Phrase Expands Language Intelligence for Humans and AI Agents
The new enhancements reflect a broader shift toward global content operations where humans and AI agents increasingly work side by side.
→ Continue ReadingArtificial Intelligence
Detecting AI-Generated Content
A hybrid linguistic approach
By Wojciech Wołoszyk and Marta Domaszk
T
he rapid advancement of artificial intelligence (AI) writing tools has created new challenges in verifying text authenticity. As language models improve, the differences between human-written and AI-generated content become harder to spot, raising important questions about text authenticity in professional contexts. This article presents a hybrid methodology for determining whether specialized texts were written by humans or generated by AI.
Our approach combines computational analysis with linguistic expertise to provide a probability assessment rather than a binary classification. For example, a conclusion might state that “this text was generated by AI with a probability of 94%.” This nuanced approach is particularly valuable in professional contexts in which content authenticity has significant implications for quality assurance, compliance, or intellectual property.
Advertisement
The Hybrid Methodology
Our methodology integrates automated analysis with research-based linguistic assessment:
- We conducted a comprehensive review of current research on AI text detection, examining studies by Georgiou (2024), Schaaff et al. (2024), Fraser (2024), and Muñoz-Ortiz et al. (2024).
- We discovered that Open Brain AI, a tool originally developed for analyzing neurological disorders, could be repurposed for text authentication. This platform analyzes multiple linguistic dimensions simultaneously but does not itself evaluate text authenticity.
- Based on research literature, we identified key distinguishing parameters for detection and assigned appropriate weights.
- We validated our approach through empirical testing on diverse text samples of varying lengths and domains.
The Open Brain AI platform was initially created by Charalambos Themistocleous to help neurologists, clinicians, and researchers analyze speech and language patterns associated with neurological conditions. Its ability to analyze text across six main areas — phonology, morphology, syntax, lexicon, semantics, and readability — provides the detailed measurements that form the foundation of our authentication methodology. Repurposing this clinical tool for text authentication represents an innovative cross-disciplinary approach to AI detection.
The practical implementation of this approach follows four steps:
- Analyze the text using Open Brain AI.
- Extract the comprehensive set of linguistic parameters.
- Apply the weighted formula to these parameters.
- Calculate the probability of AI generation.
Key Linguistic Parameters for Detection
Statistical and Probabilistic Parameters
We examine several text characteristics to compare human- versus AI-generated content (see Table 1). Human writing has higher perplexity (57.3 versus 37.8) and burstiness (0.61 versus 0.38) values than AI-generated text, reflecting more unpredictable and varied patterns. These differences stem from the fundamentally different processes behind text creation. Human writing emerges from creative cognition, which naturally produces more varied and less predictable linguistic patterns. AI-generated text, despite its sophistication, remains constrained by the statistical probabilities learned during training. The statistical signature of AI text — with its more predictable patterns and lower entropy — provides a reliable signal that persists even as language models become more advanced.
Morphological Features
One of the most reliable indicators of AI authorship is the distribution of different word types. Human writers use approximately 12.8% more verbs and 27.6% more adverbs, creating more dynamic language. AI systems employ 21.3% more nouns and 20.6% more adjectives, resulting in more descriptive but less active text.
Figure 1 shows the morphological distribution comparison; human-written texts have higher proportions of verbs, adverbs, auxiliaries, and adpositions (shown as negative percentage differences in red), whereas AI-generated texts demonstrate higher proportions of nouns, adjectives, and pronouns (shown as positive percentage differences in blue). These patterns serve as reliable indicators for identifying AI-generated content across various specialized domains.
This morphological fingerprint reveals a fundamental difference in how humans and AI systems approach language construction. Human writing tends to be more action-oriented and dynamic, and AI-generated text often produces more static, description-heavy content. This distinction likely reflects the different cognitive processes at work — humans naturally emphasize actions and their modifications, while AI systems, based on their training patterns, may prioritize entities and their attributes.
Syntactical Features
Sentence structure analysis reveals that human writers show more variation in how they construct sentences. They use more prepositional phrases and determiners, creating more diverse language patterns. AI-generated texts tend to follow more consistent, predictable patterns in sentence construction.
Human writers also exhibit what Fraser (2024) describes as “natural syntactic inconsistency” — variations in sentence construction that reflect the cognitive processes of human thinking. In contrast, AI-generated texts often display more consistent syntactic patterns, creating a subtle but detectable uniformity that can be identified through careful analysis.
When examining human-written text, we often find greater structural variability — some sentences are short and direct, while others are complex with multiple clauses. AI text tends toward a more uniform sentence structure with less natural variation. This “beneficial inconsistency” in human writing serves as a distinctive marker that helps differentiate it from AI-generated content.
Lexical Features
The content-to-function word ratio differs significantly between human and AI texts; see Figure 2. Humans maintain a balanced ratio (averaging 0.98), whereas AI-generated texts show a higher proportion of content words (averaging 1.37). This difference creates a subtle but detectable “heaviness” in AI-generated content, with greater emphasis on information-carrying terms at the expense of natural language flow.
Human writers also demonstrate greater vocabulary diversity with a type-token ratio of 55.3 compared with 45.5 for AI-generated texts, indicating that humans use a broader, more varied vocabulary, even when discussing technical topics. In specialized domains, these lexical differences become particularly noticeable. Human experts writing in their field of specialization typically employ a rich, varied vocabulary with nuanced terminology that reflects deep subject knowledge. AI-generated specialized texts often rely more heavily on common domain-specific terms and exhibit less lexical creativity within the specialized domain.
The lexical category, weighted at 25% in our methodology, provides the most reliable indicators for AI detection, according to Georgiou’s research. The consistent patterns in vocabulary diversity and content-to-function word ratios offer strong signals that help distinguish AI-generated content, even when other aspects of the text have been modified.
Phonological Features
Interestingly, the sounds of language also provide clues about text origin. As Figure 3 shows, AI-generated texts consistently show higher frequencies of various consonant types, typically 20% to 23% higher than in human-written texts. This pattern likely reflects statistical biases that emerge from language model training.
These phonological differences often persist even when content has been substantially edited, making sound patterns one of the more robust indicators in our authentication methodology. Even when an AI-generated text has been revised for content and structure, the underlying phonological distribution often remains detectably different from typical human writing patterns.
Semantic and Coherence Features
Human writing typically shows more coherent development of ideas with natural logical connections. AI-generated text, despite improvements, may still show inconsistencies in how meaning develops across longer passages. Muñoz-Ortiz et al. (2024) found that human writers express a wider range of emotions, particularly negative emotions, whereas AI content tends to be more positive overall.
In longer texts, human writers typically maintain more consistent references to entities and concepts, with fewer instances of ambiguous pronoun usage or coreference errors. AI-generated content, while improving rapidly, still demonstrates less semantic consistency across extended passages, particularly when complex topics involve multiple related concepts.
Human writing also tends to show more natural topical transitions, with ideas flowing logically but not mechanically from one to the next. AI-generated content often displays more formulaic transitions that follow predictable patterns, creating a subtle but detectable difference in how ideas progress through the text.
Frequency Analysis of Characteristic Terms
Certain phrases appear far more frequently in AI-generated content than in human writing, as Table 2 shows. Introduction phrases like “it is worth noting that” appear 4.6 times more often in AI-generated content, and terms like “embrace” and “leverage” appear 4.1 times more frequently.
This frequency analysis provides one of our most practical detection methods. The consistent overrepresentation of certain terms creates a detectable linguistic signature that often persists when other aspects of the text have been modified. While any individual term might appear in human writing, a pattern of multiple overused terms in a single text provides a strong indicator of potential AI generation.
Again, these linguistic markers reflect the statistical patterns learned during AI training. Words and phrases that appear frequently in training data tend to appear more often in the output, leading to their overrepresentation. Human writers, by contrast, typically employ a more varied approach to language, avoiding excessive repetition of specific terms or formulaic expressions.
Calculating AI Probability
Our approach uses a weighted system that integrates multiple linguistic indicators. We assign weights to parameter categories based on their reliability as indicators:
- Lexical parameters (25% weight) — vocabulary choices and diversity
- Morphological parameters (20% weight) — word types and parts of speech
- Syntactic parameters (15% weight) — sentence structure and grammatical patterns
- Phonological parameters (15% weight) — sound patterns and distributions
- Readability and coherence parameters (15% weight) — text flow and readability metrics
- Qualitative patterns (10% weight) — subjective linguistic features
We determine how closely each parameter matches typical human or AI patterns, assign a value from 0 (completely human-like) to 1 (completely AI-like), multiply by the assigned weight, and calculate the weighted average.
This is expressed mathematically as: P(AI) = (Σ(Wi × Ii)) / Σ(Wi) × 100%
Where:
- Wi is the weight of a given parameter.
- Ii is the AI indicator value for that parameter (0–1).
Table 3 shows a sample calculation from an actual analysis, where probability: 0.9375 × 100% = 93.75%. This high probability value of approximately 94% indicates an extreme likelihood that the analyzed text was generated by AI, based on the convergence of multiple independent indicators.
Our weighting system isn’t arbitrary; it’s based on extensive research showing which factors are most reliable for detection. Lexical parameters receive the highest weight because research by Georgiou (2024) found these to be the most reliable indicators, while morphological parameters have the second-highest weight based on studies by Schaaff et al. (2024) and Opara (2024).
The multidimensional approach helps overcome the limitations of any single indicator. While individual parameters might sometimes yield false positives or negatives, the weighted combination of multiple independent signals provides a more robust assessment that reduces the likelihood of misclassification.
Advertisement
Accuracy and Limitations
The reliability of our methodology varies with text length, as Figure 4 shows. For texts under 100 words, accuracy is limited to about 67%. As text length increases, accuracy improves dramatically, reaching 96% for texts exceeding 1,000 words.
This relationship between accuracy and text length reflects the fact that longer texts provide more linguistic data to identify patterns. For professional applications dealing with specialized texts (typically exceeding 500 words), our methodology provides highly reliable authentication with accuracy exceeding 90%.
Our validation testing confirms that, when properly applied, the methodology reliably distinguishes between human-written and AI-generated content. The multidimensional approach maintains effectiveness even as AI technologies evolve, as it focuses on fundamental linguistic differences rather than model-specific patterns.
Key limitations include the following:
- Effectiveness is limited with very short texts (under 100 words).
- Extensive human editing of AI-generated text can obscure some AI patterns, though many linguistic markers persist even after moderate editing.
- Evolving AI technology may change some linguistic markers.
In addition, while our methodology works well across multiple languages, different languages require specific calibration for optimal detection. Currently, English and Polish versions are the most refined, with research underway to extend the methodology to additional languages.
Conclusion
This hybrid methodology for text authentication integrates computational analysis with linguistic expertise to provide reliable probability assessments of AI involvement in text creation. By examining multiple linguistic features simultaneously — from statistical and probabilistic parameters to specific word usage patterns — our approach identifies the subtle but consistent differences that distinguish human writing from AI-generated content.
The methodology, derived from comprehensive analysis of current research and validated through empirical testing, achieves high accuracy rates, particularly for specialized texts exceeding 500 words. As language models continue to evolve, a multidimensional approach to text authentication will remain crucial. While specific linguistic markers may change, the fundamental differences in how humans and AI systems generate text provide enduring signals that can be used for authentication.
For professionals dealing with specialized texts where authenticity matters, this methodology offers a reliable tool for verification, contributing to greater transparency and accountability in an era of rapidly advancing language technologies.
Wojciech Wołoszyk is a lawyer-linguist, legal translator, and court expert on legal linguistics. He is the CEO of IURIDICO, chairman of POLOT Association of LSPs, and a coordinator for the World Law Dictionary PL Chapter. He specializes in legal terminology and secure language technology integration in translation workflows.
Marta Domaszk is the Head of Quality Assurance at IURIDICO Legal & Financial Translations and a member of the management board. She is an EU and legal translation specialist with 15 years of experience in the translation industry. She holds a degree in translation studies from the University of Warsaw.
Advertisement
Related Articles
Voiseed Announces the Appointment of Dr. Páraic Sheridan to Board of Directors
[Milano, September 14, 2023] — Voiseed, an innovative start-up specializing in expressive synthetic voices for Games & Media localization, today announced the appointment of Dr.…
→ Continue ReadingAlibaba Releases Qwen3 Models for Open-Source AI in Coding, Reasoning, and Translation
The Qwen3 AI models from Alibaba bring new capabilities in agentic coding, reasoning, and translation, expanding open-source AI development worldwide.
→ Continue Reading