Fuzzy logic detector spots AI-generated text with 100% accuracy

As large language models churn out essays, reviews, tweets and financial commentary at industrial scale, researchers have been racing to build detectors that can tell machine-written prose from human writing. Most of the winning approaches so far rely on deep neural networks—powerful, but notoriously opaque. Now a team at Graphic Era Hill University in Dehradun, India, has taken a deliberately different path: they let a fuzzy logic system, the same rule-based reasoning framework that powers everything from washing machines to subway brakes, do the detective work. Writing in the International Journal of Data Science and Analytics, Pradeep Kumar Kaushik, Raj Kishor Bisht and Ashok Kumar Sahoo describe a detector that scored a perfect 100 percent accuracy on two of five domain-diverse datasets, while remaining fully explainable at every step.

The core problem the researchers confronted is linguistic uncertainty. Human writing is messy: a single sentence can be simultaneously simple in vocabulary yet complex in structure, or dense with content words yet easy to read. Classical binary logic struggles with such ambiguity, forcing every sentence into a rigid yes-or-no category. Fuzzy logic, by contrast, embraces partial truth. Instead of declaring a text simply human or machine, it assigns degrees of membership to overlapping linguistic categories and then reasons over them using transparent if–then rules. This makes the framework a natural fit for a task where the boundary between human and AI prose is inherently graded rather than sharp.

The team selected six stylistic and linguistic parameters as inputs to their system: lexical variability, content word ratio, lexical density, keyword density, the Coleman–Liau readability index and the Flesch reading ease score. Each metric captures a different facet of how a sentence is built. Lexical variability measures the richness of vocabulary in use, while the content word ratio tracks the proportion of meaning-bearing words relative to function words such as articles and prepositions. Lexical density and keyword density gauge how much informational payload a sentence carries, and the two readability indices estimate how cognitively demanding the text is for a reader. Together, these six numbers form a stylistic fingerprint that machine-generated and human-written sentences tend to occupy differently.

Mapping these raw metrics into fuzzy categories is where triangular membership functions come in. Rather than drawing hard cutoffs—for instance, declaring that a lexical density above 0.6 is always high—the researchers defined overlapping triangular regions labeled low, medium and high. A given sentence’s measured value can partially belong to two adjacent categories at once, with membership degrees summing to one. This soft partitioning preserves the ambiguity that binary classifiers destroy. Once the six inputs are fuzzified, a set of if–then rules fires: if lexical density is high and readability is low, then the sentence is likely machine-generated, and so on. The rules were designed to reflect known stylistic tendencies of large language models, which often produce unusually uniform, informationally dense and smoothly readable prose.

The output of the rule engine is a human-likeness score, or HL score, a single number that quantifies how closely a sentence resembles authentic human writing. Unlike a deep learning detector that spits out a probability with no visible reasoning, the HL score can be traced backward through the rule base: one can see exactly which linguistic parameters pushed the verdict in which direction. This traceability matters enormously for real-world accountability. In academic integrity cases, journalistic fact-checking and legal disputes over authorship, a detector that can explain itself is far more defensible than a black-box neural network whose confidence cannot be interrogated.

To stress-test the framework, the researchers evaluated it on five datasets, each containing 500 sentences and drawn from radically different linguistic domains: a Play Store app reviews dataset, a financial sentiment dataset, a Twitter sentiment dataset, an IMDB movie reviews dataset and a social media dataset. Four of the five are publicly available on Kaggle, while the fifth can be obtained from the corresponding author on reasonable request. This domain diversity is crucial. Many AI-text detectors perform well on one genre but collapse when the writing style shifts—from movie reviews to financial filings, for example. A robust detector must generalize across the stylistic spectrum of everyday digital communication.

The results were striking. On the IMDB reviews and Twitter sentiment datasets—collections characterized by lengthy, structurally complex sentences—the fuzzy framework achieved 100 percent accuracy, correctly classifying every single sentence. On the remaining datasets, performance varied, reflecting the challenge of short, formulaic texts where human and machine writing styles converge. The pattern suggests that the six linguistic parameters are most diagnostic when sentences are long enough for stylistic tendencies to fully express themselves. In short, casual snippets, the stylistic gap between human and AI prose narrows to near-invisibility, and even the best detectors face fundamental limits.

The researchers argue that their findings position fuzzy inference as a robust and effective alternative to black-box deep learning detectors. Deep learning approaches, including bi-LSTM networks with attention mechanisms and multi-level contrastive learning systems cited in the literature, can achieve high accuracy but offer little insight into why a text was flagged. The fuzzy framework, by contrast, is flexible and comprehensive: new parameters can be added, rules can be tuned per domain, and the entire reasoning chain remains visible. This adaptability is particularly valuable as language models evolve and the stylistic signatures of AI text shift with each new generation of models.

The implications extend well beyond academic curiosity. The study situates AI-text detection within a broader landscape of concerns about disinformation, fake news and AI-generated spam reviews—problems that have grown acute as generative models become accessible to anyone. A systematic literature review on fake news during the COVID-19 pandemic, cited by the authors, underscores how urgently valid, reliable and accountable digital communication needs tools for verifying textual authenticity. Plagiarism detection, fake news identification and spam review filtering are all listed among the paper’s keywords, signaling the practical arenas where this technology could soon be deployed.

There are caveats. The evaluation datasets were modest in size, each holding 500 sentences, and the perfect accuracy on two datasets should be interpreted in that light. The framework’s reliance on surface-level stylistic metrics also means it could potentially be evaded by a writer—or a language model—deliberately mimicking human-like variability and readability profiles. Nonetheless, the study demonstrates that fuzzy inference frameworks can effectively capture complex stylistic variances between AI-generated and human-written text, offering a transparent, adaptable and computationally lightweight tool for a problem that will only grow more urgent. As the arms race between text generators and text detectors intensifies, sometimes the answer is not a bigger neural network but a smarter set of rules.

Subject of Research: A fuzzy logic-based framework for detecting and differentiating AI-generated text from human-written text

Article Title: A fuzzy logic-based framework for detecting and differentiating AI-generated text from human-written text

Article References: Kaushik, P. K., Bisht, R. K., & Sahoo, A. K. (2026). A fuzzy logic-based framework for detecting and differentiating AI-generated text from human-written text. International Journal of Data Science and Analytics, 22(1), 346. https://doi.org/10.1007/s41060-026-01338-9

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01338-9

Keywords: fuzzy, logic-based, framework, detecting, differentiating, AI-generated, text, human-written, scientific research