Why Each AI Chatbot Has Its Own Distinctive Writing Style - And Why European Educators Must Pay Attention

Forensic linguists have long used stylistic analysis to identify human authors. The same techniques now reveal that AI models exhibit consistent, differentiated patterns that can be quantified with statistical rigour.

A recent comparative study analysed hundreds of essays on diabetes generated by both ChatGPT and Gemini, applying the Delta method, a forensic technique originally developed by the corpus linguist John Burrows. Researchers calculated linguistic distances between writing samples, and the results were unambiguous. A 10% sample of ChatGPT's essays scored 0.92 against ChatGPT's full dataset, but 1.49 against Gemini's output. Gemini showed equally distinct patterns, scoring 0.84 against its own samples and 1.45 against ChatGPT's work.

These Delta scores confirm that each system writes with a statistically identifiable voice. The patterns emerge most clearly in trigrams: three-word combinations that expose each model's stylistic preferences. ChatGPT gravitates towards formal, clinical phrasing such as "blood glucose levels," "individuals with diabetes," and "characterised by elevated." Gemini, by contrast, opts for conversational expressions: "high blood sugar," "blood sugar control," and "the way for."

The idiolects are not random quirks. They emerge from each model's training data, architectural choices, and fine-tuning processes. Professor Mirella Lapata, a computational linguist at the University of Edinburgh and one of Europe's leading voices on natural language generation, has noted that the stylistic consistency of large language models is a predictable consequence of optimising on large, stylistically uneven corpora. The model learns not just language, but the modal register of its training set.

Why This Matters for European Education

The implications for European schools and universities are direct and pressing. Academic integrity bodies across the EU and UK have spent the past two years chasing AI detection tools that are, frankly, unreliable when used in isolation. Stylometric analysis offers a more principled complement to blunt-instrument detectors such as Turnitin's AI classifier.

Understanding that ChatGPT tends towards textbook formality while Gemini adopts conversational tones gives examiners meaningful contextual clues. A university essay that reads like a clinical guidelines document, yet was submitted by a first-year student, warrants a different conversation than one that reads like a lifestyle blog. Neither is automatically dishonest, but both deserve scrutiny calibrated to the model's known voice.

The EU AI Act, which entered into force in August 2024 and is being overseen by the newly established AI Office in Brussels, explicitly addresses transparency in AI-generated content. Article 50 of the Act requires deployers of AI systems that generate synthetic text to ensure that output is labelled as artificially generated. Stylometric identification tools could serve as a technical backstop where labelling fails or is deliberately circumvented.

Dragoș Tudorache, the Romanian MEP who co-led the European Parliament's negotiations on the AI Act, has consistently argued that transparency in AI-generated content is not a bureaucratic nicety but a foundation of democratic literacy. Stylometric research provides the empirical grounding that makes such transparency enforceable in practice.

The Practical Breakdown: How ChatGPT and Gemini Differ

The stylistic contrast between the two dominant models maps onto recognisable registers that any teacher or content editor will recognise immediately:

• Medical and technical terminology: ChatGPT favours clinical precision, using "glucose," "elevated levels," and "pathophysiology." Gemini prefers accessible equivalents: "sugar," "high blood," and "how it works."
• Sentence structure: ChatGPT defaults to formal, academic construction. Gemini constructs shorter, more conversational sentences aimed at a general audience.
• Audience assumption: ChatGPT writes as though addressing a professional or advanced student. Gemini writes as though addressing an informed member of the public.
• Hedging and qualification: ChatGPT applies layered caveats in a manner consistent with academic writing conventions. Gemini tends to state conclusions more directly.

These differences have real consequences for how AI-assisted work presents itself in an educational context. A student using ChatGPT to draft a biology essay will produce something that superficially resembles a textbook excerpt. A student using Gemini will produce something that reads more like a well-researched magazine article. Both may be equally plagiaristic in intent, but they require different detection strategies.

Beyond Detection: Broader Applications Across the European Sector

The relevance of AI idiolect research extends beyond academic integrity checks. Several practical applications are emerging across European institutions and commercial operators:

• Educational integrity: Universities and examining bodies can use stylometric fingerprinting as one layer of a multi-method detection approach, alongside metadata analysis and viva voce examination.
• Content authentication: Publishers, news organisations, and regulatory bodies can trace AI-generated copy to specific systems, supporting compliance with the AI Act's transparency requirements.
• Brand and tone alignment: European companies deploying customer-facing chatbots can select models whose natural register matches their communications strategy, rather than relying on costly prompt engineering alone.
• Version tracking: Developers and procurement teams can monitor whether a model update has materially shifted stylistic behaviour, which matters for regulated sectors such as financial services and healthcare.
• Multilingual research: Does ChatGPT maintain its formal register when operating in French or German? Early evidence suggests models may develop language-specific sub-idiolects, a question of direct relevance to multilingual EU institutions.

Research groups at ETH Zurich, which has been active in AI safety and interpretability work, and at the Alan Turing Institute in London are well placed to extend stylometric studies into European language contexts. The field is young, and the methodologies proven in English need rigorous replication across the EU's 24 official languages before the findings can be applied with full confidence in, say, a French grandes écoles examination or a German Abitur.