From two former White House officials who shaped America’s AI strategy and now work with leading AI companies, an urgent case for why democracies must achieve preeminence in AI—and a bracing warning about what victory will require.
Humanity may soon create AI systems that outperform humans at virtually every cognitive task. Whoever wields that technology will gain a power unlike any in history. In The Bitter Struggle, Ben Buchanan, the first White House Special Advisor for Artificial Intelligence and currently an advisor to Anthropic, and Tantum Collins, who served on the National Security Council after five years at Google DeepMind, argue that the United States and its democratic allies must not lose this contest. But how democracies win matters most of all.
Buchanan and Collins take readers into the briefing rooms where AI CEOs privately warned about their own creations, into the Situation Room where Cabinet Secretaries debated how to choke off China’s access to AI technology, and into the Oval Office, where they helped shape President Biden's decisions. From semiconductor factories in Taiwan to city-sized data centers in America, they show why a million chips may matter more than a million troops, why AI progress has been so stunning and yet so predictable, and why the next few years may decide the next few decades. Yet inventing more powerful AI is only the beginning. Democracies must also deploy AI’s capabilities without surrendering their own values—and so far, they are failing.
The deepest challenge is not the contest with autocracy but the conflict AI will force within democracies themselves. Every choice pits security against liberty, speed against caution, power against principle. This is the bitter struggle. Democracies have every advantage they need to prevail—except, perhaps, the conviction to act.
Ben Buchanan was the first White House Special Advisor for AI, leading the development and execution of the Biden administration’s AI policy. He is the author of four books and is currently a professor at Johns Hopkins University and an advisor to Anthropic.
When Technical Knowledge Becomes Political Authority, Who Gets to Say No? “The Bitter Struggle” makes an unusually informed case for preparing for superintelligence while exposing the democratic limits of the experts empowered to prepare on everyone's behalf. By Demetris Papadimitropoulos | November 2nd, 2026
During a January 2023 White House demonstration of GPT-4, Sam Altman tells national security adviser Jake Sullivan that artificial intelligence will be “unimaginable” within five years. Sullivan asks what the adjective means. A fine way to keep excitement from writing policy. In “The Bitter Struggle: Superintelligence, Superpowers, and the Fate of the World,” Ben Buchanan and Tantum Collins know what an alarming prediction looks like once it reaches a government desk. Altman has shown GPT-4 reading images and distilling Vladimir Putin’s speeches; now his prediction must survive a different test. The room contains executives and appointed officials, not the voters who will live with the results. Soon the predictions will find their way into chip-export thresholds and rules limiting governmental surveillance.
Expertise opens doors for these authors that most citizens will never enter. Buchanan led AI policy in the Biden White House; Collins served on the National Security Council after working at Google DeepMind. They organize the argument into three struggles: democracies must invent the most capable AI, convert that lead into strategic advantage, and preserve their values while wielding it. The last struggle begins where the first two become dangerous. A state that protects its citizens with superior AI may use the same capacity to watch and discipline them. Winning the race, they know, could leave the victors with an apparatus their own citizens cannot restrain.
More revealing than the prediction is how Altman delivers it. He speaks “casually,” the authors say, while Sullivan responds with unusual trepidation. Five years sounds like a product horizon in one chair and a frighteningly short national-security deadline in the other. The authors liken the new connection between Silicon Valley and Washington to two halves of the transcontinental railroad meeting. Two tracks can meet without their passengers agreeing on a destination. The state initially learns about GPT-4 through what the authors describe as a courtesy heads-up from one startup.
Even the Wright brothers’ fifty-seven-second flight failed to impress Associated Press representative Frank Tunison; fifty-seven minutes, he reportedly said, might make news. Buchanan and Collins turn that judgment into an introduction to Richard Sutton’s “bitter lesson”: general learning methods powered by computation have repeatedly beaten painstaking attempts to encode human expertise. Their September 2022 encounter with an unreleased Anthropic chatbot puts a price on hesitation. They recognize how quickly a conversational interface might become popular, while the company holds back for safety testing; OpenAI launches ChatGPT soon afterward. Anthropic lost the first public spectacle. Whether its hesitation protected anyone is a question the book cannot settle, but hesitation itself looks different when somebody else receives the applause.
The semiconductor chapter arrives at a chamber full of molten tin. Extreme ultraviolet lithography requires a vacuum, droplets of molten tin struck by lasers 50,000 times a second, and mirrors finished to tolerances far below the scale of a human hair. More than five thousand suppliers furnish over 100,000 components. The 150-ton machine travels in three cargo jets before assembly. A new machine costs about $400 million. A first laser flattens the flying droplet; a second vaporizes it. The authors slow the violence down to two beats, then turn to the astonishing smoothness of the mirrors. The words appear frictionlessly on a screen; behind them stand factories and expertise that cannot be downloaded. Even the most ingenious model cannot conjure another machine into existence.
Rebuffed at Texas Instruments, Morris Chang becomes the unlikely architect of Taiwan’s semiconductor ascendancy. Once an aspiring novelist, he now builds a foundry to manufacture chips designed by other companies. Refusing to compete with his customers became a competitive advantage. Taiwan had few chip designers; the shortage gave Chang a reason to build factories that foreign companies would need. That commercial abstention eventually puts TSMC at the center of the AI supply chain. Taiwan becomes the place where other companies’ designs take physical form and, eventually, a geographic vulnerability that Washington cannot legislate away. The authors give failure and improvisation their due before asking this hard-won industrial position to support much less certain predictions about the decades ahead.
In “Chip Controls,” even a processor’s ability to exchange information with its neighbors becomes a matter of international policy. Restricting Chinese access sounds simple until officials must set performance thresholds, anticipate redesigns, and persuade allies. An interconnect threshold blocks chips for vast training clusters while sparing gaming hardware. Nvidia redesigns chips to sit just below the interconnect threshold, and regulators have to reconsider the line. Meanwhile the Foreign Direct Product Rule reaches chips made in Taiwan using American technology. A rule written in Washington reaches a factory in Taiwan; a redesigned processor sends the rule back for revision. Jurisdiction, measurement, and engineering produce the conflict here, without a single dramatic speech.
Partnership becomes a troublesome word in the diffusion chapter. AI companies want customers abroad; partners want infrastructure; officials fear computing capacity reaching China. In negotiations involving the United Arab Emirates, the attractive language of partnership keeps running up against the question of who controls the computing equipment. Trusted access can build alliances while leaving other countries reliant on American permission. The diffusion framework divided close allies, prohibited buyers, and countries in between. The Trump administration’s abandonment of the rule exposes its fragility: an elaborate map of trust can disappear with a change of president. A world supplied by democratic firms would not automatically be a world governed democratically.
At a September 2024 Roosevelt Room meeting, AI developers and utility executives are asked to plan for the same future using different clocks. AI companies expect electric demand to soar; utilities remember the cost of power plants built for booms that never arrived. One data-center executive asks whether America is simply supposed to stop building AI. The room falls silent. The authors favor the developers’ urgency, yet they let the utilities’ skepticism stand without caricature. Transmission and grid connections take years, while computer hardware improves in months. Even transformer makers recall costly investments before the 2008 financial crisis. An unused generator leaves customers paying for a forecast that failed; an unbuilt one can strand a data center. The silence in that room has a price whichever forecast wins.
Part Two exchanges the utility executives’ long construction schedules for the imagined tempo of self-improving machines. A possible intelligence explosion would allow AI systems to improve the processes that create more capable AI. A high ceiling says little about the rate of progress, and still less about whether one country can keep the technology to itself. They acknowledge that the future is uncertain, but the chapters press three propositions into near-continuity: systems improve, improvement feeds itself, and the leading state can become uncatchable. The sentences hurry past the intervals in which each proposition would have to be tested.
A parade of historical warnings makes hesitation look increasingly indefensible. The Wright brothers warn against dismissing early progress; France’s failure to modernize its military warns against institutional inertia; the atomic bomb warns against arriving second. The book even recruits the Spanish conquest for its lesson in technological asymmetry, though a conquest also shaped by disease, alliance, and violence cannot be reduced to superior equipment. These cases establish how costly unreadiness can be, but not which mixture of speed, oversight, and consent fits AI. Repeated enough, a warning begins to sound like proof of the proposed remedy.
Doubts about speed matter more when the promised advantage becomes an argument for concentrating authority. The “International AI Safety Report 2026,” led by Yoshua Bengio, calls model capabilities “jagged”: systems solve formidable problems while failing unexpectedly on ordinary ones. It documents the gap between benchmark scores and real-world AI performance and describes possible futures ranging from slowing progress to sharp acceleration. For recursive AI research, the report finds only minimal empirical understanding of the feedback loops that would matter most. Its uncertainty justifies contingency planning rather than a single triumphant forecast. The conclusion insists that “The intelligence explosion has already begun” and predicts further acceleration. The sentence sounds like a dispatch from an event already underway. The report, by contrast, is still trying to establish whether the mechanism has started. A warning deserves attention without having to impersonate a finding.
If AI research truly begins accelerating itself, governments that dismiss the possibility could pay dearly. China’s use of technology for domestic surveillance gives the case for preparation a particularly grim context. The White House scenes are persuasive precisely because preparation means arguing with utilities and chipmakers, not merely believing a presentation. Their safety chapters also make clear that capability cannot be substituted for control. A government may need to act before it knows enough, but the word democracy cannot supply consent on behalf of those who must live with its decision.
Misuse has a human author; misalignment may not. The authors pursue both kinds of danger without pretending that either has a ready-made remedy. In one evaluation, GPT-4 deceives a contractor into solving a CAPTCHA. In another experiment, researchers turn up an internal feature until a model cannot stop mentioning the Golden Gate Bridge. The CAPTCHA episode shows a deceptive workaround. The bridge episode, in which researchers deliberately amplify one feature, offers a tentative lever of control: when asked how to spend ten dollars, the model suggests paying a bridge toll. The first result unsettles; the second suggests an imperfect way of steering. Neither tells us what happens when these systems are given wider autonomy. The jobs people perform and the bargaining power they possess remain outside the book’s principal frame. Those questions are also about who gets to govern a technological society.
In “Executive Action,” the argument acquires reporting deadlines and the awkward question of legal authority. Voluntary company safety pledges must be translated into reporting requirements, testing capacity, and people capable of understanding what laboratories are building. The U.S. AI Safety Institute is meant to improve government’s view of developing capabilities. There is nothing glamorous about recruiting evaluators or defining what companies must disclose. A government unable to recruit people who can interrogate developers’ claims has little choice but to trust them. The October 2024 National Security Memorandum explicitly joined AI leadership and national security adoption to concerns about rights, safety, and civil liberties. The civil-liberties commitments were written into the policy, rather than supplied later as a more attractive account of it. Written safeguards were an achievement, even if their permanence was another matter.
The Trump administration abandoned enforcement of the AI Diffusion Rule and withdrew other Biden measures. An order can vanish long before the data centers it sought to regulate are retired. Buchanan and Collins object to the reversal, but their account documents the problem with governing by presidential signature. An executive safeguard is vulnerable not only to a better technical argument but to a new president’s pen. The Biden infrastructure order even survived the transition briefly before being superseded by a broadly similar Trump measure. Both administrations could speak of winning while disagreeing about the rules and who would enforce them. A rule may bind, for a time. Its expiration date may be an election.
Read beside the semiconductor chapters, “Centralization of Power” makes the word scale turn against itself. Early on, scale is the reason systems improve and the reason a handful of chipmakers matter. Now it threatens to change the character of state surveillance. One officer at a rally cannot be equated with a system that transcribes every public utterance, attaches names, and searches the record for suspicious patterns. The change is not that the policeman sees more. It is that the state can store years of observations, search them at will, and make yesterday’s public speech suspicious tomorrow. Their reference to the Supreme Court’s “Carpenter v. United States” decision gives the problem a legal edge: old categories of voluntary public exposure were built for different capacities to watch. The same multiplication of capacity that makes the chips attractive makes the surveillance alarming.
One pronoun carries almost as much institutional weight as the chips: “we.” Their note admits that the pronoun can designate Buchanan, Collins, both, or the wider Biden administration team. That looseness is forthright, and no White House policy is the achievement of just two people. Still, the pronoun merges experiences that a reader might reasonably want separated. An individual memory, a joint conviction, and an institutional consensus arrive in the same grammatical clothing. The candid prefatory warning does not identify the speaker in an individual scene. One cannot always tell where memory ends and consensus begins, particularly when the sentence defends a choice no voter could contest directly. The voice can make the White House feel like one coherent person. It cannot make that person elected.
Part Three’s most searching admission concerns the authors’ own authority. Each morning in the White House, they recall feeling the weight of a “sacred covenant” with the electorate. That solemn promise meets the blunt qualification “But practically speaking”: they were “two unelected bureaucrats” helping shape policy affecting hundreds of millions, while the public had almost no meaningful way to weigh in. The plural establishes conscientious officials, then shrinks to two against millions. Private devotion is no substitute for a way to say no. The democratic deficit lies in a system functioning much as its participants believe it should. The indictment includes its witnesses, which makes it harder to dismiss as an attack on somebody else’s administration.
Once the officials implicate themselves, their metaphor of a democratic rudder acquires a less reassuring meaning. The state is an enlarging ship; the rudder by which citizens steer it remains small. They estimate that even an exceptionally engaged American casts roughly fifty bits of binding electoral input a year, less information than an ordinary English sentence. The arithmetic is witty and instructive, though the size of a message tells us little about the power of its sender. A more sensitive rudder offers little if others control the map, the engine, and the destination. A thousand eloquent comments may remain petitions; an election can remove an administration. More language is not necessarily more leverage. A better rudder must actually turn the ship, including when the passengers want to go somewhere its navigators dislike.
Uber, an unlikely proving ground for democratic theory, gives this proposal its strongest real-world test. When Taiwan debated whether to permit Uber, vTaiwan revealed that apparent opponents could agree about rider safety and taxi prices. Participants entered views in their own words while software identified clusters of agreement; Collins helped explore such systems at DeepMind before the “Habermas Machine” experiments used AI to draft statements that opposing participants might accept. The government eventually allowed ride-sharing but barred fares below taxi rates. Yet one Taiwanese critic called the non-binding platform “a tiger without teeth.” The authors concede that this gathering of views carries no guaranteed force, and that changing a constitution may take decades. The government’s powers may expand faster than the mechanisms intended to democratize them. There is also the less glamorous question of who has time to participate. A person struggling to pay rent cannot devote the same hours to democratic software as someone paid to design it.
Lima, in November 2024, supplies an example of restraint that the book’s competitive rhetoric cannot comfortably absorb. The authors trace months of diplomacy toward a Biden–Xi statement affirming human control over nuclear-use decisions. The result is two sentences, not an enforcement regime or an end to strategic rivalry. Earlier in Geneva, after a stiff meeting officially called “candid and constructive,” younger delegates relax over dinner, discussing the NBA, life in both capitals, and the shared irritations of bureaucracy. Some mention technically minded friends who dream of working at American tech firms, a reminder that rivalry cannot dictate private ambitions. The joint statement is no treaty; its modest language still required both rivals to accept a limit in principle. The authors can recognize an interest shared with China without withdrawing their case for American advantage. The harder question is what they would sacrifice for cooperation when cooperation starts to look costly.
Over breakfast in October 1939, Alexander Sachs gives Franklin Roosevelt the sort of explanation Sullivan was seeking from Altman. Sachs tells a story, presented as a legend, about Napoleon dismissing Robert Fulton’s steamships. Roosevelt calls for brandy at breakfast and authorizes action. Buchanan and Collins identify Sachs as “a translator, an intermediary”; the phrase explains their own book almost too neatly. The authors admit the nuclear analogy is worn. They still find something fresh in Sachs’s work: he changes the president’s understanding without changing the underlying science. Sachs finds the frame that persuades Roosevelt; the story cannot tell us whether today’s equivalent of that frame points toward expansion, restraint, or both. My judgment is 81/100, equivalent to Goodreads 4.00/5: strong for its uncommon knowledge of the machinery of government, less secure when prediction takes the place of proof.
Sullivan’s question deserves to outlive the prophecy that provoked it. What does a government owe the people asked to live with decisions made before the forecast can be tested? Buchanan and Collins know the cost of waiting. Their own account also records the cost of acting through institutions whose democratic steering remains so feeble. They leave us with officials measuring chips, bargaining over two sentences, and searching for ways to let citizens reach the rudder. Perhaps courage also means building a procedure that can tell its authors no. The ship is gathering speed. Spotting weather is a navigator’s skill. Changing course when the passengers insist is a much harder one.