Nuclear cascades or more of the same? Why meliorists may have gotten it right: A commentary on Tetlock et al. (2023)

Tetlock et al.'s balanced contribution to the debate over the possibility of long-range geopolitical forecasting provides a useful roadmap at a critical time. An especially challenging geopolitical juncture compels heightened attention to systematic efforts of this sort that both identify limits on expert judgment and offer ways to overcome them. The task may be extremely difficult—skeptics abound—but is nevertheless vital for a social science true to the mission of enhancing peace and security. Tetlock et al. report findings from a previous study suggesting that expertise in nuclear weapons improved accuracy in predicting long-range proliferation. In particular, they argue, experts did not cry wolf; they exhibited low False-Alarm rates. Meliorists could find some reassurance there. And yet this field of inquiry has also seen significant and repeated overestimation of the odds of nuclear proliferation cascades. In 1963 President Kennedy foresaw the potential of between 15 and 25 nuclear weapons states by 1975. Yet rather than actual proliferation cascades, it is only predictions of imminent cascades, chains, and dominoes that have proliferated since (Potter & Mukhatzhanova, 2008). Such predictions have failed to materialize thus far for well over 60 years, a significantly long range. Radical Skeptics might find this vindicating. Understanding the sources of over-predicted proliferation on the one hand, and of the more accurate tally in Tetlock et al.'s findings on the other, may shed light on these two distinctive (past) readings of the future. It may also further improve the Meliorist's case for long-range predictions.

Over-predictions of nuclear proliferation may run the epistemological-ontological gamut, but I focus here on one systematic source of bias leading to massive False Positives for over 60 years. This record is especially, though not uniquely, the domain of a brand of neorealist theory alluring for its simplicity—“it's all about systemic anarchy and balance of power.” Anarchy presumably renders all states insecure, compelling self-help while acquiring nuclear weapons provides security, helps balance power, delivers stability, and minimizes the chances of war (the classic and most impressive locus is Waltz, 1979, 1981). Yet this analytical foundation has proved fatally flawed in its predictive tally. The massive number of predicted False Positives and anomalies in neorealist studies include Ukraine, Poland, Germany, Japan, and many, many more cases.1 If one abides by the theory's core tenets, anarchy, uncertainty, and self-help should have led most if not all states to acquire nuclear weapons. Yet an overwhelming majority (191 states!) have renounced them while nine have acquired them. Even more modest predictions (Waltz, 1981) of 18 to 30 nuclear weapons states have not materialized; and even acutely vulnerable states (e.g., Vietnam, Egypt, Taiwan, and many others) have not matched those theoretical priors.

Other empirical anomalies for the theory stand out. Offers of “hegemonic protection” had disparate effects, with some states renouncing nuclear weapons but not others. In line with the theory's core tenet, alliances cannot explain nuclear abstention in an anarchic context: self-help dominates, making alliances feeble by definition. Indeed, “nuclear weapons make alliances obsolete” (Waltz, 1979, 1981). The empirical record confirms that some allies abstained from acquiring them but not others, and that most states restrained their nuclear ambitions in the absence of superpower guarantees. Comparable structural conditions can (and have) lead to very different nuclear choices. Whereas the theory predicts that multipolarity leads to nuclearization, multipolarity has led to different outcomes across time and space (e.g., East Asia vs. the Middle East vs. Latin America). Predictive insolvency has thus manifested itself at all levels of aggregation—systemic, regional, bilateral, and individual cases—and for a large number of cumulative years that offered the opportunity to test the argument's validity. This record is especially odd for a theory operating at the heart—the inner sanctum—of states' security dilemmas, a predictive domain that loads the dice in its favor as the most auspicious terrain for corroborating its tenets. If the theory cannot make it here, some could wonder whether it can make it anywhere, borrowing from Levy's (2022, p. 442) “reverse Sinatra inference.”

Speaking of such inferences, a neorealist argument consistent with its tenets must easily place Japan among “most likely” cases for acquiring nuclear weapons. Japan has endured repeated and long-lasting cumulative shocks: not one, not two, but three of its rabid neighbors have acquired nuclear weapons (Soviet Union/Russia, China, and North Korea). Yet, in an astonishing blow to neorealist predictions (i.e., countries putatively go nuclear when adversarial neighbors do so) the perfect candidate for reactive nuclearization failed to acquire its own deterrent for 70 years (and counting). This is a very long-term predictive horizon indeed, no matter what happens next (Deacon & Solingen, 2023). Too many presumed thresholds have already been crossed and even a broken clock is right twice a day. As argued, invoking the US alliance to salvage predictive failure constitutes a major logical departure from first neorealist principles, where alliances can never be trusted and are clearly theorized to lack heft anyway.

Why so many empirical failures of prediction, quantitative and qualitative? First, Tetlock et al. find that the more bias-resistant and logically coherent the forecasting agents, the more accurate their predictions. Alas, neorealism suffers variously from explanatory inconsistency, logical incoherence, or unfalsifiable predictions. Many outcomes can be made to fit vague ad-hoc notions of security maximization (a posteriori). At best the approach is under-determining and its corollaries are inconclusive and open-ended: Japan should acquire nuclear weapons; not acquire them; substitute for them conventionally; not substitute; rely on the alliance; not rely; and so on, all in the name of the same first principles of security maximization under anarchy. There isn't a singular structural logic at play: acquiring nuclear weapons can effectively lead to higher insecurity and vulnerability. Second, neorealist arguments are time-invariant, stemming from a putatively inexorable, ahistorical “balance of power” untouched by evolving world-time. Third, as an invisible hand, “balance of power” is surely among the most invisible of them all, under-defined, unmeasurable, and unquantifiable; a wobbly, vague concept with low intersubjective reliability. Fourth, Tetlock et al. also expect that “accuracy will fall off more slowly for experts who demonstrate a command of predictively useful knowledge in the early periods.” This intuition is borne out by the fact that neorealist accounts typically hinge less on deep knowledge of any cases than on simplified “Occam razor”-style assumptions. This theoretical penchant for extreme parsimony leads to information asymmetry between neorealist-based predictions and those relying on richer knowledge of cases. Coming to term with these shortcomings, in recent decades studies began challenging previously unquestioned confidence in this theory as the driving force of all nuclear decisions (Ogilvie-White, 1996; Solingen, 1994a, 1994b).

Having identified a key intellectual culprit of over-predictions, I turn to alternative approaches that may account for the superior predictive tally reported in Tetlock et al.'s study. Meliorists can rejoice in their initial findings that experts leaning on tolerance of ambiguity and dissonance, scope-sensitivity, bounded claims, and self-criticism may generate predictions that have proven better endowed to withstand the test of time. To the extent that these approaches avoided False Alarms of proliferation cascades, they already bring an epistemic edge to the table. The list of hypotheses under this rubric is diverse and more extensively discussed in Solingen (1994a, 2018a). In this brief essay, I confine myself to an argument I developed in International Security (1994a), later elaborated in Nuclear Logics (Solingen, 1994a, 1994b, 2007). An influential meta-study on failed proliferation chains by Potter and Mukhatzhanova (2008) wielded it as an example that obviated over-predicted cascades. Another study (Sil & Katzenstein, 2010) considered Nuclear Logics' core argument to have been the most undertheorized in this field hitherto as well as an analytically eclectic—yet clearly formulated—framework for estimating nuclear scenarios.

At its core, this framework highlighted that conceptions of security threats were not simple derivatives of abstract, elusive balances of power but were, instead, highly influenced by domestic models of political survival. These models hinged on competing orientations to the global political economy and offered important clues for estimating nuclear choices since the inception of the Nonproliferation Treaty in 1968. Internationalizing models were predicted to lower the likelihood of nuclear weapons' acquisition based on identified causal mechanisms upholding that relationship. Indeed, for over 50 years every decision to renounce nuclear weapons by states that had previously entertained them was nested in broader shifts toward internationalizing models. By contrast, most defiant nuclear courses since 1968 unfolded under inward-looking models hostile to integration in the global political economy. This pattern found support across different regional security contexts and diverse associations with hegemons and alliances. That it does so in a theoretically unfriendly terrain is especially suggestive; nuclear choices are arguably a “least likely” arena for corroborating the explanatory and predictive heft of domestic political economy models. And yet, despite the regularities the approach uncovered, it had been omitted from the canonical causal repertoire. That omission led to an overestimation of other presumed causal variables, especially neorealist ones.

Models of political survival explain both synchronic variation in nuclear preferences across states and overtime variation within the same state; varying state compliance with nonproliferation commitments; varying readings of security dilemmas as more (or less) intractable; variance in ranking alliances higher than self-reliance or vice versa; variance in relative receptivity to external sanctions and inducements; why and when external coercion and inducements may be effective; and why nuclear ambitions were renounced where one might have expected them. And yet, as Sil and Katzenstein note, domestic models were not portrayed as a holy grail that displaces other perspectives; the title “Nuclear Logics”—in the plural—made clear that, while making a compelling case, this argument is not the only game in town. Sil and Katzenstein deem it a pragmatic, self-conscious effort to avoid the trap of becoming so deeply embedded in a single paradigm that makes it impossible to “look at the world with fresh eyes.” Instead, the proposed models provide a useful prism that helps weigh and re-order the relative relevance of other drivers, including structural power, interests, norms, and institutions. The models refract, interpret, and filter external stimuli, offering a nimble analytical category to accommodate the reality of multiple causation. The very definition of security is endogenous to these models rather than an independent variable, as neorealism would have it.

Furthermore, the framework identified important caveats and scope conditions, making the association between models and nuclear choices neither deterministic nor inevitable but probabilistic and eminently falsifiable. One temporal scope condition was that the post-NPT “world-time” or “second nuclear age” posed different systemic constraints than those preceding it and possibly superseding it. Temporal sequencing also matters: nuclear reversals might be much harder when nuclear weapons acquisition precedes the inception of internationalizing models (e.g., China, India, Israel). Abandoning nuclear weapons once they already exist is more costly politically than ending a program before fruition, as prospect theory would submit. Another scope condition stipulated that the relative incidence of internationalizing versus inward-looking models at the regional level would inhibit or exacerbate domestic preferences regarding nuclear weapons in either direction. The framework also specified that the choice for or against alliances (which might/might not substitute for an indigenous deterrent) was also endogenous to domestic models. Sensitive to the fact that hegemony has its limits, this take is in notable agreement with Waltz's (1981) who rightfully argued that “in the past half-century, no country has been able to prevent other countries from going nuclear if they were determined to do so.” But this assertion evokes, once again, a conundrum for neorealist logic: Why were most countries not determined to do so in the first place in a world of anarchy, insecurity, and zero-sumness?

Political survival models are a dynamic category, entailing no linear or irreversible trajectory in either direction. Its predictive scenarios specified the logical conditions under which scenarios might obtain, the kind of evidence that would question or validate expectations, and the outcomes that would corroborate or falsify the theory. Even where past predictive performance over a relatively extended temporal range may have been reassuring, much to our dismay social theories don't typically work for eternity and may be less useful henceforth. Tetlock et al. acknowledge that easy questions posed during a placid period of history (slow-motion variables with low base rates of change) may help the case for Meliorism and long-range predictability. This suggests that the world-time under which a predictive study is conducted (e.g., the heart of the Cold War vs. the zenith of globalization or “end of history”) may affect aggregate forecasts by experts and nonexperts alike. Net predictability hinges not only on the skill or bias of forecasters but also on historical world-time. Alas, the contemporary geopolitical landscape may alter the relative stability of the last 30–40 years. Global and domestic shocks have rendered too many political, economic, technological, and other variables highly uncertain. Russia's attack on Ukraine compounded those shocks inducing nearly chaotic perturbation of significant relevance to nuclear futures. Ukraine had renounced the nuclear weapons it possessed in exchange of security guarantees from major powers including Russia (another anomaly for neorealism?). Putin proceeded to invade Ukraine in 2014 and 2022 and shower nuclear threats on several targets in and beyond Ukraine. Finally, a putative “third nuclear age” driven by artificial intelligence ambitions might return the “cascade” scenario forecasted by President Kennedy and so many others since the 1960s. But in line with the cautionary advice in Tetlock et al., one must also consider the distinct possibility that the emerging world time might retain some of the basic underlying conditions that obviated cascades in the past. Path dependence can persist. It may be premature to proclaim a wholesale Zeitenwende in nuclear forecasting. Even so, discontinuities and inefficient history (March & Olsen, 1998) could burden longer-range predictions under the current geopolitical juncture, pregnant with the highest levels of uncertainty in decades.

An enduring allegiance to failed epistemologies and ontologies is a sure path to fuel Radical Skeptics' convictions. Similar errors stemming from reductionist conceptualization have led to repeated falsified predictions of NATO's demise for over 70 years and much more. To be sure, neorealism is not the only source of prediction error. A related approach associated with technological determinism also exhibits a substantial number of “most likely cases” that renounced nuclear weapons and “least likely cases” that pursued them (Miller, 2017). As Lewis (2016) argued, whether or not states decide not to acquire nuclear weapons has often far more to do with restraint than with technological barriers, given a 65-year-old technology. As the sheer size of failed predictions became evident, new studies pledging neorealist roots sought to rectify some of the noted deficiencies. In so doing, that work departs from first neorealist principles in several ways: it suddenly endows alliances with colossal analytical heft and attaches a large number of contingencies that throw “Occam razor” out the window (Debs & Monteiro, 2016). That is a commendable step that could enhance the potential for “adversarial collaborations” and improved predictions. Retaining other first (neorealist) principles, however, may also retain logical inconsistency. A consistent focus on the logic of anarchy trumps the idea that states can substitute self-help for external protection. Furthermore, continued reliance on fuzzy, elastic, and highly subjective definitions of balance of power, self-help, insecurity, and power itself make predictions even more elusive and possibly unfalsifiable. Scenarios hinging on external threats—not a figment of neorealist imagination—could do better when considering the domestic political landscapes that heighten or dampen pressures to respond to threats with nuclear weapons' acquisition. Important work along the lines of classical and neoclassical realism (Wohlforth, 1993) does just that.

Adversarial collaborations proposed by Tetlock et al. would have to overcome deep cleavages in this field of inquiry, including methodological ones, that may have hindered progress hitherto. Such collaborations could, however, facilitate the crafting of “coherence hypotheses” that bear on the connections between the empirical accuracy of forecasts and the logical rigor with which forecasters make cumulative-risk judgments. “Viewpoint-variable” hypotheses can consider systematic bias stemming from the connections among ideology, forecasts, and forecasting accuracy. Other rules of thumb for converging on adversarial collaborations include greater attention to complex systemic effects, replacing analytically impoverished and grossly deficient—if parsimonious—forecasts. Hypotheses must be cast in falsifiable terms, with greater precision and better specification of threshold and boundary conditions; avoid circularity and ex-post-facto rationalizations; stipulate the kind of evidence that would challenge or corroborate their expectations; assess findings against competing theoretical claims; consider ways to discover, dissect, and assess indirect causal pathways, even if they are more difficult to work with; and devote attention to temporality and world time, or historical context. Above all, heightened tolerance for fallibility—especially for longer haul predictions—will remain forecasting's best friend, all the more so under current conditions of intensive and extensive domestic and systemic geopolitical uncertainty.