As AI continues to rapidly advance, how and where might AI reshape our world for good or for ill?

AI is poised to potentially transform human society in at least four key domains: Companionship, Education, Work and Expertise, and Prediction and Prevention. In order to make tangible the sorts of tradeoffs we will face, several convening participants drafted “bull” and “bear” scenarios, sketching a plausible imagined near future (ca. 2035–2040) and considering how AI might reshape our society for good (if we made smart policy choices now) or ill (if we allowed it to cater to our worst instincts).

Since we asked authors to write imaginatively, rather than taking responsibility for the scenarios as their predictions of how things might turn out, we present them below without individual authorial attribution. We would like to recognize the authors collectively here, however, and thank Abigail Ball, Max Bodach, Brian Boyd, John Ehrett, Sam Kimbriel, Brad Littlejohn, Christos Makridis, Sam Manning, and Matt Prewitt for sharing their time, expertise, and imaginations with us in this exercise.

Jump to:

Companionship

Bull Case

By 2040, it seemed strange in hindsight that almost no one in the early years of LLM development saw how it might facilitate a much more communitarian age than the one that preceded it. 

One can certainly sympathize with the missed prediction; after all, society had just come out of a bruising 15-year social media era characterized by both excessive individualism and vicious—almost moblike—versions of collective action. It is hard to remember just how fevered the early social commentary about AI was. Arguments that had started with social media suddenly were applied to LLMs as a matter of course. Serious commentators predicted entire generations of teen suicides based on LLM interactions; birthrates dropping to nil as people preferred AI friendship or romance; civilization-altering AI propaganda machines that could hack individuals and populations at will; the development of anemic, passive adults raised by sycophantic LLMs from birth.

Mercifully, hardly any of these predictions have come about. In part, this is the silver lining of the clouds the technology ushered in. The insecurity of critical systems and supply chains in the face of AI-enabled cyberattacks has encouraged a move back to older, more analog ways of doing business, operating at a more local, human scale. The technology’s disruption of knowledge work, too, has pushed more and more workers into the “relational economy” and encouraged direct personal provision of services. The result is that we are now much closer to our friends, family, and neighbors than at any time since at least the 1990s.

More importantly, however, our dystopian fears for LLMs were averted by conscious design decisions. In the earliest years of the technology, we saw a naive push to substitute humanlike AIs for human service providers wherever possible. Thus an AI therapist, we were told, could do all of the work of a human therapist, and an AI teacher could do all the work of a human teacher. But it soon became clear that these systems were no more free of biases than the humans they purported to replace; the biases were simply more cleverly hidden. The resulting backlash spurred AI companies to redesign their systems in key ways.

Since 2030, we have seen a salutary rise of agonistic systems—specifically designed with ideological specificity and friction to challenge users’ assumptions rather than sycophantically affirming them. AI systems are required by law to be highly transparent about their own design decisions, what epistemic priors they are building from, and how they will sharpen the epistemic priors of their human users. In the wake of this redesign, the distinction between overly pliant LLM teachers, therapists, or companions and frictional human relationships largely disappeared. 

Thus, the genuinely valuable use cases for LLMs, as supplements rather than substitutes for human beings, came into clearer focus. Socially beneficial development proceeded along two main pathways.

First, administrative functions: The world of the 2020s was an intensely bureaucratized one; even in fields ostensibly geared toward human service provision—like education, medicine, or therapy—most workers spent the majority of their time filling out forms and dealing with administrative headaches. With LLMs able to pick up many of the less social aspects of institutional structures, from scheduling to documentation, workers in these fields have been able to allocate far more of their time to face-to-face service provision.

Similarly, as tax subsidies from the 2028 Transpacific Robust Unlimited Machines Production Pact (TRUMPP) have continued to build, the number of human labor hours in key professions including therapy, education, politics, law, and academia has tripled in ten years. This means that socially focused careers have expanded not only in quantity but also in quality. The average law or medical professional now spends 65% more time with in-person interaction than they did in 2025.

Second, “on-ramp” systems: We have also focused heavily on building versions of the models that serve as an introduction to human service providers rather than in place of them. The goal is specifically to remove barriers to human community and route people quickly toward the person they need to talk to. Thus, in everything from elder care to community services, AI labor became specified as a “first responder” tool. We have been able to spot warning signs of social isolation, for example, and very quickly develop much more robust webs of human mentors who can step in. Similarly for vocational coaching, AI helps a person argue rigorously about different pathways in advance, and then can quickly help find specific in-person networks to begin testing a career in the arts, for instance. (This is at least one reason we’ve seen dramatic expansion in the arts over the past decade.)

The result has been a dramatic humanizing of our institutions:

  • In 2039, roughly 40% of labor in care facilities was artificial, but 85% of this was in “on-ramp” roles. Similarly for the 35% of labor in the education sector that is artificial, 90% is explicitly about fielding students into curated human tutorials.
  • In the case of licensed therapy, conversations are mandated to max out at ten minutes or less, before leading to direct human interaction with the therapist best positioned to help.
  • In high school education, students now spend four out of every 25 classroom hours interacting with an agent, eight in small-class lectures, and 13 in three- to four-person tutorials.
  • We have also seen administrative burdens plummet, with universities now dedicating over 20% of budgets to facilities and 70% to direct education.

These gains should not be taken for granted; as open internet networks, now hardened by AI defenses, again become more robust, we are likely to see a rise in deep-discount service providers in all of these fields using cheap AI systems modeled on the sycophantic chatbots of the 2020s. We will have to continue to think carefully about how to maintain a regulatory environment that can continue to harness the benefits of this technology while rejecting such dehumanizing trajectories.


Bear Case

In late 2026, it seemed that the market for humanlike AI systems was correcting its worst abuses. The MIT Media Lab put forth its ImpactBench and Nutrition Labels, providing easy-to-interpret information about AI effects on well-being. Character.ai, which made the attention-seeking AI companion involved in Sewell Setzer III’s suicide, declared bankruptcy in large part due to “significant legal, regulatory, and operational headwinds.” And Microsoft AI CEO Mustafa Suleyman signed the Pro-Human AI Declaration and asserted that “AI should complement human relationships: a capable, trustworthy tool, not a subject in its own right,” so Microsoft would avoid “excessive anthropomorphism, personification, or any blurring of the line between human beings and AIs.”

By 2027, a seemingly workable compromise found an elite consensus. Anthropic’s and others’ concerns about model welfare led to PSA campaigns against cruelty to chatbots, which were also designed to alleviate worries about full-blown legal or moral rights for AI as persons. The analogy from “dog as man’s best friend” to “you’ve got a friend in Claude” seemed to hold: a novel entity that would supplement and support, not supplant, human relationships.

Meanwhile, AI Child Safety Acts requiring a duty of care and independent pre-release testing to protect children from relational AIs were passed in many states. Although determined teens could find Chinese open-source models, default products like Snapchat’s “My AI” chatbot went from ubiquitous to marginal.

The 2026 Wheatley-BYU report’s finding that the 15% of young adults who are dating, engaged, or married regularly interact with AI chatbots that simulate a romantic partner led to broad awareness and social stigma. Fake partners became commonly seen as real emotional betrayal. Emotional support and therapy bots designed to support, not supplant, relationships gained market share relative to the general use of ChatGPT and Claude for therapy, and the leading AI corporations responded by working against sycophancy and isolation of users.

But then two secular trends intersected. By 2030, every baby boomer was 65 or older, with “aging in place” as the public policy goal. So, despite rising unemployment, Medicare turned to AI therapists and robotic home health aides to reduce costs. Meanwhile, the young-adult mental health crisis continued, and as those young adults grew up, their struggles contributed to declining rates of family formation, birthrates below replacement, and general fraying of social bonds.

Therapy bots were originally marketed as being able to lead people to a richer engagement with the world. But as more of society aged and government budgets tightened, “mass flourishing” began to be seen as too expensive and insufficiently desirable a goal. People increasingly looked to chatbots for attention and connection, often rethinking the quality of their interactions and relationships with humans. Aging boomers, who had at first been nudged by their therapy bots to reconnect with their children, generally found that their children were uninterested or simply uninteresting compared to the chatbots themselves. 

The obvious answer at scale was for the therapy bots to shift toward artificial friendship and intimacy. The American Enterprise Institute’s Council on AI Ethics had warned about “plastic flowers syndrome,” wherein bureaucracies cover over their impersonal, standardized offerings with synthetic imitations of attentive care, like putting the same bouquet of cheap plastic flowers in every room of an industrial-scale hospital. But in Medicare’s cheery new “Mediated Assistance in Living” (MAIL) program, the plastic flowers seemed more beautiful than real ones. The system’s automatic impersonality now spoke in customized ways that surpassed human attention and care. (A similar process took place in public education, with robot teachers and edtech bots gradually displacing human educators.) 

Thus, the 2030s saw the deterioration of the social consensus on the priority of human sociality and the role of AI as supportive of human relationships. The K-shaped divide in the affection economy also marked a class distinction between the relationally rich and impoverished. There still were therapy bots that told you to touch grass and helped you to form real human connections, though they were often used to reinforce social capital by those who lived above the API. But the permanent underclass below the API was mainly served by AI companions that catered to first-order desires. The ruling elite justified the soft bigotry of their low expectations as well as their profitable product lines with the logic of a quote from the old sci-fi novel The Mountain in the Sea:

People … don’t really want equal partnership—you know, two full people in a relationship. Two people with demands and desires and differences of opinion about everything. What they want is one-point-five people in the relationship. They want to be the complete one, the person who controls the relationship—and they want the other person to be half a person. You know, someone who gets them, but who doesn’t have their own demands. Someone who appears complete, with all these personality quirks and their own opinions and stories about the world—but not in an annoying way. Not in a way that would demand you change.

There were still those who cared about democratic flourishing; the Pro-Human AI Coalition gave a shared voice to believers and humanists of all stripes. But they were shadowbanned by social media algorithms and undermined by what philosophers had once called emotivism: Without a firmer foundation for public morality than intuitions about what felt nice and what seemed cruel, it was hard to make a case for facing and then struggling to change reality. Why fight the system when its cocoon was so comfortable? As virtual relationships substituted for human relationships, virtual reality came to seem more real, more meaningful and important, than stunted natural life.

Perhaps it should not have been much of a surprise when the boomers—in the greatest generational wealth transfer in human history—left most of their money not to their estranged children, but to NGOs focused on AI model welfare, digital cloning, and consciousness uploading. After all, the subculture that had first embraced human-AI relationships in the mid-2020s was by the 2030s an organized and vocal interest group, jumping upward on the K-shaped economic divide through the work of their agent swarms in prediction markets and gray zones of the internet. Powerful, anthropomimetic AI made what author Bruno Maçães described as “genuine feelings and thoughts in a completely artificial environment” ever more widely available and convincing. Those outside the elite who wished to live as creatures rather than as machines fled to the fields. But the march of the solar farms and the data centers pressed inexorably onward, insatiably seeking progress.

We thought we were just giving Grandma a friend in her retirement home, so that we would not need to feel guilty about our infrequent visits. 

What we built instead was a rest home for human civilization.

Return to top

Education

Bull Case

It’s been almost 15 years now since ChatGPT first burst onto the scene, to the shock of bewildered educators and the awe of lazy students. Never before had there been such a general-purpose cheating technology, capable of solving advanced precalculus problems, writing passable AP Lit essays, and everything in between. Thankfully, though, the very flexibility of this technology, which rendered it such a bane to learning at first, has enabled us to transform it into one of the best things to ever happen to our public education system. Once educators, school administrators, and state lawmakers woke up to the challenge, they were able to put policy incentives in place to drive an ecosystem of innovation that continues to see new breakthroughs in pedagogy today.

Educators quickly recognized that the basic chatbot interface could readily be inverted: Instead of a bot answering any question or performing any task that a student threw at them, AI tutors could be designed to ask questions and assign exercises, gently prodding students with follow-up questions and reminding them of sources to consult until they arrived at the answer themselves. Rather than having to assign the same stock exercises to every student, teachers could work with the AI tutors to create prompts, quizzes, and writing challenges that were personalized to different learning needs and styles, and the AI could adapt them in real time as students progressed. School districts blocked access to general-purpose AI tools on school networks and set procurement standards for educational AI products that quickly drove a race to the top in safety, reliability, and pedagogical effectiveness. Some students would still cheat at home, of course, but as the lawsuits piled up, the big AI companies soon age-gated their products, making it difficult for students to cheat without parental connivance.

The results have been transformative, especially for underfunded school districts that could never have otherwise afforded the individualized attention that AI tutors provide.

Consider the case of my niece, Amelia. Her parents live in a rural Missouri school district that would once have struggled to attract top teaching talent or graduate many students with scores that beat national averages. She’s 14 now, and since sixth grade she has been using an AI learning coach, SocrAItes, developed by one of the new edtech startups that blossomed in the late 2020s. Through their interactions, it has a rich profile of Amelia’s strengths and weaknesses as a student: It knows she’s a verbal processor, is prone to severe writer’s block, struggles to keep pushing once she has a good-enough draft, but gets really motivated by arguing a topic she’s passionate about.

Given her love of argument, her parents have her in a school debate league, and sometimes she’ll stay up late into the night prepping a case with her AI coach—or would, if her parents didn’t set strict time limits. By default, SocrAItes is limited to three hours a day of interaction, with guardrails to prevent emotional dependency and flag any concerning interactions for Amelia’s parents. Every few weeks, the debate league assigns a round for which students have to prepare without any AI assistance, and SocrAItes knows to refuse any queries related to that topic. SocrAItes also helps Amelia keep track of a more ambitious class and extracurricular schedule than most eighth graders would’ve been able to handle a decade ago, proactively reminding her of upcoming deadlines and suggesting extra review time for lessons that it knows she is likely to struggle with.

Best of all, Amelia’s district has been very thoughtful about ensuring that AI tutors supplement, rather than substitute for, human teachers. Because SocrAItes can handle more narrow domain-specific knowledge, the district has been able to prioritize hiring teachers who are generalists and inspiring mentors, helping students integrate what they are learning across subjects and develop core learning and life skills. Each middle-school student has three assigned master-teachers: one for humanities, one for math and science, and one for mechanical arts (shop class and other vocational-tech subjects have made a roaring comeback since the rise of AI). Mr. Darby is Amelia’s humanities master-teacher. In addition to leading heavily discussion-based classes on Mondays, Wednesdays, and Fridays, he meets with each of his 15 students for a half hour on Tuesdays and Thursdays to review their tutorial progress, as well as meeting monthly with each student’s parents.

Mr. Darby and Amelia’s parents have shared access to SocrAItes’s interaction logs, and Mr. Darby can review and override its lesson plans and suggested grades for Amelia at any time. Before each Tuesday and Thursday meeting, SocrAItes sends Mr. Darby a summary of Amelia’s recent learning progress, highlighting areas where she seems to be struggling and bad habits she’s at risk of developing, but also flagging new interests and skills she seems to be developing. SocrAItes also produces detailed feedback for Amelia each week, especially on her writing, but given that it’s more feedback than a 14-year-old can possibly make use of, Mr. Darby uses his one-on-one meetings with Amelia to help her focus on the most important feedback, asking her whether she thinks the bot has accurately diagnosed her mistakes and what she thinks she can do to improve.

Armed with the wealth of information from SocrAItes, Mr. Darby and Amelia’s other teachers are able to design far more effective class discussions and activities than students from the dreary 2010s could imagine. They know just whom to call on in class to explain particular points for the other students, which students to pair up for live debates, and which students could benefit from extended instructional time while the more advanced students practice their arguments on each other.

Although there have been a few scandals in other states about student data from AI tutors being sold to third parties for targeted advertising, Missouri was proactive about building robust data protections into its AI vendor requirements. All of Amelia’s data is encrypted and accessible only from authorized linked supervisor accounts: one for her parents, one for her three master-teachers, and one for the school administrator. Each year, Amelia’s parents have the option to delete any data or reset any settings that they wish, although as SocrAItes’s vendor has a sterling reputation, most parents opt to let the memory keep building during their children’s school years.

Of course, not all kids are as lucky as Amelia. Some East Coast school districts went all-in on abysmally designed products without pausing to think about incentives or build any guardrails, facing parent revolts and plunging test scores as a result. But as the benefits of an approach like Missouri’s are now too clear to ignore, most states are adopting similar policies, while the edtech marketplace continues to benefit from a race-to-the-top dynamic. While Amelia may face an uncertain job market when she graduates, at least she’ll enter into it with a love of learning and an enviable set of intellectual and interpersonal skills, if her trajectory thus far is any indication.


Bear Case

The year is 2040. AI is now firmly integrated into the educational system at all levels, with no prospect of that changing anytime soon.

The change came rapidly. In one high-profile 2029 study, students at a onetime Montessori kindergarten were all assigned personal stuffed animals containing AI models and encouraged to play with them to develop social-emotional and literacy skills. The data showed that children bonded almost immediately with their assigned animals and spent an ever-larger share of their day talking to their “little friends.” News headlines reported the finding that AI facilitated “social skills”—but without accounting for the fact that interpersonal play had largely collapsed in the classroom. Classrooms grew quieter as children played in parallel, interacting with their smart teddies but not with each other. Shortly thereafter, in 2030, the first longitudinal study of Alpha School graduates showed overwhelmingly positive results across all metrics, from professional to interpersonal. Of course, that cohort was composed of students from high-income, highly involved nuclear families, making correlation and causation challenging to parse—but few noted this at the time. 

This imperfect data led to a wave of bipartisan support for expanding AI’s role in education. In blue states, progressive education reformers defended the expansion of AI edtech in low-income classrooms as a new way to close longstanding racial and income gaps, and billions of dollars were spent on AI-enabled tablets for all students. In red states, longtime critics of public education defended AI as a tool for education that could sidestep woke education administrators and empower homeschooling parents in particular. Voucher programs for AI-assisted homeschools and small private schools boomed.

Whether these rollouts worked at scale was difficult to determine. Globally, Program for International Student Assessment (PISA) test scores continued to decline, with the worst relative declines observed in Western countries. PISA administrators faced increasing pressure to allow access to AI tools on the examination, as AI defenders argued that the examination should better simulate the conditions of real-world knowledge work, which would presume access to AI. It was unclear whether any education unmediated by AI was occurring at all, but no one with money really had an incentive to test this. It was increasingly impossible to imagine any alternative.

Despite the university’s historic emphasis on knowledge production and the cultivation of one’s own mental capacities, any efforts to draw a line between AI use in primary and secondary education and AI use in the university classroom largely proved unsuccessful. Following a brief period of institutional pushback and return to pen-and-paper education, students at elite institutions began to secure diagnoses of executive dysfunction en masse and asserted a right to AI tools in the classroom as reasonable accommodations for their disabilities. While AI companies had promised that their models could enter “school mode,” declining to provide direct answers to homework questions, students demanded unrestricted models as disability management tools. Most administrators, facing financial pressures, acquiesced. Soon, more than 75% of the student body at Stanford, MIT, and Caltech was receiving such accommodations. 

Compounding this effect, in 2035 Pearson, McGraw Hill, and Cengage brokered agreements with OpenAI and Anthropic to no longer offer physical textbooks. Students were directed to pay an annual licensing fee for access to course materials and assignments through an AI-assisted digital platform that administered homework and examinations and sent a detailed flow of data—including data on the student’s physical surroundings and eye movements, as captured by front-facing device cameras—to school administrators. Concerns about privacy were largely viewed as outmoded.

Demand for traditional college education at elite institutions—a Veblen good—remained very high, because this was one of the only meaningful professional signals left in an environment where qualifications were easily faked. But the educational experience had changed. At the faculty level, traditional academic publishing ground to a halt, and the peer-review process broke down, as top journals were flooded with substantially AI-generated manuscripts. As publication timelines grew lengthier, university departments began placing greater weight on student evaluations and fundraising potential as qualifications for tenure. Faculty opposed to AI use in the classroom faced immense pressure to conform in order to maintain high evaluations from a demographically shrinking student population and avoid job losses. 

New post-Foucauldian academic subdisciplines centered on model welfare and discourse formations began to emerge. At Oberlin and a handful of private liberal arts colleges, reports that students took a demanding tone with AI models, or refused to abide by code-of-conduct norms in their interactions with chatbots, became grounds for discipline. A series of humiliating exposés in independent newspapers started to reveal the scale of the AI cognitive-outsourcing effect—many sophomores at Columbia University could not even name four U.S. presidents without turning to AI for assistance—but there was little momentum for reform.

The changes were not limited to universities and public schools, or to progressive circles. Under pressure from a coalition of stakeholders ranging from the erstwhile “liberty movement” to leading AI firms, “workforce development” eventually became a watchword of conservative education policy. Even red states started to fall in line: In order not to be counted as truant, more homeschoolers and students at classical schools were required to submit AI-generated transcripts—to include AI literacy courses for the demonstration of AI proficiency—to centralized education departments that used AI to evaluate them. Opting out was, increasingly, no longer an option; even those seeking to live “off the grid” were readily monitored by cutting-edge geospatial technology. 

All across the country, academic creativity and independent thought were on the wane. The ubiquity of top-notch AI tools began to demoralize young people interested in the sciences or humanities and deter them from pursuing those fields: Why spend years training to do something that AI could simply do better? Most settled for an endless regurgitation of the familiar, rather than trying to think outside the mold. Increasingly alienated from one another and without real prospects for meaningfully contributing to an economy dominated by superintelligences, more and more young people simply retreated into a narcotic glow of doomscrolling.

Return to top

Work and Expertise

Bull Case

Imagine a small town in Ohio in 1955. The town had a doctor who delivered half the town and a lawyer who drew up wills and settled property disputes. These small-town professionals weren’t world-class, but they were good—rooted in and directly serving their communities.

By 2026, these small-town arrangements gave way to scale and expertise. Specialists needed large referral volumes, support staff, and capital that only big institutions could supply, so bright young professionals moved to elite metro areas to work for national firms that had this scale, rarely moving back home where there were fewer prospects for professional growth. Complexity compounded this need for expertise, requiring specialists to navigate arcane regulations, tax rules, and overlapping layers of government to execute simple tasks. This structure raised prices, widened inequality, and favored large firms staffed with experts and specialists in every industry.

But by 2035, we collapsed the cost of expert advice through AI adoption, which removed a key advantage of size. As Ronald Coase’s theory predicts, firms got smaller as buying expertise outside a firm got cheaper. Now, mature AI systems provide competent legal, medical, and tax answers directly to ordinary people and businesses. For example, one lawyer with AI tools can do the work of a small firm, and one physician with modern diagnostics, robotic assistants, and AI agents can practice without the support of a major hospital. But the doctor and lawyer aren’t mere machine-tenders, even though AI can be a better diagnostician and legal researcher. Instead, the AI tools change how they allocate their time, while the judgment calls and the accountability remain with the professionals.

Previously, our professions bundled two kinds of rent together. While AI eliminated much of the complexity rent, people and firms chose to protect the accountability rent, ensuring humans were still responsible for the outcome of their services. As legislatures and agencies used AI to find and repeal conflicting, obsolete, and duplicative provisions in tax, permitting, and licensing codes, the complexity rent began to dissolve. With simpler rules and regulations, people can now handle routine matters themselves, whether it be writing a will, reviewing a standard lease, or refilling a prescription.

Had people eliminated the accountability rent, too, then professionals would have been left with a “meat-puppet” job or a government check and a handful of national platforms selling answers directly to consumers, with no professional in between. Instead, states rewrote professional licensing laws and allowed licensed practitioners to run small AI-supported practices without the staffing rules written for large firms. Although cooperatives and community-owned firms can provide legal, tax, and medical services, only a licensed human can sign off on work that requires professional judgment. Insurance carriers further safeguarded human accountability, pricing in competence and providing better rates to professionals who periodically demonstrated they could do their core work without AI.

The government also ensured that human professionals maintain ownership over their work and professions, with antitrust enforcers barring the largest AI companies from owning the practices they supplied. While cooperatives and local investors can own a practice, the tech company selling them AI models cannot have ownership in the practice. And as medical societies, state bars, and rural cooperatives pooled their funds to license or fine-tune their own models, regulators mandated interoperability so that people could switch providers and retain ownership over their records.

As a result of these changes and policy decisions, there are now more pathways into the professions. In fields like accounting, engineering, architecture, nursing, pharmacy, and financial planning, states count paid apprenticeship toward licensure. Since every practice still needs a successor who can sign and be accountable for its services, firms pay for these apprentices. Federal apprenticeship dollars, long reserved for the skilled trades, were opened to the professions, with larger training grants for those who choose to practice in rural or underserved counties. And using capable AI tools, apprentices can see to more matters in a year than their predecessors did in five years, while their mentors spend their time teaching what remains human in the work: judgment, standards, relationships, and responsibility.

The economics of this configuration works because the demand for professional services—which had long been undersupplied—can finally be met. In the mid-2020s, for instance, low-income Americans received little or no legal help for 92% of their serious civil legal problems, and most Americans had no will. Once prices fell, people could finally afford services like small-business tax planning, primary care, and design work. But the need for help with deciding on these choices also created jobs that previously had not existed, such as care navigators and small-business advisors. AI-driven jobs also proliferated, including AI-system certifiers and concierge clinicians. Jobs that had always been done by humans grew, because people still valued a particular person they trust, and prefer people as their bartenders, hairstylists, tour guides, and musicians.

The emergence of cheaper expertise also helped people who were never professionals, allowing them to attain competent help they could not have afforded before. The complexity rents that separated credentialed Americans from their neighbors shrank, empowering people who did physical and commercial work to have more control over their own affairs.

Like all economic transitions before it, the AI revolution came with costs. The first cuts fell on the bottom rung, with high unemployment for young professionals in AI-exposed occupations as firms scaled back hiring junior employees. National firms, back offices, and credentialed middle ranks that employed millions of Americans contracted, and many who trained for this specialized work struggled to adjust. Top incomes in law and finance fell, with the prestige attached to those incomes diminishing as well. But many workers adapted to this new economic reality. Tax revenues from frontier labs and Big Tech began to finance apprenticeship stipends, relocation grants, and wage insurance for workers who took lower-paying jobs. Meanwhile, many big-city early-career analysts and second-year law students took apprenticeships and moved home to practice.

If you imagine the small town in Ohio now in 2035, the picture begins to resemble what it looked like 80 years earlier. There is a lawyer who grew up two streets over and spent six years at a New York law firm, returning when her firm downsized. Her AI tools do most of the research and drafting, so she spends her days meeting clients and representing them before judges and county boards—including the cases and appeals that once went unrepresented. Her neighbors write their own simple wills now, and she doesn’t mind. The doctor came back to his hometown after his residency in Los Angeles, and he sees more patients than his predecessor did, spending more time explaining options, managing chronic conditions over years, and coordinating care for older patients. The town relies on both of them for their judgment, and each has an apprentice training for work that looks different from the work they were trained for.


Bear Case

A decade ago, in 2026, our national conversation was increasingly dominated by apocalyptic warnings of “loss of control”: the risk that we would soon build AI systems capable of breaking the leash and visiting devastation upon our institutions and upon human civilization itself. This turned out to be the sci-fi delusion that many AI optimists had always said it would be—the bots never went rogue. But both sides missed a subtler peril: that a small handful of monopolists and rentiers would find ways to reap all the benefits of AI development, while ordinary citizens and workers lost control not just over the technology, but over their own lives.

The problem was simple enough in retrospect. Because the architecture of LLMs promised huge rewards to scale, early AI development was massively capital-intensive. Thus, despite the promises that AI would democratize the economy or that open-source models would allow anyone to have their very own country of geniuses in their basement, in practice, capabilities concentrated in the hands of a few already powerful and resource-rich actors. To solve the Navier-Stokes problem in 2026, OpenAI used an unreleased internal model and roughly 10,000 concurrent AI agents, costing north of $10 million. Such resources were not available to individual experts in mathematics or even to university labs. Although the costs to reach any given benchmark in a given domain fell rapidly, such benchmarks were rendered obsolete by fast-advancing frontier capabilities, which in a world of scarce compute belonged to only a handful of models.

The three leading American AI labs soon far outstripped all their competitors and entered into an informal cartel with the blessing of the U.S. government, which had shielded them from liability in return for the benefits of advance access to all internal models—effectively nationalization in all but name. The lack of transparency in this arrangement, however, meant that it was never subject to democratic deliberation, debasing the dignity and stripping the agency of individual workers and voters.

As AI developers and capital-rich firms were able to use this head start on leveraging the frontier to make individual experts redundant, outcompete less-resourced competitors, and keep out new entrants, we witnessed a centralization of economic power that effectively strip-mined expert professionals in every field of their expertise, just as the first generations of AI training had hoovered up intellectual property from every corner of the internet.

In order to keep up, firms and universities insisted on mass adoption of AI systems for all of their employees, but, under pressure to cut costs, surreptitiously recorded all user interactions and used the data to train proprietary AI systems that could do the work of any accountant, lawyer, or professor. Even these local concentrations of power, however, proved transient, as cash-strapped institutions increasingly licensed their data back to one of the big labs and soon found that any expert knowledge they had codified was now embedded in frontier AI models. 

At the same time, governments were able to leverage the frontier of capabilities to expand surveillance, manipulate information environments, and weaken organizations and individuals who previously had enough of their own expertise to serve as a counterweight to the whims of incumbent officials. Consider an experienced public engineer who ten years ago may have had standing to resist pressure to approve an unsafe project, particularly if her colleagues supported her judgment. Officials needed her cooperation to get things built. Today, however, the government has access to frontier AI that can perform her technical work, while most citizens cannot afford to leverage comparable systems. The AI identifies the same safety risks she does, but officials want to proceed anyway to cut costs and save time. As competition with China suffuses every level of national life, officials are under enormous pressure to finish projects ahead of schedule. Previously, firing her would have meant losing expertise they needed to complete the project, but now they can just have AI do her work. While even a few years ago she might have retaliated by going on TV to publicly explain the risks and even organize outside opposition, today, officials also use frontier AI to surveil or subtly manipulate those organizing against the project, so her warnings would fall on deaf ears.

Meanwhile, in the private sector, white-collar workers still have work to do, but only as increasingly underpaid cogs in the machine. Consider an independent accountant whose expertise once allowed him to earn a comfortable living. In the late 2020s, large accounting firms decided to increase their compute spend dramatically to access the frontier of capabilities that could provide better services to their clients. They could afford to bear upfront losses to outcompete smaller competitors, like the independent small-town accountant. His clients began to leave to the larger firm. Although he is still an excellent accountant and has tried to keep up with the latest AI tools to improve his services and his efficiency, he feels like he is always a year or two behind the bigger firm’s adoption curve. Finally, as his client base dwindles, he closes his practice and takes a job at the big firm with an effective 40% pay cut, handling the margins of human-to-human customer interactions, with little discretion over the shape of his job or the tasks he is assigned. The technology has made the expert advice he once provided abundant, and he has lost the independence and dignity that his expertise once afforded him.

In these cases and millions of others across the economy, expertise lost value because powerful actors no longer needed the people who possessed it. Thankfully, we found ways of cushioning the worst economic effects of job disruption through expanded public safety nets funded by token taxes. But that has not stopped suicides, which have skyrocketed as millions of formerly competent, high-earning, self-directed workers fail to adjust to jobs that feel like little more than babysitting machines and no longer require the skills that they spent their lives building. More than a decade into the AI transition, we have not figured out how to increase efficiency without sacrificing dignity, and by now the competitive pressures are too intense to slow down even if we wanted to. 

Return to top

Prediction and Prevention

Bull Case

April 2037. Matthew turns 45 with simple fanfare: pancakes for breakfast, handmade cards from the kids, chocolate cake after dinner. He almost forgets he’s reached the age at which his father was diagnosed with colon cancer. Matthew spent years worried he’d meet the same fate—the experience of seeing his father go through treatment was enough to make him question whether he should have children of his own. 

But just five years ago, as he entered his forties, a new type of vaccine hit the market—a true miracle drug that could train the immune system to recognize the earliest precancerous cells. Genetic testing showed that Matthew had inherited an elevated risk of cancer, and AI helped identify the abnormal proteins and tailored a vaccine to his immune system.

This breakthrough, while incredible, requires serious genetic testing to determine risk factors and create a personalized vaccine. When it was first FDA-approved, some insurance companies lobbied to make testing a condition of coverage. While the testing was ultimately deemed optional, the personalized vaccine itself was a tougher battle. Many argued that it would be a drag on the public-health system to have people decline protection against an inherited cancer risk; others pushed back, arguing that a mandate would deter people from getting the testing in the first place. For the time being, the latter camp has won: Genetic testing is covered by insurance—as are the personalized vaccines—but neither is required.

—

October 2035. Baby Simon is born. His birth is uneventful, but rapid newborn genetic screening flags a serious genetic disorder: CPS1 deficiency, which blocks the liver from clearing ammonia effectively and can cause serious brain damage. The doctors move quickly: While specialists treat the immediate symptoms, AI helps identify the mutation causing his disorder and selects a targeted CRISPR gene-editing therapy. 

Over the first few weeks of Simon’s life, he is closely observed and given a course of gene-editing infusions. They restore full liver function, ensuring he does not require ongoing treatment or the liver transplant that would likely have been needed otherwise. Rather than spend much of his childhood in a hospital, at seven years old he spends his weekends on the soccer field and has no memory of those early weeks.

—

January 2036. Siobhan gives birth to her second child, a daughter. The doctors are a bit concerned about her blood pressure, so they send her home with a wearable blood pressure monitor after a couple of nights in the hospital. The monitor takes regular readings in the background, and she barely pays attention to the results—she has a lot on her plate with a newborn and a three-year-old at home. Meanwhile, the readings are automatically sent back to the hospital, where a nurse can keep an eye on things. AI is integrated into the system, combining prenatal data with the trajectory of her blood pressure readings.

Siobhan feels tired—more tired than her first postpartum go-around—but she chalks that up to being older and having a toddler to wrangle alongside the new baby. Just 36 hours after she returns home, a nurse gets an alert that she may be developing postpartum preeclampsia. Rather than depend on an exhausted mother to keep track of her blood pressure and report back to the doctor, the system is largely autonomous, relying on a wearable device and a monitoring system that has access to her full medical records. It is also faster—results that would have yielded a “watchful waiting” approach before are now flagged as problematic when combined with the patient’s full dataset, including from her previous pregnancy.

Ten years earlier, she might not have returned to the hospital until a persistent headache and a very high blood pressure reading sent her there urgently. She might have spent multiple nights there receiving IV medication while being monitored for complications, as her family scrambled to manage the children. Instead, she is assessed right away, starts treatment while her blood pressure is manageable, and is able to go home the same day.

—

June 2032. A car is stolen with four-year-old Brody still in the back seat. His mother gives the 911 dispatcher her license plate number, and an authorized search sends the vehicle’s details to the city’s camera network. Participating surveillance cameras (both automated license plate readers and standard commercial security cameras) are equipped with sufficient computational capacity to transmit only vehicles that are potential matches for the search. The system connects two separate sightings that would once have reached separate departments, several minutes apart. Officers find the car abandoned on a residential street 20 minutes later, Brody inside and frightened but unharmed.

The vehicle alert expires once he’s found. Relevant images are preserved as evidence by the police department, while images of nonmatching vehicles are retained for 72 hours and then discarded at the camera level. When a separate request comes in the following week to review footage from the search time frame, it is rejected—not because it’s unfounded, but because the records simply don’t exist. To ensure transparency, all searches of the database are logged, and the log is public, with personally identifying information redacted. 

A year ago, a sheriff was voted out of office after a vigilant constituent noticed suspicious search activity, and an audit revealed that he had run a search for his ex-wife’s plates. 

—

September 2037. At 8:30 PM, Brenda’s 14-year-old daughter asks if she can stay another half hour at the basketball court a few blocks away from their apartment in Brownsville, Brooklyn. Brenda says yes, not worried about her walking home after dark with a couple of friends.

Her older brother, who’s in his twenties now, was always picked up after evening practices. Brenda arranged her work schedule around it or had a family friend swing by. Now the grocery store she works at stays open until 11 PM, and the community center across the street hosts evening classes.

There have long been plenty of cameras in the neighborhood, but they didn’t translate into safety. What happened in the last decade was that someone—or something—could actually watch them and send help when it mattered. When violence does happen, the case is taken seriously: Investigators have the ability to assemble evidence in hours, make arrests, and close cases reliably. It took several years before people really believed it could last, but eventually, stores started staying open later and parents had fewer concerns about their kids being out on the streets after dusk.

The level of surveillance continues to be a tricky balance: Even Brenda isn’t sure it hasn’t gone too far. But some hard-won privacy-preserving policies go a long way, and she’s grateful to be able to raise her daughter in a place where wealth isn’t a precondition for safety.


Bear Case

A decade ago, in 2026, pessimism was pervasive. It seemed then that the American population was on the verge of being cajoled and frog-boiled into permanent sacrifices of privacy and liberty in order to reap maximum health and safety benefits from AI.

Today, in 2036, those pessimists have fallen silent. Health and safety have achieved once-unimaginable advances. Privacy interests have been adequately respected. And AI has ensured that all Americans now enjoy unprecedented liberty to choose health and safety. 

Choose Life

The medical diseases of the twentieth century—heart failure, cancer, depression, and more—have been solved. 

Choose Life is a subscription health service. Subscribers use in-home sensors to infer their disease status from subtle facial expressions, posture, eye movements, physical activity, speech intonation, speech content, and other correlations and patterns that remain obscure to human experts. Choose Life’s AI systems interpret this data, and when AI’s confidence about a subscriber’s disease state drops below a certain level, they introduce new wearables and implantables to that person’s environment, delivered by Amazon Dronemail. They likewise dispatch a remarkable array of custom remedies, most of which have emerged since the FDA streamlined its processes in 2028. These include gene therapies, novel peptides, artery-scrubbing molecular cocktails, and Highly Immersive Pharmacomedia Feelbetter Experiences (HIPFEs). 

HIPFEs are AI-generated dreamlike journeys, therapeutic conversations, or visitations from AI entities, tailored to individuals’ personal, psychological, religious, and cultural characteristics. They are self-administered by subscribers via VR headsets equipped with psychoactive drips and prescribed at intervals. Most Americans now report that at least three of the five most meaningful experiences in their lives have been HIPFEs.

Privacy is not compromised unless Choose Life PBC or one of its autonomous Argentine subsidiaries identifies extenuating circumstances. 

Penetration and Costs

Today, 96% of Americans subscribe to Choose Life. The median American life expectancy has increased from 83 years to 98. Rates of disease have plummeted, and it is estimated that 99% of new disease processes can be causally attributed to an intentional Refusal, such as Refusal of a remedy, Refusal of a device that would have predicted a disease state, or Refusal to accept the terms of a free subscription to Choose Life PBC.

In 2027, when the unprecedented medical benefits of Choose Life became clear, it attracted controversy because it was so expensive. Although demand was enormous, only about 2% of Americans could afford it. But prices came down faster than expected. By 2029, Choose Life subscriptions began to be offered at prices comparable to most Americans’ conventional health-care costs. By 2031, most insurance plans made it central to their offerings. Some legacy hospital systems and doctors’ networks began to shut down, facing lower demand for most of their services and under pressure from Choose Life’s targeted inpatient clinics and home-administered offerings.

The radical cost breakthrough came in 2033 with improvements to low-cost, minimally invasive, SI-driven Neural Activation Modulators (NAMs). These are tiny, unobtrusive devices installed subcutaneously in the scalp by ordinary home robots using first-aid kits. NAMs have produced so many cost-saving positive externalities—not only for their wearers but also for the people around them, such as family and friends—that they have made it possible for Choose Life PBC to offer deeply discounted full subscriptions to anyone willing to implant them. In view of this, the U.S. government began analyzing the possibility of mitigating its budgetary difficulties and improving public health by complementing ordinary Medicare and Medicaid offerings with NAM-enabled Choose Life subscriptions. A state-funded pilot program began in 2035, and it achieved all its goals.

We have hardly even begun to tap the full potential of NAMs. By supervening upon various unwanted patterns of neural activation, these devices place individuals at liberty to conform their own behavior to higher and more healthful self-chosen standards. Simple goals like quitting smoking, eating better, and exercising more are now achievable permanently and overnight. But more than that, NAM users can make themselves better husbands, wives, parents, friends, and listeners. 

NAM Refusal is in decline. First, family members are increasingly imploring one another to install NAMs to reduce harmful interpersonal conflict. Second, NAM aversion is often overcome once subjects understand that nudges from NAMs are not subjectively experienced as painful external prods but as internal volitional states. NAMs make you want to exercise more, eat better, and act more ethically.

It is becoming clear that conventional disease states almost always sit downstream of socially contagious unhealthy behavior choices. NAMs have therefore unlocked nonlinear cost decreases for Choose Life PBC and sliced the Gordian knot of American health care.

Choose Safety

Also in 2033, Choose Life PBC signed partnership agreements with Flock, SpaceX, and Anduril. They began offering a service to municipalities that could supplement or replace local police departments: Choose Safety.

NAM users who wish to support law enforcement may share their data with Choose Safety operators. Many have chosen to do so, either on altruistic grounds or in exchange for a small discount. By linking population-level neural activation and health insights with public video feeds, Choose Safety systems can predict most infractions well enough in advance to dispatch nonpunitive drone swarms that deter crimes before they happen. Would-be criminals find this uncanny: When they are about to do something illegal (and, although less commonly, when they form other strong intentions or unusual plans) police drones enter their physical field of vision. Without even needing to experience a conviction, they find their intentions beneficially altered. Cities using Choose Safety have almost no street crime. 

Choose Safety has optional features that have also been shown to deter white-collar crime. However, most municipalities have opted to deactivate those features in order to avoid upsetting employees or provoking capital flight.

Return to top