Expert Says There Is No Error-Free Way to Detect AI-Generated Texts
Modern neural networks have learned to convincingly imitate human writing, making detection increasingly difficult.
Distinguishing human-written text from machine-generated content is becoming harder, and no universal solution exists. That assessment was shared with IT Russia by Yury Chekhovich, a PhD in physics and mathematics, an expert in academic ethics, machine learning, and AI, head of Laboratory No. 42 at the Institute of Control Sciences of the Russian Academy of Sciences, and founder of the academic integrity verification service Domate.
At the core of the Domate service is a multi-stage text analysis. Dozens of algorithmic modules examine information extracted from the uploaded document from different angles. At the final stage, the results produced by these modules are combined into an overall assessment report.
He explained that modern detection tools are tuned to look for markers such as unusual words rarely used by humans; shifts in the frequency of parts of speech, such as an excess of verbs or adjectives compared with natural language; syntactic and stylistic distortions; and differences in word order. In academic and scientific texts, AI algorithms also tend to produce sentences that are excessively long and overloaded.
Another important marker in academic writing is the list of sources cited by the author. Because generative AI often produces fictitious references, their presence in a bibliography can point to the use of machine-generated content.
Analytics Matter More Than “Catching” AI
According to the expert, it is already clear that no absolute, error-free detection criterion exists, because detection algorithms operate on probabilities. In practice, a detector can only indicate that a fragment of text resembles those typically produced by generative models. At the same time, modern neural networks have become highly skilled at imitating human style, and generative technologies continue to evolve.
