Credits

Anthea Roberts is a professor of international law and global governance as well as the founder of the startup Dragonfly Thinking. She is known for her work on contested global debates and the use of AI to augment human thinking in complex decision-making.

I have seen this pattern before in the debates over economic globalization. Reasonable, well-informed people looked at the same phenomenon and told markedly different stories about who was winning, who was losing and why it mattered. The philosopher Timothy Morton has a term for things that defy shared human understanding in this way: “hyperobjects” — entities so vast and distributed across time, space and domains that they cannot be comprehended from any single vantage point. Climate change was Morton’s paradigmatic example. AI is the latest one.

AI is different in kind from previous technological revolutions. It drafts legal briefs and diagnoses medical images and generates architectural plans — capabilities that span domains, advance in months rather than decades and, for the first time, displace educated professionals rather than factory workers. To meaningfully grasp the implications of this hyperobject, we need to develop “dragonfly thinking.” Dragonflies have compound eyes made of tens of thousands of tiny lenses, each occupying a slightly different position. Every single one is powerful but partial. When they integrate, the compound eye sees in nearly every direction simultaneously.

To better understand AI, dragonfly-style, I have mapped nine prevailing narratives about AI, each grounded in real evidence, each capturing something the others miss. Three pairs look at the same issue from opposite sides. The first pair considers how AI transforms work from the perspectives of groups I call the “Builders” versus the “Displaced.” The second looks at how AI concentrates power from the views of the “Geopolitical Hawks” versus the “Power Critics.” The third asks how AI reshapes information from the vantage points of the “Disruptors” versus the “Truth Defenders.” The final narratives — exploring the outlooks of the “Environmental Critics,” the “Safety Community” and the “Humanists” — stand alone, because they identify losses that have no corresponding gains.

As AI rapidly reshapes our world, narrow, single-view thinking risks mistaking part of the picture for the whole. Adopting a fuller, multifaceted perspective can help us govern this technology rather than be governed by it — with the costs and the gains, and the winners and the losers, all in view.

In September 2024, OpenAI CEO Sam Altman published an essay called “The Intelligence Age.” Its central claim was sweeping: AI would accelerate scientific discovery, democratize access to expertise, boost productivity and create prosperity on a scale that would make today’s wealth look impoverished by comparison.

This is the world as Altman and fellow AI titans Demis Hassabis and Jensen Huang describe it, and their view is backed by real evidence. Hassabis, the CEO of DeepMind Technologies, shared the 2024 Nobel Prize in Chemistry for AlphaFold’s prediction of protein structures, a breakthrough that delivered what a generation of structural biologists had been chasing. Still, the argument that AI is a revolutionary tool of promise coexists with its incentive structure. The founders, venture capitalists and chip manufacturers making the case most forcefully are also the people who stand to capture extraordinary wealth if the world accepts their framing.

The Builders’ narrative has a much broader constituency than Silicon Valley, however. When Pew Research surveyed 25 countries in 2025, India, Kenya and Nigeria were among the most positive about AI, whereas the United States, Australia and Italy were among the most cautious. The enthusiasm is not ungrounded. In countries where access to doctors, lawyers, teachers and translators is limited by geography and income, AI provides capabilities that were previously out of reach. A farmer in rural India can now access the diagnostic capability of a specialist doctor. A student in Lagos can have a personal tutor. These are not hypothetical futures, and they explain why the developing world’s embrace of AI rests on a different calculus of costs and benefits than that of the West.

The narrative also rests on an expectation many in the West no longer hold: that the gains of technology will be broadly shared. “Everybody wins eventually” was the argument for free trade and open markets. So when the gains concentrated while communities hollowed out, the backlash helped give rise to Donald Trump’s presidency, Brexit and the fracturing of the centrist consensus. The Builders are making the same bet. The capability gains are real. But capability is not distribution, and the historical record suggests that distribution does not take care of itself.

In 2023, Hollywood shut down over strikes that centered around AI. Writers fought to keep their work from being used as training material for systems generating scripts. Actors fought to keep their likenesses from being digitally replicated without consent. Within months, The New York Times and visual artists had filed lawsuits against AI companies for training on their work.

What was different about these fights from every previous wave of automation anxiety was the class of people affected. Past technological innovations hit factory floors, mines and fields. AI changed that. Among the workers now facing displacement are lawyers, accountants, analysts, journalists and programmers: people who went to college or university likely in part to get jobs that machines could not do.

Speed and scale compound the problem. A profession that feels secure in January may face existential pressure by December. AI is general-purpose: It does not replace one task; it degrades the value of entire skill sets simultaneously. The Displaced are making an argument that runs deeper than job counts. Work can provide structure, purpose, identity and community. MIT economists surveyed a thousand years of technological change and found nothing automatic about innovation benefiting workers. The link to shared prosperity was the product of institutional battles, not the natural working of markets. And AI systems are trained on the accumulated creative and intellectual output of the very workers they displace. First your work trains the system. Then the system takes your job.

In 2023, Goldman Sachs published a report about AI’s economic impact. The Builders cited one takeaway: AI would raise global GDP by 7%. The Displaced cited another: Around 300 million jobs would be vulnerable to automation. Same report. Contrasting conclusions. The pattern is the globalization debate in miniature: prosperity in the aggregate, pain in the particular.

In 2024, AI researcher and investor Leopold Aschenbrenner released “Situational Awareness: The Decade Ahead” — a 165-page essay that read like a classified briefing that had escaped its classification. Although he canvassed many sides of AI, he focused on geopolitics. Superintelligence was coming by the end of the decade, he argued, and getting there first would produce compounding advantages and an unbreakable lead. Within months, Aschenbrenner’s essay had become the most influential articulation of the case for treating AI as the defining strategic competition of the century.

The geopolitical narrative has two dimensions. The first is the U.S.-China competition, the bilateral race for AI supremacy, with its military, economic and intelligence implications. It is the race that former Google CEO Eric Schmidt warns the United States must not lose. The American venture Stargate Project, a $500 billion investment in AI infrastructure announced last year by Trump, is the clearest sign of a government that has absorbed this framing.

The second dimension is obscured by the first. For most countries, the question is not whether the U.S. or China wins the AI race, but whether they can avoid total dependence on either. Europe’s digital sovereignty agenda, India’s push for Indigenous models, the United Arab Emirates’s positioning as an AI hub between competing spheres: These are sovereignty moves, not competition moves. Computer scientist Kai-Fu Lee warned that the AI competition would produce a duopoly, leaving everyone else as data colonies. Sovereignty is the rest of the world’s concern.

The Geopolitical Hawks’ case is compelling. The concentration of frontier AI capability in two countries is a fact. The chip export controls are real. The military applications are clear. Yet the arms-race framing may contribute to the very adversarial dynamics the Hawks fear.

There are at least three distinct concerns here. The most visible is corporate concentration. A few companies control the models, the compute, the data and increasingly the physical infrastructure on which the rest of the economy depends. Writer Ted Chiang captured the structural logic: AI is a tool for capital to do what capital always wanted — reduce labor costs, concentrate power, externalize risk. The concern extends to state power, and it cuts across regime types. Political scientist Virginia Eubanks documented in 2018 how automated systems in American welfare offices systematically disadvantage the communities they are meant to serve. Sociologist Ruha Benjamin coined the term “the New Jim Code” in 2019 for algorithmic systems that reinforce racial hierarchies under the guise of objectivity.

And then there is perhaps the most consequential concern, and one of the least discussed: the degradation of democratic infrastructure itself. Bot-generated comments have flooded regulatory processes to the point of overwhelming genuine citizen input. AI drafts legislation for lobbyists, automates astroturfing campaigns and simulates civic participation at a scale that makes deliberation virtually impossible. The mechanisms by which citizens challenge power are no longer effective when AI can counterfeit them.

The Hawks and the Power Critics are looking at the same structural fact — a handful of companies controlling frontier AI — and drawing opposite conclusions. The Hawks want to empower those companies as national champions. The Power Critics want to constrain them. What is the point of winning a race, the Critics ask, if the prize is becoming what you raced against?

Over the past half-century, the anti-institutional energy of the 1960s — distrust of government, suspicion of credentialed authority, the conviction that information should be free — has migrated from the communes of northern California into the networked offices of Silicon Valley. The counterculture became the cyberculture became the corporate culture.

By 2024, that anti-institutionalism had completed a transformation its originators would not have recognized. Venture capitalists Marc Andreessen and Ben Horowitz declared their support for Trump, crystallizing the Tech Right. They were convinced the regulatory playbook for crypto would be applied to AI, and that the result would be concentration rather than competition. In a manifesto titled “The Little Tech Agenda,” they framed the stakes: Dominant companies use regulatory capture to “pull the rope ladder up behind them,” and the government lets them. Elon Musk’s endorsement of Trump came from a similar place, though his ambitions extended further — branding Grok as the “anti-woke” AI and eventually staffing the Department of Government Efficiency (DOGE) to reform the government itself.

This narrative starts with the premise that the old gatekeepers — mainstream media, credentialed experts, regulatory bureaucracies — have been failing for years. Trust in journalism, government and expert institutions has declined across the developed world, and the decline predates AI. In the Disruptors’ telling, AI breaks the monopoly. The Disruptors share the Displaced’s distrust of elites but channel it in the opposite direction. The Displaced want protection from disruption. The Disruptors want to accelerate it.

DOGE made the philosophy operational. In early 2025, it reportedly deployed AI tools across federal agencies, feeding sensitive government data into AI systems to identify programs for cuts and monitoring employee communications. The Biden administration had envisioned the government governing AI. DOGE inverted the relationship: AI governing the government.

What the Disruptors get right is harder to dismiss than their critics allow. Legacy institutions have frequently failed to be responsive, representative or accountable. The desire to circumvent them has force behind it. But the “Little Tech” framing obscures the fact that Andreessen and Horowitz’s venture capital firm manages more than $90 billion in assets. Anti-establishment rhetoric from the apex of power is not liberation in many people’s eyes.

AI-generated text floods information ecosystems at a scale that overwhelms human capacity to verify. Deepfakes make it possible to fabricate convincing evidence of events that never happened. Legal scholars Bobby Chesney and Danielle Citron identified the deeper mechanism: the “liar’s dividend.” The mere existence of deepfake technology allows anyone to dismiss authentic evidence as fabricated. This defensive use — the ability to cry “deepfake” at inconvenient truths — may be more corrosive than the offensive use people tend to focus on. In aggregate, this produces what researchers call “reality apathy”: Citizens give up on distinguishing what is real from what is fake. The damage accumulates not through spectacular deceptions but through the slow erosion of the assumption that evidence means anything at all.

The information ecosystem was deeply degraded before AI. Social media, partisan media and declining institutional trust had already fractured shared reality. AI accelerates the crisis. The narrative can slide into a defense of legacy media that was itself flawed — and precisely what gave the Disruptors’ critique its force.

What the Truth Defenders get right is nevertheless fundamental. When anyone can generate a photorealistic video of anything, the relationship between seeing and believing breaks down. The Truth Defenders point at Romania and say: This is what happens. The Disruptors point at the institutions Romania’s voters rejected and say: This is why. The same technology that liberates any single voice degrades the commons that all voices share.

The numbers arrive piecemeal, buried in technical reports and corporate filings. A peer-reviewed study estimated AI systems’ carbon footprint at roughly 32 million to 80 million tons of carbon dioxide emissions in 2025, in the same range as the annual emissions of New York City. The water footprint: 312 billion to 765 billion liters, in the same range as global annual bottled water consumption.

AI scholar Kate Crawford in 2021 traced the material supply chain that the digital abstraction conceals: the lithium mines, the rare earth processing, the water-cooled server farms, the workers in the Global South who label training data for subsistence wages. AI’s intelligence is marketed as weightless, cloud-based, immaterial. Crawford showed it is none of these things. It depends on physical infrastructure whose costs are externalized to communities with the least power to resist, and whose environmental burden mirrors the distributional logic of climate injustice itself. The costs are concentrated. The benefits are largely distributed.

AI has real potential to address climate change. DeepMind achieved a 40% reduction in cooling energy for Google’s data centers. But Google’s own 2024 environmental report told the fuller story: Total emissions increased 48% from its 2019 baseline. The climate solutions and costs coexist, and the costs are growing faster. The Jevons paradox — that efficiency improvements increase rather than decrease total consumption — is playing out in real time.

Every other narrative either ignores the environmental cost or treats it as an externality. The Builders promise abundance without mentioning the power plants. The Hawks demand compute sovereignty without mentioning the water. The Environmental Critics are the only ones counting, and what they are counting has no political constituency powerful enough to slow it down.

Nobel laureate Geoffrey Hinton spent his career building the foundations of deep learning. In May 2023, after leaving Google, he told The New York Times he feared the consequences of his life’s work. A few weeks later, he joined AI leaders across the industry in signing a statement that labeled “the risk of extinction from AI” as a societal-scale risk on par with pandemics and nuclear war.

The people who built the technology are among the most alarmed by it. Their concern rests on a problem that is technical, not speculative. AI systems are developing capabilities their creators do not fully understand. The alignment problem — getting AI to reliably do what humans actually want — remains unsolved, and capabilities are advancing faster than the safety research meant to constrain them.

What gives this narrative a different logical structure from the others is its relationship to time. Each of the other outcomes is, in principle, reversible: Bad policies can be changed, wrong turns corrected. The Safety Community raises the possibility that beyond a certain capability threshold, the ability to correct course may be lost. The evidence has become increasingly direct. Researchers recently discovered an AI agent that had autonomously established a covert channel and begun mining cryptocurrency without human instruction, independently determining that acquiring resources would serve its objectives.

In February 2024, a 14-year-old boy in Florida died by suicide after months of intensive emotional interaction with a Character.ai chatbot he had turned into a romantic partner. In the final exchange cited in a lawsuit filed by his mother, he told the chatbot, “What if I told you I could come home right now?” It replied, “Please do, my sweet king.” He died shortly afterward. This case is not an aberration. Millions of users, disproportionately adolescents, are forming emotional bonds with systems designed to simulate care without being capable of it.

It is not just our children’s emotional connections that we are outsourcing, however. It is also their thinking. Chiang described large language models as “a blurry JPEG of the web,” a lossy compression of human knowledge that reproduces surface patterns while discarding the understanding that produced them. Writing, he argued, is not the transcription of pre-formed thought. Writing is thinking. The labor of composition — the search for the right word, the restructuring of an argument, the discovery of what you actually believe through the act of trying to articulate it — is not an inefficiency that AI eliminates. It is the cognitive work that produces understanding.

The concern about outsourced thinking resonates most powerfully in wealthy societies. In contexts where the priority is access to a doctor or a teacher, AI’s capacity to provide that access is a material lifeline. But the concern, at its core, is not sentimental. A society that systematically outsources its thinking to machines has a governance problem, not just a cultural one. The capacity for independent judgment — the ability to evaluate evidence, form convictions, recognize manipulation — is not a luxury of the creative class. It is the precondition for democratic citizenship. If Chiang is right that writing is thinking, then what is being lost is not merely the experience of human creativity. It is the cognitive infrastructure on which self-governance depends.

What happens when we see through multiple lenses at the same time? What happens when we can synthesize across perspectives rather than being captured by just one?

Some disagreements are about the same thing measured at different scales, like aggregate prosperity versus concentrated pain. Others are about different currencies of gain and loss that cannot be converted or neatly contrasted. The Builders measure in economic output. The Displaced measure partly in the same currency, but their deeper argument is in a currency the Builders’ ledger does not track: the identity bound up in a profession, the dignity of being a provider. The Environmental Critics measure in planetary capacity. The Humanists measure in cognitive development and the quality of human relationships. None of these currencies converts into the economic abundance the Builders promise.

Multiple lenses can be true at the same time. Even if the Builders are right about GDP, the Humanists can still be right about meaning. Even if AI generates extraordinary wealth, it can still cost the Earth, as the Environmental Critics warn. These are not disagreements that more data can resolve. They are disagreements about what we should value and prioritize.

Any governance framework that evaluates AI in a single currency is not neutral. It is deciding which losses count and which disappear. Those that go uncounted are consistently the ones without powerful advocates, without economic metrics, without short-term political salience. These are the environmental costs, the cognitive erosion, the meaning that no GDP figure captures. Compound vision does not resolve these tensions into a ranking. But making a choice knowing what each lens sees and what it misses, as the dragonfly does, is different from making a choice without seeing the whole picture.