Last updated on January 14th, 2026 at 11:46 pm
Artificial intelligence isn’t “just another technology.” In Paul Scharre’s Four Battlegrounds: Power in the Age of Artificial Intelligence, AI is framed as a fight over who gets to shape the future order—because leadership in AI can translate into economic leverage, military advantage, and influence over global rules.
Scharre grounds this competition in four decisive arenas—data, compute, talent, and institutions—and shows how the fiercest rivalry is between the United States and China, with Russia playing a consequential (if different) role.
Table of Contents
1. Background
Paul Scharre frames Four Battlegrounds: Power in the Age of Artificial Intelligence (W. W. Norton, 2023) as a geopolitical and national-security story about who will shape the “new digital order.”
The opening is explicitly grounded in his experience with military robotics: he describes encountering an explosives-disposal robot during a 2007 deployment to Iraq, which he says triggered his long-running focus on autonomy, later leading to Pentagon work on autonomy policy and participation in UN talks on autonomous weapons.
From there, the book’s core organizing idea is that national AI power is decided across four “battlegrounds”—data, compute, talent, and institutions—rather than by algorithms alone.
Scharre emphasizes that the decisive edge often comes from how effectively societies and bureaucracies convert these inputs into real-world capability (including in defense adoption), not simply from who publishes the best papers.
2. Four Battlegrounds
Scharre’s central claim is simple: in the “cognitive revolution,” the key resources in the global struggle for AI dominance are computing hardware, data, and AI talent—and the ability of institutions to mobilize them effectively.
He describes AI as ushering in a new “digital order,” with nations fighting for control over that future.
1. Data
Modern machine learning is hungry. Data is not merely an input; it can become an advantage—especially in domains where feedback loops exist (collect data → improve model → deploy at scale → collect more data). Scharre ties this directly to state power and domestic control, especially in surveillance.
A striking example is China’s rapid buildout of camera and facial-recognition infrastructure. Citing IHS Markit research, Scharre notes that the combined market share of two Chinese companies—Hikvision and Dahua—was about 77% of the world market for video surveillance equipment.
This matters because widespread deployment produces a flood of labeled real-world imagery—exactly the kind of data that strengthens computer vision systems.
He then quantifies the scale: China’s “Skynet” and “Sharp Eyes” efforts aimed to deploy as many as 500 million surveillance cameras nationwide by 2021, which Scharre describes as more than half of the world’s surveillance cameras.
Even more pointedly, he writes that China is expected to have 73% of all the world’s facial recognition cameras.
The economic gravity is also large. Scharre reports the Chinese security and surveillance industry was estimated at 70 billion yuan in 2019, and projected to exceed 100 billion yuan by 2025. Those figures show why “data” is not just a technical asset—it becomes an industrial base.
But Scharre also emphasizes a second, darker consequence: AI-enabled surveillance isn’t only repressive; it can entrench repression, making it harder for people to organize, resist, or defend freedoms. This turns the “data battleground” into a contest over the relationship between citizens and the state—and potentially the global balance between freedom and authoritarianism.
In Scharre’s framework, data dominance isn’t only about better models; it’s about scale, feedback loops, and governance choices—who is watched, who controls the data, and what the data is used for.
2. Compute
If data is fuel, compute is the engine. Scharre treats compute as a strategic resource because advanced AI increasingly depends on large-scale training runs and production infrastructure.
He cites research showing just how dramatic this has become: from 2010 to 2022, compute used in cutting-edge machine learning research projects increased ten billionfold, with an effective doubling period of about every six months.
This speed matters strategically: it compresses the time between “new technique” and “new capability,” and it rewards actors who can reliably access specialized chips, massive clusters, and high-end cloud systems.
Compute is also tied to supply chains and industrial capacity. Scharre highlights a long-term U.S. shift away from domestic manufacturing: U.S. semiconductor manufacturing capacity fell from 37% in 1990 to 12% in 2021. The implication is not simply “chips are important,” but that compute access can become geopolitically fragile—vulnerable to chokepoints, export controls, or concentrated production.
In military contexts, compute is entangled with procurement, cloud modernization, and operational speed. Scharre describes U.S. defense leaders launching “billion-dollar cloud computing projects” to build AI infrastructure.
And he stresses the institutional side: the U.S. Department of Defense spends about $280 billion per year on procurement and R&D, which is “more than any other government in the world,” yet much of that money goes to traditional platforms rather than AI—and effective AI adoption can still be blocked by process.
In Four Battlegrounds, compute is the scarce accelerator. When compute grows this fast and concentrates in specific hardware and clouds, it becomes a pillar of national power—like oil in an earlier era, but built from silicon and infrastructure.
3. Talent
Talent is the battleground where education systems, immigration policy, research ecosystems, and corporate labs collide.
Scharre points to a specific, quantifiable “brain gain” dynamic: he cites a MacroPolo analysis of AI researchers publishing at NeurIPS. The study found less than a third of Chinese AI researchers who studied in the U.S. returned to China, while more than half came to the United States, and over 90% of those who came chose to stay.
He frames this as a major geopolitical advantage for the United States—if it can preserve its attractiveness and capacity to absorb global talent.
But talent also creates security and values dilemmas. Scharre writes that AI collaboration cannot be treated as “apolitical” when AI is being used to oppress citizens and enable atrocities—he calls the naïve view “morally bankrupt” in that context.
At the same time, he argues that entirely cutting off U.S.–China research ties is not necessarily in U.S. interests, because some cooperation (for example, certain AI safety work) might be mutually beneficial—though he is careful to note that “better” AI is not always a net positive if it improves repression.
Scharre’s talent argument is not “more engineers = win.” It’s that who attracts and retains top researchers shapes innovation speed, military capacity, and even the ability to set standards—while also forcing hard choices about security and ethics.
4. Institutions
If data, compute, and talent are the ingredients, institutions decide whether a nation can actually turn them into durable power.
Scharre’s institutional lens covers governments, militaries, regulators, standards bodies, and the invisible machinery of implementation—procurement rules, oversight, legitimacy, and social trust. In democratic societies, he argues, choices about AI surveillance are shaped through interplay among citizens, government, media, and civil society.
That process can be messy, but it grants legitimacy when it is “inclusive, transparent, and representative.”
By contrast, authoritarian systems can silence debate before it begins, limiting citizens’ ability to challenge abuses. Scharre’s warning is that AI can “supercharge repression,” enabling monitoring at a scale that would be impossible with humans alone.
Institutions also determine military innovation. Scharre describes how JAIC invested in a cloud-based AI platform—the Joint Common Foundation—and worked on improving data management practices so organizations would become “AI-ready.”
Yet he emphasizes a blunt barrier: DoD’s “biggest challenge” isn’t only data, talent, or algorithms—it’s the acquisition system. He summarizes this even more sharply later: the U.S. Defense Department’s biggest adversary in AI competition is its own bureaucracy.
Institutions are the “multiplier battleground.” Two countries can have similar AI models, but the one with better procurement, governance, legitimacy, and standards-setting can convert AI into real-world advantage faster—and with fewer destabilizing failures.
3. AI Tech Rivalry of the USA, China, and Russia
Scharre is explicit: many countries are pursuing AI for national advantage, but “nowhere is the competition fiercer than in the accelerating rivalry between the United States and China.”
Russia is not framed as an equal peer in overall AI ecosystem scale—but it remains strategically important in military competition and risk (especially around safety, escalation, and nuclear stability).
China vs. USA
1. China’s Strategic AI Ambitions
Scharre describes China as pursuing AI leadership with national-level intent and follow-through. He writes that China “launched a national-level AI development plan with the intent to be the global leader in AI by 2030.”
He contrasts U.S. strategy drift with China’s consistent emphasis.
Following the 2017 national AI plan, Chinese leaders issued implementation plans, including a “Three-Year Action Plan to Promote the Development of New-Generation AI Industry” and a “Thirteenth Five-Year Science and Technology Military-Civil Fusion Special Projects Plan,” both released in 2017.
Scharre also quotes Xi Jinping’s framing of AI as a strategic lever. In a 2018 Politburo session, Xi said accelerating AI development is an “important strategic handhold” to gain initiative in global science and technology competition and a strategic resource driving leapfrog development.
China’s ambitions aren’t only policy documents—they connect to industrial scale (surveillance markets, cloud ecosystems) and to the data/compute/talent flywheel described earlier.
2. The United States’ Position
Scharre argues that the United States has enormous strengths (especially in attracting talent and building world-leading firms), but struggles with coherence and execution at the national level.
He writes that the Trump administration released high-level AI documents in 2019, including an executive order on AI leadership and an updated R&D plan, but these didn’t move significantly beyond the Obama administration’s 2016 documents.
He highlights the National Security Commission on AI (NSCAI), chaired by Eric Schmidt and Bob Work, which released a 700+ page report in early 2021—and Scharre discloses he served as a consultant to the commission.
But he stresses a core institutional weakness: the commission could only advise; it had no authority to implement decisions.
Scharre also explores the uncomfortable reality that U.S. and Chinese AI ecosystems are deeply intertwined. He lists concrete indicators that sparked U.S. alarm: since 2007, he cites an estimate of 500 Chinese military scientists studying in the United States; and from 2006 to 2017, Chinese military scientists coauthored over 1,100 research papers with U.S. scientists. Those figures feed “decoupling” pressure, especially amid fears around “military-civil fusion.”
At the same time, Scharre warns against simplistic decoupling: engagement in international standards and norms can be essential precisely because it prevents authoritarian-aligned actors from writing the rules uncontested.
3. Russia’s Role in the Tech Rivalry
In Scharre’s telling, Russia is not portrayed as matching the U.S. or China in AI ecosystem scale, but it is central to military AI risk and strategic stability.
One major theme is safety and reliability. Scharre cites the NSCAI warning that “Russia and China are likely to field AI-enabled systems that have undergone less rigorous TEVV… and may be unsafe or unreliable,” and that the U.S. should emphasize rigorous TEVV and the escalation risks of unsafe systems.
He also describes U.S. defense leader Jack Shanahan advocating that the United States, Russia, and China engage in discussions—especially around the danger of “brittle algorithms” in nuclear decision chains.
Scharre links institutional performance to battlefield outcomes as well. He points to Russia’s 2022 invasion of Ukraine as an example of how military failures can stem from institutional issues—poor planning, poor leadership, and a system where leaders’ need to preserve their image discourages bad news from traveling upward.
While this passage is not “about AI” directly, it supports Scharre’s broader claim: institutions can be the difference between theoretical capability and real-world performance.
Vladimir Putin said “whoever leads in AI will rule the world”.
4. Major Socioeconomic and Geopolitical Implications of AI
1. AI as a General-Purpose Technology
Scharre rejects the simplistic framing of AI as a single “weapon.” He says AI is a general-purpose technology much like electricity. That matters because the largest national advantage may come not from one killer military app, but from broad adoption across society—efficiency, healthcare outcomes, productivity, and economic growth.
He also stresses compounding effects: productivity improvements stack over time, shaping national power.
To quantify plausible labor-market impact, Scharre cites a 2017 McKinsey Global Institute analysis: roughly half of all job-related tasks in the U.S. economy could be automated using existing technology; less than 5% of jobs could be entirely replaced; but six in ten jobs could be significantly transformed, with automation replacing 30% or more of tasks in those roles.
This frames AI as an economy-wide transformation engine—one that can widen gaps between nations that integrate AI effectively and those that don’t.
2. Military Superiority through AI-Enhanced Warfare
Scharre’s military analysis is not science fiction; it’s grounded in operational mechanics—targeting cycles, automation bias, brittle systems, and escalation risk.
He gives a concrete U.S. example: in 2019, the U.S. military demonstrated automation accelerating a “kill chain.” A satellite detected a mock enemy ship; a surveillance aircraft confirmed identification; target data moved to command-and-control; coordinates relayed to a naval destroyer—all automated except the final human decision to strike.
But he warns that more automation can create new dangers. Automation can be brittle, reduce flexibility, and humans can succumb to automation bias.
He describes a broader “security dilemma” dynamic where states’ efforts to increase security provoke others to react—potentially driving a race to the bottom on safety, accelerated tempo of war, and overreliance on brittle AI that raises accident risk.
Scharre’s language becomes stark when describing the end-state risk: AI could usher in a battlefield “singularity” where war moves too fast for human control.
3. Economic Dominance through Technological Innovation
Because AI is general-purpose, economic power flows to nations that can deploy it broadly and manage disruption. Scharre emphasizes that the biggest national advantage may come from adopting AI across society—health, productivity, and growth—not only from military applications.
That returns us to the four battlegrounds: the nation that has compute at scale, data pipelines, talent depth, and institutional capacity is positioned to capture compounding economic gains.
4. Geopolitical Power and Strategic Alliances
AI competition also reshapes alliances and rulemaking. Scharre notes that more than fifty countries have signaled intent to capitalize on AI for national advantage. In that world, standards and regulations become a form of power.
He uses the EU’s privacy regime to illustrate. He writes that the GDPR gave Europe a “first-mover advantage” in establishing global privacy norms—and that when he spoke with Chinese lawyers debating a consumer data privacy law, GDPR was their default starting point.
This is a quiet but powerful geopolitical mechanism: the rule-setter shapes everyone else’s design constraints.
5. Authoritarianism: Control Over Global Governance
Scharre’s argument about authoritarianism is not merely that dictators abuse technology (they can do that with batons and jail cells). The claim is that AI changes the scale and precision of control.
He writes that AI can “supercharge repression,” enabling intelligent surveillance networks at a scale impossible with humans. And because surveillance systems generate data that improves the models, repression can become self-reinforcing.
Scharre also describes how earlier U.S. assumptions about the “dictator’s dilemma” proved wrong.
He quotes President Clinton arguing that without full freedom to think and create, China would be disadvantaged in the information age; but Scharre writes U.S. policymakers underestimated China’s ability to “break the dictator’s dilemma” and use information technology to repress citizens while permitting economic freedoms. He then describes China’s buffers against liberalizing forces: the Great Firewall, increased surveillance, censorship, propaganda, and crackdowns on activists and dissidents.
When you connect this institutional strategy to the data battleground metrics—500 million cameras by 2021, 73% of the world’s facial recognition cameras, and dominant market shares for surveillance equipment—Scharre’s warning becomes concrete: authoritarian governance can be “locked in” by AI.
6. Misuse of AI in Mis/Disinformation and Synthetic Media
Scharre treats synthetic media as a direct threat to public truth, social trust, and democratic resilience.
He writes that deepfake video and audio “continue to improve,” and that long-term trends point toward fakes “indistinguishable from reality,” undermining truth.
1. The scale problem: synthetic content is cheap, fast, and persuasive
Scharre emphasizes that synthetic media is not hypothetical—it’s already widespread. He cites a report estimating 14,678 deepfake videos online in 2019, and that 96% were pornographic.
Even if the exact counts shift over time, the book uses these figures to show the direction of travel: synthetic media spreads first where incentives are strong and barriers are low.
He also includes technical detail showing how quickly the “real-time deception” frontier is moving.
He describes a voice conversion system that can modulate a speaker’s voice with a delay of as low as 10 milliseconds, and notes that delays under 30 milliseconds are effectively unnoticeable in conversation.
The strategic implication is chilling: persuasion and impersonation can occur live, not just in edited clips.
2. Detection is hard—and getting harder
Scharre highlights the cat-and-mouse nature of deepfake detection by citing the Deepfake Detection Challenge launched by Facebook in 2019.
The challenge included over 100,000 videos, drew more than 2,000 participants, and produced over 35,000 models. Even then, the winning model achieved about 65% accuracy.
In other words: even with a large benchmark dataset and strong incentives, detection did not reach near-perfect reliability—while generation quality keeps improving. That asymmetry is exactly what makes synthetic media dangerous as a disinformation tool.
7. Comparison with other major AI books
1. Four Battlegrounds Versus AI Superpowers (Kai-Fu Lee, 2018)
Kai-Fu Lee’s AI Superpowers: China, Silicon Valley, and the New World Order (2018) is also structured around U.S.–China competition, but it is more economy-and-industry centered—AI’s business deployment, job impacts, and the pace of commercialization—than a four-factor theory of national power.
Scharre overlaps with Lee on the idea that “leadership” isn’t only about cutting-edge research—implementation and adoption matter—but Four Battlegrounds builds a broader statecraft frame: compute supply chains, talent flows, and institutional capacity (including defense acquisition bottlenecks) become first-class strategic variables, not just context.
2. Four Battlegrounds Versus The Age of AI (Kissinger, Schmidt, Huttenlocher, 2021)
The Age of AI: And Our Human Future (2021) is pitched as a high-level guide to how AI changes “knowledge, politics, and the societies in which we live,” with a strong philosophical/governance orientation.
By contrast, Scharre’s contribution is more operational and competitive: he repeatedly treats AI as a foundation of military, economic, and political power that depends on tangible inputs—especially compute and institutions—rather than primarily on a shift in epistemology.
In short: The Age of AI leans “what this means for humanity and governance,” while Four Battlegrounds leans “how nations actually compete, build, and deploy.”
3. Four Battlegrounds Versus Life 3.0 (Max Tegmark, 2017)
Tegmark’s Life 3.0: Being Human in the Age of Artificial Intelligence (2017) is explicitly framed around how AI may affect “crime, war, justice, jobs, society and our very sense of being human,” and is commonly read as a big-picture exploration of future pathways and risk/benefit tradeoffs.
Scharre’s book intersects with Tegmark most strongly where AI affects security and freedom, but it is less speculative about distant futures and more focused on near-to-medium-term power competition and control points—e.g., compute scaling and the strategic implications of access to advanced chips and the institutions that can exploit them.
4. Versus Weapons of Math Destruction (Cathy O’Neil, 2016)
O’Neil’s Weapons of Math Destruction (2016) is a warning about algorithmic harms—models that “increase inequality and threaten democracy”—and it focuses on opacity, unfairness, and societal damage from data-driven systems.
Scharre’s lens is different: he is not primarily writing an ethics-of-algorithms critique, but a geopolitical account of how AI capability becomes national leverage. Still, they connect on a practical point: governance and institutions matter because AI’s effects (good or bad) are amplified when embedded in large-scale systems.
5. Versus Rise of the Robots (Martin Ford, 2015)
Ford’s Rise of the Robots: Technology and the Threat of a Jobless Future (2015) is centered on automation’s labor-market shock and economic restructuring, emphasizing how technology undermines work and reshapes society.
Scharre addresses economic power too, but Four Battlegrounds is less a jobs book and more a state capacity / competition book: who can mobilize data, compute, talent, and institutions to gain advantage (including military advantage) is the primary plot line.
8. Personal Reflections
One of the strengths of Four Battlegrounds is that Scharre doesn’t write as a distant observer. He opens with personal experience that shaped his view of autonomy and risk.
He recounts watching an explosive ordnance disposal team in Iraq deploy “a little robot” rather than a human in a bomb suit—and the realization clicked: send the robot to defuse the bomb and spare human life.
He then describes working at the Pentagon on emerging technologies and helping push investments in military robots, believing the machines could help save lives and defend the nation.
That experience pulled him into the hardest questions: autonomy in weapons, and the legal, ethical, and strategic dimensions of delegating decisions to machines. He writes that he helped draft the Defense Department policy on autonomy in weapons and participated in UN diplomatic discussions on autonomous weapons.
What changed, he argues, is that the robotics revolution gave way to something deeper: the AI revolution—powered by exponential growth in data and compute—reshaping power dynamics and potentially human freedom.
This personal thread matters because it explains the book’s tone: Scharre is not arguing for panic, but for clarity about tradeoffs. He pushes for engagement on standards and norms (to influence outcomes), while insisting that engagement must include pushing back against abuses and standing up for democratic values.
And he ends with a sober institutional diagnosis: despite ethical controversies around military AI programs, the bigger long-run obstacle is often bureaucracy—contracts and acquisition systems clogged in ways “lethal to innovation.”
9. Conclusion
In Four Battlegrounds, Paul Scharre’s message is that AI power is not decided by a single breakthrough model or one dramatic weapons system. It is decided by who wins—consistently—across four battlegrounds:
- Data, especially at national scale and in high-feedback domains like surveillance.
- Compute, where growth is explosive (ten billionfold from 2010–2022) and access is strategic.
- Talent, where the U.S. advantage depends on remaining the top destination and retaining researchers.
- Institutions, where legitimacy, procurement, standards, and governance determine whether resources translate into real advantage.
The fiercest rivalry is the U.S.–China contest—because China has declared a goal to lead by 2030 and backs it with policy implementation, while the U.S. has struggled to convert reports into coordinated action.
Russia’s role, meanwhile, looms largest in military risk and stability—especially where unsafe systems and brittle algorithms could accelerate escalation.
Finally, Scharre’s warning about synthetic media is not abstract: deepfakes already proliferate, real-time voice manipulation can happen at near-imperceptible latency, and detection remains imperfect even at scale.
AI will shape the next global order—but in Scharre’s telling, the outcome is not predetermined. It will be built—by choices—in these four battlegrounds.
10. FAQ: Four Battlegrounds: Power in the Age of Artificial Intelligence (Paul Scharre)
1. What is Four Battlegrounds about?
Paul Scharre argues that modern geopolitical power in AI is shaped by competition across key inputs—not just “better algorithms.”
He frames AI rivalry as a contest over critical resources (notably hardware/compute, data, and talent) that determine which nations can build, deploy, and scale AI systems for economic, military, and political advantage.
2. What are the “four battlegrounds” in the book?
The book’s central framework is that AI power depends on four arenas: data, compute, talent, and institutions—the resources and systems that let countries turn research into real-world capability.
3. Why does the book say data is a battleground for AI dominance?
Scharre emphasizes that large-scale deployment (especially surveillance and consumer platforms) generates massive data streams that can be used to train and refine AI systems. He points to China’s domestic surveillance buildout—over 500 million cameras deployed nationwide by 2021—as creating major data advantages for firms improving systems like facial recognition.
4. Why is compute considered the “crux” of AI competition?
The book highlights compute as a limiting factor because cutting-edge AI increasingly requires enormous training resources. It cites that from 2010 to 2022, compute used in cutting-edge machine learning research increased ten billionfold, growing exponentially and doubling every six months.
5. What are the biggest semiconductor chokepoints mentioned in the book?
Scharre describes critical chokepoints in the chip supply chain. For example, he notes that Japan, the United States, and the Netherlands control over 90% of the global semiconductor manufacturing equipment market, and that the Dutch company ASML is the only supplier of EUV lithography tools used for leading-edge fabrication.
6. How does AI talent shape national power, according to Four Battlegrounds?
The book argues that people—not algorithms—are scarce, and it cites how competitive the market is: fresh AI PhDs reportedly make $300,000 to $500,000+ in compensation, with top researchers able to make millions.
7. What does the book say about the U.S.–China “AI talent flow” problem?
Scharre frames U.S.–China competition as deeply intertwined, including research and talent linkages. He cites estimates such as about 500 Chinese military scientists studying in the U.S. since 2007, and over 1,100 coauthored papers between Chinese military scientists and U.S. scientists from 2006–2017.
8. What is China’s AI strategy goal for 2030?
The book states that China launched a 2017 national AI development plan aiming to become the global leader / “world’s primary AI innovation center” by 2030, followed by implementation plans and high-level political emphasis from leadership.
9. How does the book define AI competition: “arms race” or something else?
Scharre argues that AI is not a weapon, but a general-purpose technology (he compares it to electricity). He also warns the strategic dynamic can resemble a security dilemma, where states’ efforts to increase security trigger counter-moves and risks like racing “to the bottom” on safety.
10. What are the economic impacts of AI and automation highlighted in the book?
Citing a McKinsey Global Institute (2017) analysis, Scharre notes that roughly half of all job-related tasks in the U.S. economy could be automated with existing technology; less than 5% of jobs could be fully replaced; but six in ten jobs could be significantly transformed, with 30%+ of tasks automated in those roles.
11. What does Four Battlegrounds say about deepfakes and AI misinformation?
The book details how detection is difficult even at scale. It cites Facebook’s 2019 Deepfake Detection Challenge, which produced a dataset of 100,000+ new videos, drew 2,000+ participants, and led to 35,000+ submitted detector models—yet the best detector was only 65% accurate.
12. How does the book connect AI to authoritarianism and global governance influence?
Scharre argues that the spread of AI-enabled surveillance isn’t only about exporting hardware—it can include exporting governance models and information controls.
He cites Freedom House reporting China has held trainings with over thirty countries on cyberspace/information policy and provides examples of foreign officials trained in how to “guide public opinion” online.
13. What does the book say about the U.S. military’s institutional challenges with AI?
Scharre describes U.S. efforts to accelerate adoption through new entities like the Joint Artificial Intelligence Center (JAIC) (founded 2018) and notes internal bureaucratic resistance (“antibodies”) that can stall initiatives—even when AI projects show promise.
14. Does the book suggest ways countries can reduce AI conflict risks?
Yes. It proposes that states can pursue confidence-building measures, transparency in testing/evaluation, and norms like keeping humans “in the loop” for critical decisions, to reduce accident risks and destabilizing escalation.
