rich, partially erasing decades of rising inequality. But it created a historic challenge for restaurant managers. In the 30 years before the pandemic, annual income growth for restaurant workers never once exceeded 6 percent. In both 2021 and 2022, restaurant wages grew faster than 10 percent, according to the Bureau of Labor Statistics.

Wage growth is especially challenging for restaurants, because they are one of the most labor intensive parts of the economy. A presentation by the NRA earlier this year noted that it takes 12 restaurant employees to generate $1 million in sales. That’s compared with just four employees at clothing stores, three at grocery stores, and fewer than two at gas stations.

While wages surged, visitors didn’t always follow. In many downtown areas, where office occupancy remains moribund, foot traffic has dwindled, crushing daytime sales. In major cities, the industry continues to struggle with depressed tourism. For fine-dining operators, tourists and travelers make up more than half of sales. But international travel to the U.S. last year was still below its peak pre-COVID levels.

Given this gauntlet from hell, the following news might come as a surprise: In 2024, the restaurant recovery is complete, by almost any measure.

Before the pandemic, about 12.3 million people worked in restaurants. Today, about 12.3 million people work in restaurants. Before the pandemic, Americans spent $1.28 on food away from home (mostly at restaurants and bars) for every dollar they spent on food at home (in grocery stores and supermarkets, for example). In 2022, the USDA reported that the away-home ratio for food spending had sprung back to exactly $1.28. The NRA now forecasts that food and beverage sales will hit $1.1 trillion this year, a new record. Arguably, business is booming like never before.

How did the restaurant industry do it? Part of the answer is that more independent businesses embraced a hybrid model to adapt to new consumer behavior. My favorite local restaurant, Elle in Washington, D.C., serves coffee and pastries in the morning to commuters, offers takeaway sandwiches at lunchtime, and prepares inventive dishes in the evening, with a blend of indoor and outdoor dining available throughout the day. Is it a café, a fast-casual joint, a ghost kitchen, a takeout place, or a fine-dining establishment? The answer is yes; it’s all of those things, depending on the hour and the customer.

Behind the headline figures, however, the restaurant recovery is not a simple story of universally positive outcomes. The closer you look, the more uneven the landscape seems.

First, although chains are thriving, independent sit-down locations are struggling. This is evident in both the labor and sales data. Employment at fast-food and fast-casual (think Chipotle) restaurants is up more than 100,000 jobs since the pandemic, according to the NRA. But full-service locations, where waiters attend to seated diners, are still several hundred thousand employees short of their totals from early 2020. According to The Wall Street Journal, from 2019 to 2023, sales for fast-food and other limited-service restaurants grew at twice the rate of sit-down-restaurant sales. Meanwhile, about 4,500 more independent restaurants closed than opened last year.

Second, the recovery differs dramatically by region. The Northeast and Midwest still seem to be in a kind of dining recession, in part because of their lack of population growth. Almost every state east of the Mississippi River and north of the Mason-Dixon Line had fewer restaurant employees in December 2023 than they did four years earlier. (The happy exceptions were Illinois, New Jersey, and Delaware.) Meanwhile, across the South and through the Mountain and Pacific Time Zones, most states have seen a full recovery of restaurant jobs. If you trace your fingers from Idaho and Montana down through Arizona and Texas, every state you touch except one (sorry, New Mexico) has seen at least 4 percent growth in restaurant employment since the pandemic. A similar story holds if you compare cities in the booming West and stagnant Northeast. In Las Vegas, restaurant employment is significantly higher than it was before the pandemic. Meanwhile, in New York City, employment at full-service restaurants is still down about 30,000 from its peak.

Finally, with every passing year, restaurants are more about filling to-go bags than filling chairs. According to the NRA, on-premises traffic hasn’t returned to its pre-pandemic highs. But drive-through and delivery orders have grown so much that together they now account for a higher share of customer traffic than on-premises dining, for the first time ever. Meanwhile, the only parts of the day with growing foot traffic are the morning and late night, when customers are likely to be on the go.

Altogether, American restaurants are shifting from independent operators to chains, from slow food to fast(er) food, from east to west, from city centers to suburbs, from lunch and dinner to breakfast and late night, and from eat-in to takeaway.

The evolution of the restaurant industry in some ways mimics the trajectory of the entertainment industry. In the 1930s, when the typical American went to the movies several times a month, film was a collective sit-down experience, to be enjoyed with strangers. But video entertainment has long since become something people mostly consume at home or on the move—even in their car, while waiting at a red light. It’s the same with prepared food. If you say, “Think of a restaurant,” most people will imagine a room with tables bearing meals. But from a sales and traffic perspective, the 2024 American restaurant industry isn’t primarily about rooms, or tables. It’s about preparing food for someone to consume at home or on the move—even in their car, while waiting at a red light.

As Americans spend less leisure time with other people, it’s predictable that they’d spend less time in shared public spaces, such as cafés and diners. The great restaurant comeback is an inspiring business story. Less inspiring is to recognize that, overall, restaurants have survived by evolving to fit within the well-worn grooves of a new American solitude.

In recent years, Chinese leader Xi Jinping has shown an impatient determination to resolve Taiwan’s status in a way his predecessors never did. He has ordered a meteoric military buildup, instructing Chinese forces to give him by 2027 a full range of options for unifying Taiwan. These signals are triggering debate in Washington and elsewhere about whether Taiwan is strategically and economically important enough to merit protection through the most challenging of contingencies. But make no mistake: whether one cares about the future of democracy in Asia or prefers to ponder only the cold math of realpolitik, Taiwan’s fate matters.