Karolina Rudnicka, PhD, a linguist from the university’s Faculty of Philology, said it is currently impossible to use an algorithm to establish with certainty that a text was generated by a large language model (LLM).

‘There is no such tool. There are many algorithms that check the origin of texts, but they are not able to determine it 100%’, she told the Polish Press Agency.

The issue is becoming increasingly important as the use of AI-generated content expands. The provisions of the EU AI Act regulation concerning the Code of Practice on Transparency of AI-Generated Content recently came into force in all EU countries.

From August 2, entities using machine systems to create content are obliged to inform recipients that the material was generated or modified using AI.

Rudnicka said that the growing sophistication of AI models is also making their texts harder to identify by their language.

There are fewer dashes, distinctive styles and particular words that might previously have revealed the ‘machine’ origin of a text.

At the same time, she said, the presence of features associated with AI does not necessarily prove that a text was written by a machine.

‘The language people use can also change because of AI. If we often use chatbots, we borrow sentence structures or certain expressions from them. Changes that occur naturally in every language can significantly accelerate under the influence of LLMs’, the linguist said.

Rudnicka is the author of the paper ‘Each AI Chatbot Has Its Own Distinctive Writing Style - Just as Humans Do’, published in 2025 in the journal Scientific American.

People speak and write differently depending on factors including age, gender, native language and education. This individual style is known as an ‘idiolect’.

‘The style of expression of various language models bears the hallmarks of idiolect. In my work from 2025, I called it a +chatolect+’, Rudnicka said.

Her latest research, conducted with Thomas Juzek, PhD, from Florida State University, confirms that different AI models have different writing styles.

The researchers analysed datasets containing hundreds of short English-language texts generated by ChatGPT, Gemini, Claude and several other AI-based tools.

‘We examined the differences between them at the levels of sentence structure, text structure and lexical level. The models wrote on four topics: health disinformation, fear of math, global warming and climate change. We found that the AI language varied not only between models, but also model generations. Chat GPT 3.5, introduced two years ago, wrote differently than the currently released Chat GPT 5.5. We were also interested in whether newer versions of LLM +inherited+ chatolects from their predecessors. And we found out that they did not’, the researcher said.

Rudnicka also referred to an article published last week in the British weekly The Economist, titled ‘How to spot AI writing’. As AI-generated texts have become widespread and have an impact on many branches of the economy, the publication conducted its own linguistic research.

‘Journalists from The Economist contacted me with questions about the features of the AI language, and about the distinctive styles of various language models. Based on my research, I told them that just like writers have their own styles and favourite words, bots can have them too’, she said.

There are further linguistic differences between texts produced by individual AI models, Rudnicka said.

‘For example, the length of sentences. We can consciously control the words we use and whether we speak formally or casually. But we do not have full control over the length of sentences or the number and type of function words, i.e. words that do not have their own meaning, such as conjunctions or phrasemes (as if, for sure, in fact) - but give the text an individual character. Their analysis is used, for example, to determine the authorship of literary works, and in forensic linguistics’, Rudnicka said.

The researchers found that chatbots differ in the average length of the sentences they generate and in their punctuation.

But beyond the differences between individual models, they also identified similarities that appear to be common to AI-generated texts.

These include the use of the ‘rule of three’ — a three-part enumeration such as ‘People need health, love, and safety to be happy’ — and figures of speech such as ‘both X and Y’.

Rudnicka said people who frequently use AI may become increasingly sensitive to such features.

She referred to recent research suggesting that humans are better at detecting AI-generated texts than algorithms specifically designed for that purpose.

‘Users who frequently use LLM are sensitive to certain signs of their creativity. Therefore, I believe that in the future, frequent AI users and AI experts will still be able to distinguish texts created by humans from those generated by machines’, Rudnicka said.

Anna Bugajska (PAP)

abu/ agt/

tr. RL