It’s a compelling framework. It’s also wrong about the mechanism.
Heinlein’s Mr. Dubois lectures his students that previous democracies collapsed because voters had no skin in the game. Universal suffrage without sacrifice produced electorates that voted for comfort over survival. The Federation’s solution: make the franchise expensive enough that only the committed bother to earn it. The implied logic is selection. Service filters out the unserious.
But Gordon Allport’s contact hypothesis, the most replicated finding in social psychology, suggests something different is happening in Heinlein’s Federation. Allport found that intergroup contact reliably reduces prejudice when it occurs under conditions of equal status, shared goals, and institutional support. Military service under compulsion satisfies all three simultaneously. You cannot maintain tribal contempt for someone whose competence keeps you alive. It’s a claim about network topology, not character.
Service doesn’t test whether you’re worthy. It changes who you are. Not ideologically; no one converts to some centrist consensus. But people who serve alongside people they would never have chosen stop sorting the world into tribe and enemy.
Countries that maintain compulsory service should show lower tribal polarization than those that abandon it, and the effect should track contact specifically rather than some broader civic-duty signal.
Chagai Weiss, in a study later presented at Stanford’s Freeman Spogli Institute, tested exactly this across fifteen European countries. He found that compulsory military service measurably reduces affective polarization by preventing early partisan sorting: conscription interrupts the process by which young people increasingly live, work, and socialize only among ideological peers during the years when identity is still forming. Men exempt from service showed significantly higher affective polarization than those who served, even after controlling for ideology. A separate Erasmus University study found that conscription exposure reduced susceptibility to extreme right-wing populism, not through ideological conversion, but through reduction of group parochialism.
The American timeline is suggestive. The draft ended in 1973. The partisan sorting curve begins accelerating in the late 1970s. European countries that abolished conscription in the 1990s and 2000s show polarization increases on roughly a fifteen-to-twenty-year delay, consistent with a generational cohort aging through without the contact experience. Countries that maintained universal service; Norway, Switzerland, Finland, South Korea; show notably more durable national cohesion under stress.
Other forces were at work in the same period: cable news, geographic sorting, the Southern Strategy’s long aftermath. But the pattern is specific enough, and the mechanism clean enough, to deserve more weight than it receives. The standard accounts of polarization describe the sorting. Contact theory explains the conditions under which sorting becomes possible.
If compulsory service reduces polarization through contact, the obvious question is why no one restores it. The answer is structural.
Reduced tribal sorting is a collective good. It makes democracy more functional, propaganda less effective, and the electorate harder to manipulate through identity mobilization. It is also, for that exact reason, a threat to virtually everyone who currently holds political power. Safe legislative seats depend on maximum sorting. The propaganda ecosystem, from cable news to social media outrage cycles, monetizes tribal fear. The donor class has never wanted its children sharing barracks with the working class, which is why Vietnam-era exemptions existed and why libertarian institutions like Cato oppose compulsory service with such consistency. Identity mobilization is not exclusively a Republican tool; both parties depend on a sorted electorate.
This is Mancur Olson’s Logic of Collective Action operating as predicted: concentrated minority interests will systematically outcompete diffuse majority interests because the minority has higher per-capita stakes and the majority faces a free-rider problem. Representative Charles Rangel introduced the Universal National Service Act five separate times between 2003 and 2013. It never received a serious floor vote. A 2025 Carnegie Corporation summit convened former Defense Secretary Lloyd Austin and Republican Governor Spencer Cox to push national service from aspiration to policy. It generated coverage and no legislation. The pattern holds: bipartisan support in polling, bipartisan co-sponsorship in committee, bipartisan death on the floor.
The intervention that would repair the system is blocked precisely because it would work.
Heinlein built a fictional society where service produced functional democracy and attributed the result to moral virtue. The science suggests he was right about the outcome and wrong about the cause. The service requirement manufactures, at scale, the cross-cutting relationships that make tribal manipulation structurally ineffective. You can’t dehumanize someone you’ve depended on for survival.
We had an institution that did this, imperfectly and inequitably, for decades. We eliminated it in 1973 and replaced it with nothing. The polarization curve that followed is not mysterious if you take the contact mechanism seriously. Neither is the fact that no one in a position to restore it has sufficient incentive to try.
Heinlein’s question was whether citizens should earn the franchise through service. The better question: can a democracy survive without forcing its citizens to know each other?
But what if the profession everyone is defending was already dead before ChatGPT shipped? Not declining. Not under threat. Structurally gone, with the data to prove it.
The clearest numbers come from British publishing, which has tracked author income for decades. Between 2006 and 2022, UK author median income fell 60% in real terms, from roughly £12,000 to £7,000. The percentage of full-time professional writers dropped from 40% to 19%. All of this happened before generative AI existed as a consumer product. When a 2025 survey shows 86% of authors claiming AI has harmed their earnings, those authors are describing conditions that have persisted since the mid-2000s. AI gave them a new name for it.
The mechanism was structural. Publishers consolidated and shifted resources toward potential blockbusters and viral candidates. The midlist author, the backbone of a sustainable literary economy for decades, lost advances, marketing support, and shelf space. Chris Anderson’s “long tail” theory promised that digital distribution would let niche creators find their audiences profitably. What happened was closer to the opposite: Amazon’s algorithms preferentially surfaced bestsellers, the head grew faster than the tail, and self-publishing produced enormous volume with almost no viable livelihoods. The distribution didn’t flatten. It went Pareto: a few superstars at the top, a vast unprofitable tail at the bottom, and nothing in between.
There is no lateral move. A musician whose recordings stopped generating income could pivot to touring; being in a room with a live performer is an irreproducible experience. A character actor can build a resume over twenty years that generates consistent work. An author has no equivalent. A book is a fixed artifact. There is no night-by-night version of a novel that audiences attend. The compounding career is largely gone; each book resets to zero unless it goes viral.
By the time ChatGPT launched in late 2022, the social contract between writers and the publishing industry had already been voided. Develop your craft, build an audience, earn a modest living doing serious work: that was the promise. The economy had already broken every clause.
Two things can be true simultaneously. The structural collapse of writing as a profession predates AI, and some of the specific concerns about AI are worth fighting over. Training on copyrighted work without licensing raises legal questions that courts are actively adjudicating. The flood of low-effort AI content on retail platforms is a genuine pollution problem that degrades discoverability for everyone. These are real issues. They are also not what killed the profession, and conflating them with the structural collapse makes it harder to see what has actually happened.
AI did not murder the writing profession, but it is accelerating the decomposition. The mechanism is price compression at the commodity end: AI-generated genre novels at $2.99 competing for algorithmic placement against human authors at $9.99. A race to the bottom in content valuation for anything readable but undifferentiated. It doesn’t touch Colleen Hoover. It obliterates the mid-tier genre writer who was already barely surviving on volume self-publishing.
The deeper problem is homogenization. AI trained on existing popular fiction optimizes systematically toward what has already sold. It regresses to the mean. The midlist author’s historical function in literary culture was never just commercial; it was evolutionary. Midlist writers took formal risks. They failed more often than they succeeded, and when they succeeded, they produced the next generation’s defining voices. They were the R\&D department of an entire art form. If that experimental layer disappears, replaced by AI commodity content at the bottom and proven IP at the top, the pipeline for genuinely new work dries up from both ends. The ecosystem stops renewing itself.
The anger in writing communities is misdiagnosed but emotionally legitimate. Writers are not angry about AI specifically. They are angry about the disappearance of any path that rewards the patient accumulation of craft. They got better over years. They did the work the culture told them to do. And the economy that was supposed to reward that work dissolved underneath them, quietly, across a decade and a half, with no single catastrophe to point to. AI is the catastrophe they were waiting for: visible, nameable, resistible. It is easier to fight a technology than to fight a Pareto distribution.
This is where the homogenization problem becomes your responsibility. If AI-assisted writing means prompting a model and smoothing the output, you are feeding the regression to the mean. You are producing more of what the training data already contains. The market does not need that; it is drowning in it. But if you are using AI the way any ambitious work gets written, taking formal risks the tool would never generate on its own, pushing toward structures and voices and concerns that the model has no template for, then shaping and reshaping until the result could not have been produced by the model alone, you are doing something the anti-AI crowd cannot see and the slop merchants will not attempt.
The human contribution is not editing AI prose. It is the willingness to reach for something the training data does not contain.
Writing as a practice will survive; humans need to make meaning through language regardless of economics. Writing as a profession is probably gone for good. AI showed up after the funeral and got blamed for the body.
The only question is whether what you build next is worth reading.