The Caricature Gap: Why Synthetic Political Text Fails at the Population Level, Not the Sentence Level

Most of the energy in AI-generated text detection has gone into the wrong place. We have built classifiers that hunt for perplexity anomalies, burstiness signatures, and token-level irregularities, treating the problem as one of authorship: did a machine write this sentence? That framing made sense when language models produced stilted, repetitive prose. It makes considerably less sense now that fluency is essentially solved. A new paper, "The Algorithmic Caricature: Auditing LLM-Generated Political Discourse Across Crisis Events" by Gunjan, Sidahmed Benabderrahmane, and Talal Rahwan at NYU Abu Dhabi, argues that the durable signal is not in the sentence at all. It is in the crowd.

The authors reframe synthetic discourse detection as a Computational Social Science problem. Instead of asking whether one post looks machine-generated, they ask whether a population of generated posts behaves like a real online public. The answer, across nine politically salient crisis events, is consistently no. Generated political discourse is a caricature: recognizable in outline, but flattened, exaggerated, and stripped of the heterogeneity that characterizes actual human crowds.

A population-level audit, not a sentence-level detector

The empirical backbone is a paired multi-event corpus of 1,789,406 posts spanning nine events: the COVID-19 pandemic, the January 6 Capitol attack, the 2020 and 2024 US elections, Dobbs/Roe v. Wade, the 2020 BLM protests, the US midterms, the Utah shooting, and the US-Iran conflict. For each event the authors collect observed discourse from social platforms and then generate synthetic discourse about the same event, allowing a controlled comparison that holds topic fixed while varying the source.

Crucially, the unit of analysis is the distribution, not the document. The framework compares observed and synthetic discourse along four dimensions:

• Emotional intensity: sentiment polarity and its dispersion across the population.
• Structural regularity: post length distributions, repetition, and the shape of their tails.
• Lexical-ideological framing: how abstract or concrete, formal or colloquial, the vocabulary is.
• Cross-event dependency: whether the discourse adapts to the specific event or recycles generic patterns across events.

The authors use both mean-based gaps (does synthetic text differ on average?) and dispersion-based evidence (does it differ in its variance and tail behavior?). This second move matters more than it might appear. Two populations can share an identical mean sentiment while being completely different distributions: one tightly clustered, one spread across the full emotional range. A mean comparison alone would miss this. By instrumenting dispersion explicitly, the paper captures the property that distinguishes a real crowd from a synthetic one.

What the caricature actually looks like

The findings are consistent and, on reflection, intuitive. Synthetic political discourse is:

• More negative and less dispersed in sentiment. The model defaults to a narrow band of emotional register. Real publics, by contrast, contain sarcasm, ambivalence, dark humor, indifference, and the occasional off-topic joke. The generated population collapses this range.
• Structurally more regular. Human discourse has long tails: one-word reactions, link dumps, multi-paragraph rants, and everything between. Synthetic discourse clusters around a typical length and rarely produces the extremes.
• Lexically more abstract. Generated text reaches for general, dramatized framing. Observed text is full of context-specific slang, local references, hashtags, in-jokes, and the colloquial markers that anchor a post to a particular moment and community.

The authors summarize these differences with a single event-level divergence measure they call the Caricature Gap. The name is well chosen. A caricature is not a failed portrait; it is an accurate exaggeration of salient features with everything else suppressed. That is precisely what these models produce when asked to simulate a political crowd. They amplify the dominant emotional and rhetorical signature of an event while discarding the variance.

The most interesting result is that the gap is not uniform. It widens for fast-moving, decentralized crises and narrows for formal, institutionally mediated events. Consider the contrast: a presidential election generates a great deal of institutional language, official statements, news framing, and predictable partisan talking points. A model trained on the broad sweep of the internet reproduces that register reasonably well, because the register is itself somewhat templated. A decentralized event like a sudden shooting or a viral protest, on the other hand, produces discourse that is improvisational, local, and saturated with community-specific texture. That is exactly the material a model has the least ability to reconstruct, because the texture lives in the long tail it systematically smooths over.

My take: the gap is a moving target, not a permanent moat

This is the result I find genuinely useful, and also the one I would push hardest on. The population-level framing is correct, and it is a better foundation than perplexity-based detection precisely because it does not depend on a single model's quirks. Dispersion, tail behavior, and cross-event adaptivity are statistical properties of crowds, and they are harder to fake than the surface smoothness of a sentence.

But I do not think the Caricature Gap is a stable defensive line, for three reasons.

First, the gap is partly an artifact of default generation behavior, not a fundamental limitation. The reduced dispersion the authors observe reflects how models behave under naive prompting: they regress to the modal register. An adversary who wants to produce realistic synthetic crowds will not prompt naively. Mixture-of-persona generation, deliberate temperature variation, injected demographic and stylistic conditioning, and post-hoc resampling to match a target distribution can all widen synthetic dispersion on purpose. The honest reading is that the paper measures the gap for unsophisticated synthetic discourse. The sophisticated case is the one we actually need to worry about.

Second, there is a data feedback loop. As models train on increasing volumes of exactly this messy, colloquial, long-tailed social media data, the lexical and structural components of the gap should narrow. The model's prior over what a crisis crowd sounds like will improve. This is the open question I raised in the short version of this argument: as training data absorbs the mess, does the Caricature Gap close, or does it relocate? My inference is that it relocates. The easily measured dimensions (length distribution, lexical concreteness) will converge first, because they are the most learnable. The harder dimensions (genuine cross-event dependency, the causal coupling between an event's specific unfolding and the crowd's reaction) will persist longer, because they require not just stylistic mimicry but a model of how real publics actually process novel information in time.

Third, and most importantly, the cross-event dependency dimension hints at the real frontier. Authentic discourse is reactive and path-dependent. The conversation at hour three of a crisis is shaped by what happened at hours one and two, by which rumors spread, by which accounts went viral. Synthetic discourse generated from a static event description has no such trajectory. This is the dimension least likely to be solved by more data and most likely to require genuine simulation of information diffusion. If detection migrates here, it becomes a question about temporal and network structure, not text.

Limitations and what they imply

The paper is descriptive by design, and a few caveats follow. The synthetic discourse depends on the generation procedure used, so the measured gaps are conditional on those choices rather than properties of LLMs in general. Sentiment and toxicity scoring rely on classifiers that carry their own biases, particularly on informal political language. And the nine-event corpus, while large in post count, is US-centric and English-language, which limits claims about discourse environments in other political and linguistic contexts.

None of this undermines the central contribution. The shift from "does this text look machine-generated" to "does this crowd behave like a real one" is the right shift, and the Caricature Gap is a clean way to operationalize it.

Conclusion

The most valuable claim in this work is also the most sobering. Distribution-level auditing gives us a real, present-day edge over synthetic political discourse, because today's generated crowds are flatter, more negative, and more homogeneous than real ones. But that edge is temporal. It buys time, not permanence. As generation grows more sophisticated and training data absorbs the very messiness that currently betrays it, the gap will narrow on its easy dimensions and concentrate on its hard ones: temporal reactivity, network structure, and genuine event-specific dependency. The open question for the field is whether we can measure those harder dimensions before adversaries learn to fake the easier ones. The race is not between detector and generator at the sentence level. It is between population realism and population simulation, and that race has only just started.