In 2010, David Markowitz (then at Stanford) published a paper arguing that fraudulent scientists write more 'obfuscated' papers than honest ones—longer words, more complex sentences, more passive voice. The logic was simple: if you're fabricating data, you instinctively write in ways that are harder to scrutinize. That paper, cited hundreds of times, became foundational work on using language psychology to detect academic fraud.
Fifteen years later, Markowitz conducted a replication and extension. The conclusion: the original finding holds, but the mechanism is more subtle than we thought.
800 Fraudulent Papers vs. 800 Genuine Papers
The new study is far larger than the original. The team collected 800 retracted papers from 2013–2025 (the fraudulent group) and 800 matched control papers (the genuine group). Matching criteria included publication year, journal, and discipline, ensuring comparability of external conditions.
Computational linguistics tools extracted language features from each paper—not just the original 'obfuscation' metrics but a full analysis across more than 100 linguistic dimensions.
Core finding: fraudulent papers are indeed more 'obfuscated' than genuine ones. The original hypothesis holds.
But the more interesting question is *why*.
Three Competing Mechanisms
The original study proposed the cognitive cost hypothesis: fraudsters deliberately write obscurely, because obscure text demands more cognitive resources from readers, lowering the probability of detection.
The problem with this mechanism: if fraudsters were truly writing obscurely *on purpose*, their writing strategies should differ fundamentally from ordinary authors. Yet many retracted papers look... fairly normal at the linguistic level.
The new study tested three competing mechanisms:
1. Cognitive cost hypothesis (original): deliberate obfuscation → higher reader analysis cost → lower detection probability 2. Reality Monitoring theory: fraudsters lack real experimental memories to draw on while writing, so their language automatically skews toward the 'abstract' and 'non-specific' 3. Emotional load hypothesis: the heightened emotional burden of lying affects language production patterns
The parallel mediation model's result: Reality Monitoring theory has the strongest explanatory power.
Specifically, the largest linguistic difference between fraudulent and genuine papers is not 'longer sentences' or 'more complex words'—it is fewer adjectives.
Adjectives: The Honest Writer's Fingerprint
Reality Monitoring is a classic psychological framework: memories of real experiences contain more perceptual details—color, shape, sound, touch—than imagined events. Those perceptual details project into language as adjectives.
If you ran a real experiment, you'd write "the blue solution turned cloudy after 30 seconds"—color, state, time.
If you fabricated one, you'd write "the reaction proceeded as expected"—no perceptual detail, only abstraction.
This isn't fraudsters *deliberately* writing obscurely. It's that without real experimental memories to draw on, the brain automatically fills in with more abstract language. It's like recalling a place you've never visited: you'd say "the scenery was nice," but if you'd actually been there, you'd say "the wind at the summit was so strong you couldn't keep your hat on."
This finding shifts the mechanism of 'linguistic fraud detection' from *strategic deception* to *cognitive structural differences*. Fraudsters aren't deliberately hiding anything—their cognitive state simply differs from honest authors', and that difference naturally shows up in language.
Detection Accuracy: 60.7% and 71.6%
Beyond mechanism analysis, the study tested the practical value of linguistic fraud detection.
Using only obfuscation features, cross-validated detection accuracy was 60.7%—better than chance (50%) but far from practical.
An elastic net model using 100+ linguistic dimensions raised accuracy to 71.6%. That number doesn't look impressive, but consider that the detection targets are papers that already passed peer review—these have already been scrutinized by experts at the content level. A weak linguistic signal reaching 71.6% is notable.
The researchers are also honest about this: 71.6% accuracy means a 28.4% error rate. Applied to real manuscript screening, roughly 28 of every 100 honest papers would be flagged as 'suspicious.' That false-positive rate is unacceptable in practice.
Replication in the Open Science Era
The paper's other value is that it is itself a replication study.
Psychology's 2010–2020 'replication crisis' taught the field that many classic findings fail to reproduce under larger samples and stricter methods. Markowitz's 2010 study used a smaller sample—would it hold fifteen years later?
Answer: yes. But the effect size may be smaller than originally believed, and the mechanism more complex.
This is how science works—not a one-time 'discovery of truth,' but continuous re-examination at larger scales and across more dimensions. Markowitz replicating his own work is itself a model of scholarly integrity.
Implications for AI Detection
The study has indirect but important implications for AI-generated text detection.
Many current AI detectors follow a similar logic: extract linguistic features, use a classifier to judge 'human vs. AI.' These detectors face the same problem as fraud detection—false-positive rates are too high.
Markowitz's findings suggest an intriguing analogy: real experimental memory leaves an 'adjective fingerprint,' while fabricated content lacks it. Similarly, genuine human writing may contain 'cognitive structural traces' that AI struggles to imitate. But finding those traces requires understanding *why* human and AI language differ at the cognitive level—not just surface feature matching.
There's also a lesson about mechanism vs. correlation. If fraudulent papers have fewer adjectives because of 'missing real memory,' then any writing lacking genuine experience—AI-generated or human-fabricated—might show similar linguistic patterns. But if the adjective difference stems from other causes (disciplinary norms, writing style), using it for detection would produce many misjudgments.
Limitations and Reflections
Several limitations are worth noting:
- Sample bias: Retracted papers are only a subset of *detected* fraud. Much undetected fraud is absent from the sample—so the study measures the language of *discovered* fraud, not fraud in general.
- Time span: Papers from 2013–2025 straddle the rise of AI writing tools. Post-2023 papers may confound 'fraud' and 'AI generation' signals; the researchers don't discuss this in detail.
- Disciplinary coverage: Writing norms differ enormously across fields. Although the study matched by discipline, the 'fraud signal' in 100 linguistic dimensions may vary by field.
- Causal direction: Though the parallel mediation model supports Reality Monitoring, causality remains uncertain. Does 'lacking real memory reduce adjectives,' or 'are people skilled at abstract writing more likely to fabricate'? Finer experimental designs are needed.
Conclusion
The biggest takeaway: the traces of deception lie not in strategy but in cognitive structure.
Fraudsters aren't 'deliberately' writing obscurely—they may not realize their language differs from honest authors' at all. The difference is cognitive, unconscious, and hard to fake.
Just as pupils dilate and voice pitch shifts when someone lies—signals beyond conscious control—adjective density in language may be the cognitive equivalent of pupil dilation.
71.6% detection accuracy isn't practical yet. But the mechanism it reveals—real experience leaves perceptual fingerprints in language—points to a more fundamental direction: to detect deception, don't look at *what* was written; look at *why* it was written that way.
---
Paper: https://arxiv.org/abs/2610.13678
Author: David M. Markowitz (University of Oregon)
Original study: Markowitz, D. M. (2010). *Linguistic obfuscation in fraudulent science.*