Jürgen Habermas died on Saturday. His death has been the occasion for several substantial and interesting obituaries. So far, I prefer Gal Beckerman’s in the New York Times.
I took a seminar on Habermas in 1988, when I was a college junior. Georgia Warnke was the professor, and I have kept her useful packet of readings to this day. Habermas crystallized my early thinking about politics and philosophy and has remained a pillar for me ever since. I discuss him in most of my books, with the most general and extensive presentation in chapter 5 of What Should We Do? A Theory of Civic Life (2022) The title of that book basically captures Habermas in a phrase. I have also recorded a 29-minute introductory lecture on him.
It is misleading to treat Habermas as a proponent of rational, civil discourse. (See “Habermas with a Whiff of Tear Gas,” 2018). I suspect that more Americans have read Iris Marion Young’s critique of Habermas (“Activist Challenges to Deliberative Democracy, Political Theory, 2011) than have read Habermas itself. The late and lamented Iris Young caricatured him in that article. If Habermas wanted everyone to talk calmly all the time, then why did he conclude his two-volume magnum opus, The Theory of Communicative Action, with a celebration of disruptive social movements?
Habermas lived so long and became famous so early that his public role is itself an interesting phenomenon. Apparently, Ronald Dworkin remarked that even Habermas’ fame is famous, and it is worth asking why someone who wrote such thorny theory occupied the position of (arguably) the most influential German thinker for half a century.
I took a whole semester course on Habermas–in English, on the other side of the Atlantic–when he still had 38 years ahead of him. That is an indication of his stature. But it does not mean that he shaped the course of history, or even of scholarship.
In Postwar, Tony Judt discusses “the demise of the continental intellectual.” On May 31, 2003, Habermas plus Jacques Derrida, Umberto Eco, Richard Rorty, and several other leading thinkers published coordinated essays against the Iraq War in distinguished European newspapers. The result “passed virtually unnoticed. It was not reported as news, nor was it quoted by sympathizers. No-one implored the authors to take up their pens and lead the way forward. … The whole project sputtered out. One hundred years after the Dreyfus Affair, fifty years after the apotheosis of Jean-Paul Sartre, Europe’s leading intellectuals had thrown a petition–and no one came” (pp. 785-7).
I am not quoting Judt today to cast aspersions on Habermas, whose work was deep and broad. I suspect that changes in media and communications have reduced the influence of serious intellectuals. Besides, Habermas may never have wanted to be the new Jean-Paul Sartre. Elsewhere, I have discussed how Michel Foucault (born just three years before Habermas) deliberately shunned the role of the “universal intellectual”; and perhaps we are better off without such people. By all accounts, Habermas welcomed criticism and learned from a wide range of responses. He modeled what he advocated: listening and learning from others. I think his work will long outlive him.
If there were a single concept from economics I’d like to see more widely understood, it would be adverse selection: the tendency for markets to sort participants into worse and worse pools when one side of a transaction knows more than the other.
The idea originates in a 1970 paper by George Akerlof, “The Market for Lemons.” Akerlof modeled used car sales in a way that revealed how, under very common conditions, no one with a good used car would be willing to sell it. Sellers know more about their cars than buyers do, and so those with “lemons” will always be more desperate to sell than those with reliable vehicles. Since buyers can’t tell the difference, they will only purchase a used car on the assumption that something is wrong with it, and pay accordingly. As Akerlof puts it, “good cars may not be traded at all. The ‘bad’ cars tend to drive out the good” (489).
This is the core pattern: asymmetric information (one side knows more than the other) leads to adverse selection (the worse risks or lower-quality goods dominate the market) which can produce market unraveling (the market ceases to function well for anyone). That three-step sequence recurs across domains far removed from used cars. In each case, the same basic dynamic creates sorting problems that individual good faith cannot solve.
It is worth noting at the outset that markets sometimes generate their own responses to adverse selection. The used car market did not actually collapse: it produced warranties, certified pre-owned programs, CarFax, and lemon laws. A free-market economist would say that Akerlof identified a problem and then markets solved it through entrepreneurial innovation. This is a fair point, and it should be conceded. But the private-solution story works best in domains where the stakes are modest and the information gap is narrow enough for reputation mechanisms to bridge. Nobody invented a CarFax for health risk that prevented the insurance death spiral. Nobody developed a private warranty against neighborhood racial transition. The domains where adverse selection is most destructive tend to be precisely the ones where private solutions are weakest: the information gaps are widest, the stakes are highest, and the affected populations have the least market power to demand better terms.
Poverty and Insurance Markets
Akerlof applies the same model to car, life, and medical insurance: the more accident-prone, close to death, or sickly you are, the more desperate you will be to have insurance. But the more these “high-risk” individuals buy insurance, the higher the payouts, which drives up the average price above what healthier, younger, better drivers would choose to pay. Adverse selection takes over, and only the high-risk individuals end up insured.
The textbook version of this story is clean, but reality is messier in an interesting way. In practice, the people most likely to buy insurance are not just the sick; they’re also the risk-averse. Healthy people who worry about everything overinsure. They buy supplemental plans, max out their coverage, and renew faithfully. This actually helps stabilize insurance pools, because the worried-well subsidize the genuinely ill. The people who go without insurance tend not to be the healthy-and-rational actors of the economic model; they’re the young and cavalier, the overwhelmed, or the too-poor-to-bother. Adverse selection is real, but it operates alongside a countervailing force: anxiety. The question for policy design is which force dominates.
The underlying problem remains asymmetric information. Insurance purchasers know more about their own risks than the insurance company does. This is why life insurers try to get as much medical and genetic information about their enrollees as possible, and why car insurance companies use accident history, residential zip code, and miles traveled to price coverage.
It is also why the Affordable Care Act was designed to require “community rating”: forcing insurers to ignore most individualized risk information and instead treat communities as a single risk pool. Several distinct mechanisms work together here. Community rating prevents price discrimination against the sick. Subsidies stabilize the pool by making coverage affordable. The individual mandate discourages people from waiting until they’re desperately ill to enroll (the same waiting game that “pre-existing conditions” exclusions were designed to prevent, and that the ACA banned in favor of the mandate and financial penalty).
This is all well-known to anyone who has paid attention to health care policy over the last decade. But naming and framing the problem this way reveals why some solutions are superior to others, why even well-intentioned proposals can recreate the very dynamics they’re designed to prevent, and why dismantling these protections is so dangerous.
We are watching this play out in real time. The 2017 Republican tax bill zeroed out the individual mandate penalty, removing the main tool that kept healthy people in the insurance pool. During his first term, Trump slashed ACA outreach and advertising by 90% and cut the enrollment period roughly in half. This is a textbook recipe for adverse selection: the sick always know where to sign up, while healthy people, the ones whose premiums subsidize the pool, are the most likely to miss a shortened window or never hear about their options. By one estimate, these changes alone kept 500,000 people from enrolling.
The second term has been worse. Congressional Republicans allowed the enhanced premium subsidies to expire at the end of 2025, and the so-called “big, beautiful bill” cut over a trillion dollars in federal health care spending. New CMS rules on enrollment verification significantly disrupted automatic reenrollment, which had kept nearly 11 million people covered, though the full regulatory picture remains in flux amid litigation and rule revisions. The results are already visible: insurers’ list-price premiums rose an average of 26% for 2026, with increases exceeding 30% in states like Delaware, New Mexico, and Mississippi. For subsidized enrollees, the picture is even worse: with the enhanced tax credits gone, what households actually pay has in many cases more than doubled. The CBO projects marketplace enrollment will drop from 22.8 million to 18.9 million, with 4.2 million Americans losing coverage entirely because it has become unaffordable. Aetna has already exited the individual market, and three insurers have pulled out of Illinois, where rates jumped nearly 29%.
Each of these moves accelerates the adverse selection spiral the ACA was designed to prevent. Fewer healthy enrollees means higher average costs, which means higher premiums, which drives out more healthy enrollees. The pool gets sicker and more expensive. Insurers exit markets where the math no longer works. The people left holding coverage are the ones who cannot afford to go without it, paying more for less, in a market with fewer options. This is what adverse selection looks like when policy actively feeds it rather than fighting it.
Consider Bernie Sanders’ original Medicare for All Act. It went a long way toward eliminating selection effects by creating a universal federal entitlement with comprehensive benefits. But even this ambitious proposal preserved a significant state-administered role for long-term care, routing it through existing Medicaid channels with maintenance-of-effort requirements tied to state spending floors. This raises a familiar adverse selection worry. Long-term care is among the most expensive forms of health provision, and the populations who need it most, the elderly, the disabled, the chronically ill, are precisely those whose costs any system has the strongest incentive to contain. By leaving the administration of those costs in fiscally pressured state channels rather than absorbing them fully into the federal pool, even a nominally universal plan can create a two-track system: comprehensive federal coverage for the majority, and more constrained, state-variable care for those with the greatest needs. The risk-averse healthy people, the ones whose overinsurance stabilizes any pool they join, stay in the well-resourced federal track. The people who need the most help end up in the track most vulnerable to cost-cutting. Partial universalism can be worse than honest market segmentation, because it borrows the moral authority of universalism while quietly reproducing the sorting it claims to abolish.
Race and Real Estate
Adverse selection is part of a broader family of sorting dynamics that appear wherever people make choices under uncertainty using imperfect information. Akerlof recognized this himself: his original paper treats racial discrimination in labor markets as a case of the same phenomenon, where employers use race as a proxy for unobservable worker quality, and the resulting discrimination makes it harder for members of the discriminated group to invest in quality, confirming the original prejudice. The mechanism is simple. The outcomes are vicious.
Racialized housing markets in the US offer a vivid example of how these dynamics interact with, and amplify, deep structural racism. No economic model captures the full weight of what redlining and blockbusting did to Black communities, and there is a risk that naming these dynamics as “selection effects” makes them sound more orderly and less violent than they were. But the analytical lens is worth using precisely because it shows how racist outcomes reproduce themselves even when individual malice is not the proximate cause.
Consider how redlining and blockbusting worked hand-in-hand to prevent market-driven racial integration. (No story about race in America ever truly “begins” where we say it does; the history of slavery, Jim Crow, and white supremacy is ongoing and always in the background.)
Source: National Archives (Mapping Segregation in Washington DC)
It began when the Federal Home Loan Bank Board tried to determine whether some neighborhoods were too risky to finance. In the midst of the Great Depression, the most vulnerable neighborhoods were those primarily inhabited by African Americans: those who suffered most during the Depression were those who suffered most generally, so foreclosures were always worst in those neighborhoods. The maps drawn by the FHLBB were used by the Federal Housing Administration to dictate underwriting to private mortgage lenders. African Americans were barred from receiving federally underwritten loans. Private mortgage companies could still lend to them, but at much higher risk without federal insurance, and so they demanded a much higher premium.
A free-market critic would note, correctly, that the worst distortion here was government-created: the FHA drew the maps, and federal policy enforced the discrimination. This is true, and it matters. But the objection proves too much. Government created the information asymmetry; the market then amplified it through blockbusting, white flight, and self-reinforcing price spirals. Removing the FHA maps did not undo decades of wealth destruction. The damage compounds, and the compounding is a market process. The lesson is not that government intervention is always the answer, but that markets do not self-correct when the underlying information environment has been broken, even long after the original distortion is repealed.
Add to this the existence of “blockbusting” real estate agents, who used the threat of incoming African Americans to pressure white homeowners into selling. The fears were partly racial anxiety, but partly financial calculation: a “busted” block would lose considerable home value as wealthier whites fled and were replaced by African American buyers who had been shut out of better-financed markets. Even setting aside the racial animus, the financial incentives alone were sufficient to drive the sorting: any homeowner, regardless of personal attitudes, faced a real threat to her single largest asset. Real estate companies made a solid business of this practice, flipping houses from fleeing whites to middle-class African Americans for decades in places like Chicago.
While blockbusting sometimes looks like the just deserts of a racist society, where white people are so afraid of Black people that they willingly sell their homes at a loss, it exacerbated the underlying dynamics of segregation. The sorting was self-reinforcing: each departure confirmed the fears that motivated the next one, driving prices further down and concentrating poverty in the newly “turned” neighborhoods. The result was decades of wealth destruction in Black communities, even as the real estate agents who facilitated the churn profited handsomely.
Today we see related sorting dynamics working in the opposite direction: white homeowners return to the cities their parents fled, and expectations of neighborhood “improvement” drive transactions that accelerate displacement. The mechanism rhymes with blockbusting (speculation on racial neighborhood change) but the power dynamics are inverted. Capital flows in rather than out, and the displaced population, largely Black and largely renters, loses not just housing but neighborhoods, churches, schools, and the social networks that make a community function. They lack the market power that white homeowners had when they chose to flee. Some people may catch a windfall if they own property at the right moment of transition, but the overall pattern is the same: sorting by race and wealth, driven by information asymmetries and self-fulfilling expectations, producing outcomes that are worse for the most vulnerable.
There is a bitter irony here. Racial integration was the goal of decades of civil rights struggle. Now a version of it is arriving through market forces, and it is experienced, correctly, as displacement rather than progress. Integration through capital is still sorting. It replaces one racially stratified equilibrium with another, and the communities that fought hardest for integration bear the costs of its market-driven arrival. This is perhaps the clearest illustration of why adverse selection problems cannot be solved by markets alone: even when the market moves in the “right” direction, it moves on terms that reproduce the underlying injustice.
Test Scores and Schools
The connection between test scores and school segregation may be the clearest example of adverse selection logic outside insurance markets. Risk-averse parents cannot distinguish between two causes of low test scores: poverty and pedagogy. They tend to prefer schools whose students test well, treating scores as a proxy for teaching quality. Schools with lower test scores lose these risk-averse parents, who transfer their children elsewhere or simply move.
This has come to define real estate markets, with housing prices in desirable school districts dwarfing those in neighborhoods whose schools have historically underperformed on standardized tests. High-performing schools gradually accrue higher-income parents, while low-performing schools gradually accrue impoverished ones. Even childless homeowners prefer to live in high-performing districts, because of the relationship between school test scores and property values.
The sorting here is self-reinforcing in exactly the way Akerlof’s model predicts. The “good cars” leave the market, which makes the market look worse, which drives out more good cars. Parents who can afford to leave do; their departure further depresses scores; the next tranche of parents leaves. The schools that remain become, in effect, pools of adverse selection.
There are “community rating” equivalents: mixed catchment areas that require students of different races and incomes to attend the same schools. But the responses must be calibrated to account for consumer behavior. Parents can read catchment maps as well as any bureaucrat, and they will relocate for their children, to the suburbs or, in the most extreme cases, to private schools. After which they tend to oppose funding for the schools they have abandoned, creating a new cycle of sorting and disinvestment.
School choice advocates would argue that vouchers and charters are the solution: let motivated families escape failing schools. They are right that geographic assignment is an imperfect tool, and that trapping students in schools that do not serve them is its own injustice. But choice programs do not eliminate the sorting; they accelerate it with public funding. The families who exercise choice are disproportionately those with the information, resources, and engagement to navigate the system, leaving behind a more concentrated pool of the most disadvantaged students in schools with declining enrollment and declining revenue. This is adverse selection with a permission slip. The question is not whether parents should have options, but whether the mechanism for providing them creates a spiral that makes the remaining schools worse for the children who have the fewest options of all.
Employment and Credentials
The adverse selection dynamic also shapes the labor market. The General Education Development (GED) test was designed to help those who failed to complete a US high school diploma demonstrate mastery of equivalent material. High school diploma holders earn more than those without one, so if the value of high school is primarily learning, then exam-certified learning ought to perform just as well. Yet as Stephen Cameron and James Heckman have shown, it does not.
Part of the explanation is adverse selection. The GED has become correlated with incarceration, and criminal backgrounds are a cause for hiring discrimination and low wages. Once a credential is associated with a negatively selected population, its market value degrades, which further discourages anyone with better options from pursuing it. The pool gets worse; the credential’s value drops further; the cycle continues.
But most GED holders did not leave high school because of incarceration. They tend to be primary caregivers, to have lost a loved one, or to have been abused. These experiences also make sustained full-time employment difficult. The credential ends up carrying the weight of every reason someone might not finish high school, and employers treat it accordingly.
In 2014, Pearson VUE tried to fix this by making the GED significantly harder: computer-based, aligned to Common Core standards, with constructed-response questions replacing multiple choice. The logic was straightforward. If the credential’s value had degraded because the pool of holders was negatively selected, a harder test should produce a more valuable credential. The results were dramatic. Test-takers dropped by more than half nationally; completion rates fell 23 to 39 percentage points across every racial and ethnic group. By 2016, Pearson had to lower the passing score from 150 to 145, retroactively awarding credentials to people who had scored in between. The stated reason was revealing: students hitting 150 were actually outperforming high school graduates in college, suggesting the bar had overshot. Pass rates eventually recovered to around 80%, and Pearson’s own research claims 45% of passers enrolled in postsecondary programs within three years. But “fewer people take the test, and the ones who do are better prepared” is a description of adverse selection doing its work, with the credential’s gatekeepers actively assisting. The people deterred by a harder test are precisely the most marginal candidates, the population the GED exists to serve.
A related sorting dynamic appears in “ban the box” policies, which prohibit employers from asking about criminal history on initial applications. There is evidence that employers who cannot directly inquire into criminal backgrounds compensate by discriminating against all African American men. The structural logic is clear enough: employers substitute a racial proxy for the information they’ve been denied. But the structural logic does not exhaust the moral reality. This is racism, leveraging the machinery of adverse selection to do its work. The policy designed to help the formerly incarcerated ends up punishing an entire demographic, and the employers making these decisions are not innocent bystanders caught in a sorting trap. They are choosing the proxy. The sorting logic is relentless, yes, but the people operating it bear responsibility for the choices they make within it.
Solutions, or: Why Universalism Matters
So far as I can tell, adverse selection and its related sorting dynamics are natural byproducts of free choice under asymmetric information. This means the solutions must address one or both of those conditions. But the menu of responses is richer than it first appears.
Correct the information asymmetry. Force disclosures that help the less-informed party make better decisions. This is what insurance companies do when they demand medical histories, and what consumer protection laws do when they require used-car inspections or home disclosures. But disclosure is a double-edged tool: the same information that helps markets function can also enable discrimination.
Limit the sorting. Mandates, community rating, anti-discrimination rules, and mixed catchment areas all work by preventing people from separating into stratified pools. The individual mandate in the ACA is a pure example: you must participate whether you’re healthy or sick, which stabilizes the pool.
Provide the good directly. Where selection effects are severe enough, the most effective response may be to remove the market mechanism altogether. Public schools, single-payer health systems, and universal social insurance don’t just limit choice; they eliminate the market in which adverse selection operates. This is the strongest argument for universalism: it doesn’t try to outsmart the sorting. It refuses to sort. A fair objection: universal systems do not eliminate sorting entirely. They push it into different channels. Single-payer systems produce wait times, regional quality variation, and private supplementary markets. The NHS has private alternatives; Canadians cross the border for some procedures. But the sorting that persists under universalism is a different kind than the sorting markets produce. Under market sorting, the excluded population is the sickest and poorest. Under universal systems, the people who opt into private alternatives are the wealthiest. A system where the rich buy faster care is a far less harmful equilibrium than one where the poor are priced out of care altogether.
These three are the standard responses. But the examples above suggest that they are not always sufficient, and that more creative interventions are possible.
Change the incentives, not the information. Rather than hiding risk information from insurers or forcing everyone into a single pool, you can let everyone see the information but make it unprofitable to act on it. Risk adjustment transfers do this: insurers who attract healthier pools pay into a fund that subsidizes insurers with sicker ones. The ACA has a version of this mechanism, though it remains underdeveloped. The principle generalizes. School funding formulas that pay more per high-need student, enough to make “difficult” students a revenue source rather than a cost center, would transform the incentive structure that drives school sorting. Instead of fighting adverse selection, you make it financially irrelevant.
Change what information is legible. The essay has framed the choice as: reveal information or hide it. But there is a third option: change which information is visible. Raw test scores conflate poverty and pedagogy; value-added models that control for demographics would give parents a signal for teaching quality that doesn’t simply proxy for income. The GED carries a stigma because it is a single binary credential that cannot distinguish a formerly incarcerated person from a teenage caregiver. Richer signals, such as portfolio-based assessment, stackable micro-credentials, or apprenticeship records that carry narrative rather than just pass/fail, could break the pooling that makes the credential toxic. Adverse selection thrives on coarse information. Finer-grained information can disrupt the sorting without requiring either blanket disclosure or blanket concealment.
Design for precommitment. Adverse selection requires people to sort themselves after they know their type: sick or healthy, in a good school district or a bad one. If you can get people to commit before they know, the selection pressure disappears. Employer-provided health insurance partially works this way: you choose a job before you get sick, so the pool is not self-selected by health status. School assignment by lottery before test scores are published does the same thing. This is the Rawlsian insight as design principle: the veil of ignorance is not just a thought experiment but a template for institutions. Where you can build precommitment into the architecture, you defuse adverse selection at its root.
Recruit against the spiral. If adverse selection is driven by the exit of good actors, one response is to make staying, or entering, actively attractive to them. This reframes the ACA outreach cuts as even more destructive than the premium numbers alone suggest: outreach was not just informational but counter-selective, specifically targeting the healthy young people whose participation stabilizes the pool. Magnet programs in struggling schools follow the same logic. So do signing bonuses for teachers in high-need districts. And so does second-chance hiring: employers who actively recruit people with criminal backgrounds, GED credentials, or nontraditional career paths are not just doing social good. They are breaking the adverse selection cycle that degrades those credentials and populations in the first place. Every employer who hires a GED holder and has a good experience makes the next employer’s decision slightly easier. Counter-selection is contagious in the same way that adverse selection is; it just requires someone to go first.
Route around the degraded pool. When a credential or institution has been too badly damaged by adverse selection to rehabilitate, sometimes the answer is to build a new pathway rather than trying to rescue the old one. This is arguably what community colleges do for GED holders, what expungement does for criminal records, and what “housing first” does for homelessness services. You are not correcting the information asymmetry or forcing people to stay in a broken market. You are routing around it entirely. The risk is that the new pathway eventually suffers the same selection effects (community colleges themselves are not immune to this). But pathway replacement buys time and creates options that pure reform cannot.
Each of these approaches has costs, and none eliminates the underlying pressures entirely. Parents still move to better school districts; wealthy patients still seek private care; employers still find proxies. But the history traced above suggests that leaving markets to sort themselves, especially in domains as fundamental as health care, housing, education, and employment, reliably produces equilibria that are worse for almost everyone, and catastrophic for the most vulnerable. And the range of available interventions is wider than the usual debate between “let the market work” and “replace the market” would suggest. The most promising designs often work with the sorting dynamics rather than against them, redirecting incentives, enriching information, and recruiting the actors whose participation stabilizes the whole system.
The point of naming the pattern is to make it harder to ignore. Once you see adverse selection, you start to notice it everywhere: in the way dating apps stratify their users, in the way adjunct hiring degrades faculty quality, in the way nonprofit funding cycles punish organizations that serve the hardest cases. The concept doesn’t explain everything. But it explains a remarkable amount about why free markets, left to their own devices, so often deliver the opposite of what their advocates promise.
Although many people are using principles of game theory to analyze the Trump/Iran war and to predict the next steps, I haven’t come across an explicit model. Any model drastically oversimplifies reality but also serves to clarify assumptions.
The model that I present is essentially a Prisoner’s Dilemma. For each side, it is better to continue deadly offensive operations than to cease, regardless of what the other side does. Therefore, the model predicts that the war will continue (bottom-right box) even though both sides would be somewhat better off with a mutual ceasefire (top-left). That’s how a Prisoner’s Dilemma works.
The model presumes that both sides have the capacity to continue offensive operations–that the US won’t run critically low on munitions and Iran will retain drones, missiles, mines, and possibly sleeper cells abroad. To the extent that the US and Israel have a plan, it is to destroy Iran’s military assets so that Iran cannot choose to continue to bomb or lay mines. I cannot assess whether this is possible, but it seems doubtful. The recent reduction in the tempo of Iranian strikes may simply reflect a strategy of operating for a longer period.
The model is symmetrical, which is misleading. The Iranian leader, Mojtaba Khamenei, has already lost his father, wife, daughter, son-in-law, and 14-month-old granddaughter in a strike and could be killed himself. More than 1,000 Iranians (and probably many more) have died so far. Donald Trump is much safer, as are American citizens–presumably. On the other hand, Trump’s political fortunes are sensitive to exactly what happens in the war, whereas Khamenei and his team are trying to survive. For them, a difference in the length of the conflict or the number of casualties may be immaterial.
Another way that the model simplifies is by reducing the whole war to two parties. Israel is not shown. Nor are other major countries, such as Russia, China, Saudi Arabia, and others. Also, the two sides are shown as if they were unitary, but there are internal conflicts on both sides. In fact, each leader may care most about the struggle with his own domestic opponents. However, to some extent, that dynamic is captured by the outcomes shown in the model. For example, each side benefits domestically from being able to claim victory credibly, and each side loses domestically if it cannot.
When you watch fish swimming around in very cold water, they look fine. Human beings have a protein, TRPM8, that reacts to cold and affects our nervous system, causing discomfort or even pain when the temperature goes down. But fish do not have any TRPM8 (Yong p. 138). Thus we can infer that fish do not sense cold in the way we do.
This does not mean that we know what cold is really like, while fish do not. Nor does it mean that our pain is nothing real, as if we can make it go away by disbelieving it. Nor does it mean that we know what it feels like to be a fish. But we can perceive a difference between species.
Long before anyone knew about proteins, the behavioral difference between us and fish was obvious enough that it served as an example for several thinkers who asked whether experiences like pleasure and suffering are subjective. More deeply, they asked what happiness is.
Japanese Zen Buddhism uses the term kyogai. Often translated as “consciousness,” it literally means “boundary” or “bounded place,” deriving originally from the Sanskrit word visayah, in the sense of a pasture that has a boundary. The Buddhist Abbot Mumon Yamada (1900-1988) taught:
This thing called kyogai is an individual thing. …. Only another fish can understand the kyogai of a fish. In this cold weather, perhaps you are feeling sorry for the fish, poor thing, for it has to live in the freezing water. But don’t make the mistake of thinking it would be better off if you put it in warm water; that would kill it. You are a human and there is no way you can understand the kyogai of a fish.
I think the upshot here is humility: if things seem and feel very different to creatures that have different senses, we cannot really know how things are. We should be compassionate, but that is harder than it may at first appear because it requires knowing what another feels. It would not be compassionate to move carp to a warmer pond. Our humility must temper even our compassion.
Aristotle wants to distinguish wisdom, which is knowledge of objective truths, from practical wisdom or phronesis, which allows us to act well. For example, “straight” (using the term from geometry) always means the same thing. The line that takes the shortest distance between two points is straight, regardless of whether any creature sees it as such–or sees it at all. In fact, a line would be straight even if there were no sentient creatures. Hence geometry is a part of wisdom.
However, says Aristotle, different things are healthy and good for people and for fish, and human phronesis involves doing the healthy thing for us, not for them. The “lower animals” also have practical wisdom because they also know what to do. If we try to convince ourselves that our phronesis is wisdom because we are higher than fish, we are foolish because there are things far more divine than we are (NE 1143a).
The upshot, for Aristotle, is that each creature has its own nature, and the proper definition of happiness is acting according to that nature. This means that a fish is happy if it swims around in the cold, not because that behavior feels good to it, but because happiness is accordance with nature. One distinguishing feature of human beings is that we can also know wisdom, or glimpses of it, by studying things higher than ourselves. Thus, for Aristotle, observing the behavior of fish does not really encourage humility. It directs us to identify our proper nature and its place in the cosmos as a whole.
Now here is a passage from Zhuangzi:
Zhuangzi and Huìzi wandered along the bridge over the Hao river. Zhuangzi said, ‘The minnows swim about so freely and easily. This is the happiness of fish’.
Huìzi said, ‘You’re not a fish. How do you know the happiness of fish?
Zhuangzi said, ‘You’re not me. How do you know I don’t know the happiness of fish?’
Huìzi said, ‘I’m not you, so indeed I don’t know about you. You’re indeed not a fish, so that completes the case for your not knowing the happiness of fish’.
Zhuangzi said, ‘Let’s go back to where we started. When you said, “How do you know the happiness of fish”, you asked me about it already knowing that I knew it. I knew it over the Hao river’. (17/87–91)
I have virtually no knowledge of Taoism or its context, so it is risky for me to venture an interpretation. But I think the idea here is that neither of the men in the story can know the other, let alone the fish, and therefore all knowledge (including of one’s self) is illusory. However, Zhuangzi was right in the first place. “This” was the happiness of fish. He could not know its content or how happiness would feel to a fish, only that because fish were being fish, they were happy.
Ed Yong, An Immense World: How Animal Senses Reveal the Hidden Realms Around Us (Penguin Random House, 2022); Yamada as cited in Victor Sogen Hori, “Koan and Kensho in the Rinzai Zen Curriculum,” in The Koan: Texts and Contexts in Zen Buddhism (2000); Zhuangzi. The Complete Writings, translated by Chris Fraser (Oxford World’s Classics, p. 200). I translated Aristotle from the 1894 Clarendon edition on https://scaife.perseus.org/, but I have paraphrased here because the literal text is thorny. See also: some basics; Verdant mountains usually walk
I am no scholar of Elizabeth Anscombe, but I really enjoy her 1958 essay “Modern Moral Philosophy.” There’s so much rich provocation in her thesis that:
…it is not profitable for us at present to do moral philosophy; that should be laid aside at any rate until we have an adequate philosophy of psychology, in which we are conspicuously lacking.
— “Modern Moral Philosophy,” Philosophy 33:14, January 1958
What makes this such a great provocation is that it tees up one of my own interests, which she treats as a second thesis but really seems logically to precede the first: we ought to jettison the language of obligation, duty, and “ought” from our moral lives, because they have a metaphysical and theological root that we don’t really share, and provide instead a psychological foundation for morality. (For a good accessible introduction to Anscombe’s argument, see this overview.)
It does seem true that there is a great deal of baggage in the ways that we think about moral life that is a holdover from a medieval Christian worldview. If we are no longer entirely devoted to that worldview then we need to provide a new framework. In large part the debate about free will and determinism is still infected by anxieties from Christian theology about predestination, sin, and salvation. The concepts of “ought” and “may” are imbued with a sense of a “moral law” set down by a “moral sovereign:” a world-authoring deity who has the authority to command us. At the same time, many other cultures without a Christian worldview nonetheless seem to have a concept of obligation or duty.
Modern secular moral frameworks can at least superficially look like they require a divine juridical concept of a sovereign legislator and judge with the power to punish those who transgress His law with tortures and confinements. The divine throne is empty, but it holds the structure together. That image deserves a second look: an entire civilization organized around an absent authority, unable to either fill the seat or dismantle it. What goes in its place varies. The contractarian just-so story is that in a democracy every citizen is a co-legislator and co-executive guided by his or her own reason and moral common sense. But of course, some views predominate, and one candidate for replacing the entire worldview of Christian theology is psychology, the scientific study of the human mind.
Anglo-American secular moral philosophers are also deeply Christian, and Western, in ways that become clearer when you study Chinese, Indian, African, and ancient Greek philosophy. Because the standard way to ask the questions of moral philosophy is “What should I do?”, the answers tend to focus on the intentions behind your action or the consequences of it. But then they strip out the eternal soul and its afterlife, they strip out the omniscient solution to Plato’s invisibility ring, they remove the easy Thomistic mapping between natural law, moral law, and human law, and they leave us wondering if our own minds can even be trusted to correctly report “what we should do.” The “law conception of ethics” doesn’t work so well when you get rid of the legislators, the judge and jury, and the punishments too. “Conscience” and “guilt” seem weak without all that: the Western moral framework has replaced God the loving and vengeful Father with Jiminy Cricket, a cartoon character from a Disney movie about a puppet that comes to life.
The Standard Alternatives Don’t Work
Anscombe takes these problems up in a kind of speedy frustration, finding each one wanting for our current era. Her particular disdain for Kant’s account of maxims is I think especially important: Kantian morality is almost incomprehensible once you realize that as “strangers to ourselves” we very rarely know what we are doing, and so we very rarely can be sure of what principle we are acting upon, and then know further whether it could coherently be universalized. The secular modern thinker is left asking what it would mean to have a moral version of mens rea when we also know that we are psychologically prone to self-justification, backdating judgments, and ignorance of our own true intentions.
The Aristotelian conception of character and virtue seemed to suggest that it was not something that everyone could aspire to: that through bad luck, immoderation, cowardice, and bad judgment we would mostly all fail to find the right role models, practice hard, overcome adversity, and choose the right goals for ourselves. This has often seemed unfair to people raised on democratic and egalitarian values, and so modern American stories about “character” and “grit” and “resilience” all ignore luck and the contextual factors that make the cultivation of character in Aristotle’s sense possible. It’s not remarkable when you think about it, but the major virtue theorists have tended to be aristocrats, and they have tended to be focused on how the wealthy and powerful can teach their children the things they need to know to wield inherited privilege sustainably. (This seems as true of Confucius as it does of Aristotle.)
Community, Institutions, and the Question That Won’t Go Away
I choose to derive a different lesson from the psychological research: community matters. Family matters. Institutions matter. We know from all sorts of histories and psychological research that if those groups are cruel, or racist, or genocidal, then we are likely to be cruel, racist, and genocidal too. So getting the institutions right matters.
But from whence comes this “getting it right”? Have I smuggled in a bit of “what should we do?” from the outdated moral philosophy of a more Christian era? Perhaps. The question is whether psychology, on its own terms, can supply what the old moral frameworks supplied, or whether it just smuggles the same furniture back in under new names.
Psychology’s Trinity
If Anscombe is right about our need for a moral framework that is written anew with an eye on the philosophy of psychology, then it behooves us to think clearly about the undergirding of that discipline. What are the concepts and ideas that psychology substitutes for God, and sin, and the divine law?
I think there are three, and the theological parallels are not accidental:
Pathology and deviance replace sin. Where the Christian framework identified transgressions against divine law, psychology identifies deviations from statistical and functional norms. The sinner becomes the patient, the confessor becomes the therapist, and the threat of hellfire becomes the threat of institutional confinement and social exclusion. The structure is the same: there is a standard you must meet, and failure to meet it exposes you to coercive consequences.
Happiness and autonomy replace salvation. Where the Christian framework promised beatitude as the ultimate end of a rightly ordered life, psychology promises subjective well-being and self-determination. Self-reported life satisfaction stands in for the state of grace, and the therapeutic restoration of agency replaces the soul’s redemption. The good life is still the goal; we’ve just traded eternal bliss for a favorable score on the Satisfaction with Life Scale.
Biases and heuristics replace original sin. Where Christian theology held that we are fallen creatures whose nature is fundamentally corrupted, cognitive psychology holds that we are systematically irrational creatures whose judgment is fundamentally unreliable. We are strangers to ourselves, prone to self-deception, moved by forces we cannot see. The doctrine of total depravity becomes the doctrine of bounded rationality. In both cases, the conclusion is the same: you cannot trust yourself.
Each of these deserves scrutiny.
Pathology and Deviance
Without belaboring the arguments of Thomas Szasz, Michel Foucault, Ian Hacking, and the Mad Pride movement, the idea that deviations from the norm are treatable conditions rather than the result of human diversity has to be the single most important story in psychology over the twentieth century. The confinement of millions of “deviants” (literally those who deviate from norms) or “misfits” (literally those who do not fit properly in their economic and social role) is a story of human suffering that continues, in altered and criminological form, to this day. Just as right and wrong once derived their authority from God’s threat of hellfire, mental illness and pathology derive their real meaning and stigma from the threat of institutional confinement and social and economic exclusion.
The idea that there are still places where teenagers are locked up and “treated” for their deviant behavior, so long as the parents approve, fills me with dread and outrage. The fact that we don’t always confine the mentally ill but we have a series of social practices that exclude and demean them fills me with rage. This is not a post about my feelings, but it is a post about whether we might have become too comfortable with the common sense view that having distinctive psychological features marks one out as different and potentially justifies treatment and mistreatment with varying levels of consent. The fact that we internalize these norms and effectively stigmatize what makes us special worries me in much the same way: those judged abnormal can now find relative respite in the “normal” world if they’re willing to build themselves an asylum in their heads. Perhaps this is provocative, but I think it is also true.
Happiness and Autonomy
Here’s a simple reductio: positive psychologists study happiness through self-reporting. If we have any reason to think people are not particularly insightful into their own mental states, the whole edifice wobbles. And we have many such reasons. Anscombe’s discussion of the incoherence of the conception of pleasure is roughly parallel, but the quickest way to see the problem is this: if our goal is to increase pleasure and happiness, why are we so worried about the addictive drugs that seem to target pleasure directly? The answer we give is that we care about autonomy, not just pleasure; that psychologists want to restore the agency their addicted clients have lost. But we are less likely to endorse the second-order volitions (the desires they want to have) of someone who has different values than us, and the state still reserves massive coercive power for drug users whose pursuit of happiness takes unapproved forms. (Research on the “true self” confirms this: we attribute authenticity to the desires we approve of and treat the rest as alien impulses.) Our conceptions of agency and freedom aren’t coherent enough to bear the weight psychology places on them.
Biases and Heuristics
So then we are thrown back on the more fundamental thought that people are strangers to themselves: that while we often know what we are doing, and can be forced to give reasons for it, those reasons are often not the true reason, and in any case we do not know what it is that what we are doing does, what effects it predictably has beyond the ones we’re willing to own. The modern heuristics and biases literature under Kahneman and Tversky has been fascinating and important, but it owes a great deal to the older Freudian theories of the Unconscious and its drives, and it leaves much in Freud and Jung that we should perhaps excavate again.
What Remains
After we have imbibed the suspicion of humanity’s self-justifying psychology, what is left? How can we continue our practices of praise and blame, reward and punishment, befriending and shunning, loving and parenting and choosing, if we don’t have a principle to guide us, a reason to trust our guts, or a divine guide to shine a light on our sinful nature and lead us out of the darkness of doubt?
The goal of moral philosophy should perhaps become to provide a framework for our moral lives that is structured around soul-searching, storytelling, and individual encounters with normativity. When we start with our practices and let the metaethical concepts bubble up from those, we’ll encounter vestiges of religious and legal traditions, but we’ll also realize:
Our moral lives are not generally organized hierarchically, from the bottom up or the top down, with some grand principle at the base or some supreme authority at the pinnacle.
Our moral lives are assembled out of the lessons we have learned and the projects we have set for ourselves given the people we have loved and respected and the communities of which we are loyal members. (I’ve written elsewhere about how prejudices function as crystallized judgments in this sense: heuristic instruments for living in a world whose every relevant detail cannot be fully known in advance.)
Given all the evidence that principles can be ignored or perverted, what ends up mattering for moral life is whose name you put in that familiar bumper-sticker, “What would Jesus do?” It’s odd that there is so little moral philosophical attention to our paragons of virtue and goodness, since they play such a big role in our actual moral lives. We might ask this a bunch of different ways: Whose approval are we explicitly or hypothetically seeking? Whose life story are we trying to emulate? Who are the Disney villains we’re trying to avoid becoming, and what is their signature vice?
The Existentialist Inheritance
These questions point toward a tradition that took the empty throne seriously and refused to pretend it could be refilled. The existentialists understood that once you strip away the divine legislator, you don’t get a tidier secular version of the same system. You get radical freedom, and with it the vertigo of choosing who to become without cosmic authorization.
This is where the three psychological substitutes fail most clearly. Pathology, happiness, and cognitive bias are all ways of avoiding the confrontation with freedom that the death of God actually demands. They replace one set of external authorities with another: the clinician, the well-being researcher, the behavioral economist. The empty throne stays furnished.
Existentialism, for all its mid-century baggage, got something right that neither Anscombe’s Aristotelian revival nor the psychological establishment has adequately addressed: the moral weight falls on the individual’s encounter with meaning, and that encounter cannot be delegated to any science. The search for meaning, the liberatory possibilities of art and narrative, the oppressive structures of modern life under capitalism and bureaucracies, the opportunities and crushed hopes for revolutionary change: these are the conditions under which we actually form our moral lives. Psychology can describe some of these dynamics, but it cannot prescribe our response to them.
The question I started with was Anscombe’s: can we do moral philosophy without first having an adequate philosophy of psychology? I want to end with a related but different question. Once we have that philosophy of psychology, and we see clearly what it can and cannot do, what then? The existentialists thought we would need to confront freedom, absurdity, and the irreducible responsibility of choosing. The paragons and role models, the stories we tell about who we are and want to become, the communities whose approval we seek: these are not substitutes for moral philosophy. They are its proper subject matter, once we stop pretending that the throne was ever occupied.
“Now there was a day when the sons of God came to present themselves before the LORD, and Satan came also among them. And the LORD said unto Satan, Whence comest thou?
Then Satan answered the LORD, and said, From going to and fro in the earth, and from walking up and down in it” (Job 1:6)
Iain McGilchrist quoted this verse in a keynote that I just heard him deliver at a conference at Duke. McGilchrist ranged from neuroscience to theology in a long and rich talk. His premises were scientific, metaphysical, moral, and political, and I wouldn’t endorse them all. But his description of artificial intelligence as satanic is worth serious consideration on its own.
For me (although perhaps not for McGilchrist), Satan is a metaphor. But we need metaphors or models to make sense of phenomena like AI, and Satan provides a valuable alternative to some other metaphors, such as AI as a tool, a machine, a mind, a person, or a social organization.
The Satanic metaphor draws our attention to temptation, which is Satan’s favorite trick. It presents AI as not new but instead as an appearance of things that have been walking to and fro all along, such as greed and power-lust. It explains why AI might seem like a god to some (for instance, Silicon Valley tech-bros), since Satan is known to appear as a false savior. Large language models also speak to us as if they were people, talking sycophantically in the first-person singular, much as Satan does. (“Then Satan answered the LORD, and said, Doth Job fear God for nought?”) Finally, the metaphor poses the classic question of whether AI is an active force or rather a manifestation of human freedom.
I wrote a review of a new PBS documentary about urban debate leagues for Education Next. It was published today, and it begins:
You may have seen a movie in which teenagers experience grave injustice and then enter a prestigious competition where they prove to the world that they are smart. The competition might be the AP math exam (Stand and Deliver, 1988), the National Spelling Bee (Akeelah and the Bee, 2006), robotics (Spare Parts, 2015), or chess (Queen of Katwe, 2016), to name just a few.
Typically, one charismatic adult believes in the kids, inspires them to confront their doubts and society’s stereotypes, and leads them—through setbacks—to an exciting victory that demonstrates their dignity and character as well as their skills.
A few months ago, I saw a Reddit post advertising some kind of AI development editor. The author claimed to have written novels, paid a development editor to review them, and been unsuccessful. Then their software-engineer husband vibecoded an AI tool in a weekend that supposedly produced all the same criticisms and revision suggestions, but for $20/month. Lots of skeptical Redditors objected that this was likely a scam. Nobody could actually find the tool.
But the premise stuck with me: here’s a hyper-specialized skill that maybe a hundred people in the world do really well, for a major premium, and a performant LLM might approximate it. So what does a development editor actually do for $3,000?
What You’re Paying For
I talked it over with Claude and with a couple of friends who have actual writing careers. What emerged is that “development editor” is a consulting job that requires good aesthetic judgment, connections, and self-promotion. For that sum, you can expect a well-read, finely-honed expert to read your manuscript closely (and potentially reread it), pore over key sections, perform a set of recognized structural analyses, and then write a long, thoughtful editorial letter. Afterwards, there’s usually a meeting where the editor walks you through the documents, hands you a marked-up copy, and brainstorms solutions.
Many authors believe there’s an implicit further transaction: that the $3k earns you a ticket out of the slush pile through the editor’s rolodex, if they like the work. The idea is that the development editor is vetting you during the process, and if they like what they see, they’ll connect you with serious agents and publishers. If that’s true, then it’s less a time-intensive editorial task than a gatekeeping function wrapped around a skill wrapped around a real time commitment.
But the reality is that most people doing this work don’t have an expansive network and can’t get you a contract. The Association of American Literary Agents actually enforces this separation: agents who offer editing must refund all fees if they later offer representation for the same manuscript. The skill just is discernment.
And it’s real work. For an 80,000-word novel, a developmental editor typically puts in 40 to 80 hours of active analytical labor: reading, rereading, building scene maps, drafting the editorial letter. At $3,000, that’s $40-75 an hour for highly specialized cognitive work.
And someone online claimed AI could replace it, right now, in early 2026, with a weekend’s work.
Put frankly: that sounds like bunk. But I thought it might be fun as a kind of test for how much a nerdy non-programmer could do. I missed the last bandwagon when everyone was building apps, but this, maybe, I could try. So I used AI to build Apodictic.
Why “Apodictic”?
I’d been writing about Arendt and judgment in my free time, and I kept trying to justify this experiment in terms of testing whether LLMs could exercise authentic aesthetic judgment. Early on, I came up with the working title “anotherpanacea’s development editor,” whose initials, APDE, sound a little like “apodictic.” Kant uses “apodictic judgments” to name necessary copulas: judgments like “Necessarily, bachelors are unmarried men,” where the “are” can’t be otherwise. That’s the kind of deductive logic we expect from computers. But somehow LLMs are giving us more than that.
The name also rhymes with something I discovered about how fiction works. A book can be wildly experimental, playing with form, genre, voice, and plot in ways that surprise, frustrate, and infuriate. That can be highly enjoyable, but only if the reader is primed for it. Otherwise, experimental fiction generates one-star reviews and never finds an audience. So even difficult fiction has to communicate its intentions to its readers somehow.
Development editors think about this in terms of a contract between the book and the reader: what are you telling the reader to expect? Do you deliver? Genres might actually offer something closer to apodictic inferences. Not every mystery is a Whodunnit (it might be a Howcatchem), but there’s always a reveal economy. By asking authors to articulate their goals, Apodictic doesn’t have to just read and wonder: it can test a novel’s putative thesis and genre self-identification against the conventions of that genre.
What the tool does, in practice, is take a first pass to infer a contract. The most performant large language models can do a surprisingly good job of this. It’s like asking the AI: “What am I trying to say?” When it gets something wrong, you correct it, and then (if it works) it tells you which parts of the manuscript are doing something you didn’t intend.
One design principle mattered more than any other: the firewall. Every AI writing tool I tried wanted to rewrite my prose. I didn’t need a co-writer. Apodictic diagnoses problems and identifies classes of solution. It never invents content: no new plot events, characters, dialogue, or imagery. You’re the writer. It’s the analyst. Without that boundary, the tool would just become another way for the LLM to take over your manuscript, and the whole exercise would be pointless.
What I Learned About Writing
Building the back-end reference files was the most fun part. It was also an opportunity to dust off all the fiction and narrative nonfiction guides I’d always wanted to figure out: the detailed craft skills a competent writer cultivates that I, as an academic and even as an editor, never had a professional reason to learn. The idea that every avid reader would make a good writer is like the idea that every avid magic fan would make a good magician.
Reverse-outlining, for instance, is probably obvious to anyone who does serious work on fiction. You can learn something like it in law school, and every logic professor has taught a version of argument diagramming. But thinking clearly about the reader experience and reveal economy, or measuring the proportionality of different narrative elements: that’s what you have to know to write a good novel, not just to read one. A beat map is a cool tool, and if you’ve reread a book a few times you could probably reconstruct one. But it’s not like such things come naturally. (At least not to me.) And once you know the moves of Save the Cat, you won’t be able to watch network television without seeing them everywhere, which is less a gift than a curse.
I’ve always been tempted by the idea that narratives make arguments much like philosophy papers do, and that’s pretty obviously reductive. But the most transformative thing I learned was simpler: development editors think in terms of a contract between the book and the reader. What are you promising? Do you deliver? Within genre constraints, fiction and narrative nonfiction flourish when the author has something to say and something to refute. These guardrails are part of what a development editor looks for and enforces. It’s a simplification, but for many authors and readers it’s a necessary one.
That idea reframed everything else I was learning. The difference between Happily Ever After and Happy For Now in romance isn’t a trivial genre tag; it’s a promise with real consequences if broken. Grimdark and hopepunk aren’t symmetrical moods; they make different contracts with the reader. I’d never thought about any of this before, and it’s already starting to change how I read.
What I Learned About LLMs
Here’s where it gets humbling. The first time I ran a simple “be a development editor” prompt on a recently published novel (Dungeon Crawler Carl), a five-line prompt got probably 60% of the insights that my much more elaborate tool produced. I was shocked.
Even more humbling: it turned out the simple prompt could do a great job without even reading the novel. I picked what I thought was an obscure time-travel novel from the late 80s, Leo Frankowski’s The Cross-Time Engineer, and Apodictic seemed to do well analyzing it. It was even insightful about the book’s misogyny. But so was the five-line prompt, because the misogyny is famous, and there are plenty of discussions of it on Goodreads and in reviews that ended up in the training data. Anthropic paid a $1.5 billion settlement for ingesting pirated books, too, so the model may have had direct access to the text. This is a known issue in benchmarks like the math olympiad, but I didn’t expect the model to waste weights on obscure mid-list science fiction from forty years ago.
Among many other things, what we’re seeing in LLMs is a spectacular project of knowledge compression. They’re big files, but they contain far more general knowledge than you’d expect.
Thankfully, performance drops precipitously with unpublished writing, which is, of course, what a development editor actually works on. A few other lessons:
The models are sycophantic. Everyone knows this on some level, but it’s really hard to get them to notice a criticism and sit with it rather than explaining it away. A lot of what I built was about making sure hostile perspectives survive the review process.
Structure matters, but less than you’d think. You can get 20-30% of the depth of analysis from Claude with a simple five-line prompt in incognito mode. The elaborate plugin structures the next 70%. The AI labs mostly think this kind of scaffolding is unnecessary: as datasets grow, they naturally incorporate writing expertise from the entire English-language corpus. I kept testing the simple version against mine and almost gave up when the simple version started getting really good.
Multi-model synthesis helps. You can get better results by asking the same question of Gemini, ChatGPT, and Claude, then asking one of them to synthesize all three answers. Gemini writes a little research paper on every question; ChatGPT applies clear structural thinking; Claude writes lyrically and thoughtfully, and has enough working memory to evaluate everything together.
Vibecoding Is Real
I am not a computer programmer. I can parse basic HTML and have some rudimentary database knowledge. I learned git for this project. But I was able to build an app version of the plugin and fully set it up. It’s not totally stable (the main instability is that it uses a janky backend cloud computer rather than more advanced hardware) but it works.
One thing about vibecoding: you can do it early in the morning and late at night. Mostly it’s testing, reviewing, asking for a specific fix, and then waiting around while the machine does the work. I built the basic thing on Google AI Studio in React in an evening.
A confession on models: Codex is faster and smarter at React programming than Claude, and it’s not close. Especially after the Pentagon fiasco, Anthropic has my loyalty, and Claude works best with my personal approach to text and writing. But Codex 5.3 is just more of a stickler for coding projects right now, at least for what I’ve been building.
What I Actually Learned
I started this project to test a dubious claim from the internet. What I didn’t expect was that building the tool would teach me more about writing than twenty years of avid reading had. AI makes crazy projects possible: a philosophy PhD with no programming background can prototype a working app. And the work that editors, writers, and critics do is going to change. The models aren’t better than human judgment, but they’re good enough to sharpen it.
Try It Yourself
If you want to skip the tool entirely and try the bare-bones version, here’s the five-line prompt that gets you surprisingly far:
You are a developmental editor for fiction. Read the attached manuscript and write an editorial letter in Markdown. Identify what’s working structurally, what’s losing momentum or undermining its own impact, and provide a prioritized revision checklist. Include an adversarial stress test: inhabit hostile reader perspectives and identify what an uncharitable reader would attack. For each adversarial claim, commit to a severity rating before generating a counter-argument, and do not let the counter-argument reduce the severity. End with a “what not to touch” section.
And a three-line addendum that fixes a common failure mode where the model doesn’t actually read the whole manuscript and hallucinates the rest:
Read the complete manuscript before beginning analysis. If the text is long enough that your reading tool truncates or summarizes middle sections, read it in sequential chunks until you have covered every section. Do not estimate word count—count it or ask for it. Do not begin drafting the editorial letter until you can confirm you have read the final page.
Give it a try on something you’ve written. Start with the five-line prompt and see what it catches. If you want to see what the full tool produces, here are some sample outputs on published books you can check against your own reading: an editorial letter for Dungeon Crawler Carl (structural proportion problems, emotional ceiling, non-terminal climax), an editorial letter for Theo of Golden (episodic structure, protagonist without an arc), a targeted audit of A Court of Thorns and Roses (force architecture and erotic content), and a pre-writing pathway for a historical novel interleaving Arendt and Rahel Varnhagen. (I’d like to write this someday.)
If you want the structured version, Apodictic is here. It also runs as a Claude Code plugin and as a Custom GPT, both at no additional cost if you already have a Claude or ChatGPT subscription. And then tell me what it got wrong. The whole point of the adversarial layer is that these models want to be nice to you, and the tool is only as good as its ability to resist that impulse. If Apodictic pulled its punches on your manuscript, or praised something it should have flagged, I want to know. Drop a comment below or email me at anotherpanacea@gmail.com.
Last night, I was part of The United States at 250: A Tufts Faculty Panel. In a full room of students, Tufts historians and political scientists with various specialities addressed the question: “Where are we as a nation and what’s next?”
I offered the following argument. I have derived it from other people’s scholarship, and I am not sure it is true, but I think Americans should consider it.
We’re marking a 250th anniversary because 1776 began the period that concluded with our Constitution. However, the Constitution is now in a deep crisis. We may now be coming to the end of a 250-year period. The reasons are not named “Donald J. Trump.” These are three deeper reasons.
First, presidential republics have a fatal flaw, and none except the US–and arguably, France–has survived for a long period (Linz 1990). Whenever opposing parties control the legislature and executive, they are motivated to battle at the cost of the republic.
For most of our first two centuries, we did not have regular impasses, because the Democrats were divided into two major blocs, resulting in at least three effective parties in Congress; and most presidents could build a working majority. However, when conservative Democrats defected to the GOP, the two parties polarized. Since 1990, it has been possible to govern in the ways envisioned by the Constitution only when the same party has controlled both elected branches (6 periods of 14 total years). During the other 24 years since 1990, presidents have tried to rule by executive order and Congress has tried to undermine the current administration. We have moved ever closer to complete constitutional breakdown.
Second, the Constitution enacts three branches of government: the executive, legislative, and judiciary. Since at least 1932, we have actually had another branch: the administrative and regulatory agencies, staffed by about about 2.2 million federal employees who are understood to be insulated from politics. They follow rules, norms, and principles of their own that are not mentioned in the Constitution–for example, scientifically measuring the costs and benefits of proposed policies and publishing drafts of policies for public review and comment. Perhaps we have also had a fifth branch, the national security apparatus.
We muddled through for decades by pretending that the agencies were part of the executive branch while the White House usually deferred to them. Under a 1984 Supreme Court decision, Chevron, the courts also generally deferred to agencies’ decisions. Meanwhile, Congress intentionally gave agencies broad scope. The regulatory state was largely independent from the other branches.
However, in 2024, the Court repealed Chevron with the Loper decision, allowing courts to review agency decisions. And Donald Trump has fired and replaced many civil servants and members of so-called independent agencies for openly political reasons.
Libertarians argue that we shouldn’t have had a massive federal government in the first place. And populists of right and left argue that an elected president should be able to determine policies. A left populist may celebrate the opportunity for a Democratic president to reshape the agencies at will now that they have lost their independence. I think, however, that every country with an advanced economy has built an elaborate and quasi-independent regulatory apparatus that applies science and managerial acumen to generate benefits that voters want. We may not have that anymore.
Third, Congress no longer legislates, in the sense of passing or reforming substantive statutes. In 1965 alone, Congress passed at least 10 landmark bills that established agencies or dramatically altered national policies. As recently at the 1980s, Congress sometimes legislated by substantially cutting regulation. But Congress has arguably passed no major laws in this whole century so far.
For example, Congress has never passed legislation explicitly about the climate. Federal regulatory agencies have used 1970s Clean Air Act (written before Congress was really aware of climate change) to try to regulate carbon. Likewise, federal financial laws were passed before cryptocurrency; and the Telecommunications Act of 1996 still governs despite some minor new developments, such as social media and smartphones.
In sum, we can’t handle frequent periods of divided government; our massive regulatory state lacks a constitutional basis; and the branch in which “all legislative power” is “vested” no longer legislates.
It is possible that we will keep driving ahead, frequently bumping into the Constitution’s guardrails but somehow staying on the road for decades.
Or we could see substantial reforms–major constitutional amendments or new voting laws that change the basic structure. (For instance, proportional representation would transform Congress–for better or worse–and could be accomplished by law.) I sometimes wonder whether our incompetent and blatantly authoritarian president is a blessing, alerting people to the need for reform without successfully consolidating power.
Or we could see a collapse. The typical final act of a presidential republic is a soft dictatorship. That’s why this topic is important to discuss on our 250th.
On February 18, the Alliance for Civics in the Academy hosted a webinar on “What Counts as Success? Assessing the Impact of Civics in Higher Ed” with Trygve Throntveit, Rachel Wahl, Joseph Kahne, and me.
We discussed some of the advantages of developing reliable and consistent measurements of civic education, particularly the opportunity to learn from data and the need to be accountable. We also discussed some drawbacks and risks, including Campbell’s Law (a remark by Donald T. Campbell): “The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor.”
We asked ourselves who should use assessments, and for what purposes. For example, it is a different matter for a college professor to get feedback from the students in a course or for a university to measure student outcomes. I thought the conversation was both intellectually serious and relevant to practice.
Panelists:
Rachel Wahl: Associate Professor in the Social Foundations Program, Department of Educational Leadership, Foundations, and Policy at the School of Education and Human Development at the University of Virginia
Joseph Kahne: Ted and Jo Dutton Presidential Professor for Education Policy and Politics and Director of the Civic Engagement Research Group at the University of California, Riverside.
Trygve Throntveit: PhD, Research Professor in Higher Education and Associate Director of the Center for Economic and Civic Learning (CECL) at Ball State University.
