When OpenAI released ChatGPT in November 2022, it didn't just change how people write — it introduced a new and identifiable writing style into the world. That style comes from two things: the enormous corpus of text the model was trained on, and the RLHF (reinforcement learning from human feedback) fine-tuning process that shaped how it responds. Understanding those origins is the key to understanding why ChatGPT writes the way it does, and why it is detectable. Our free AI detector is built on exactly this research.
The Kobak Research: What the Science Actually Found
In 2025, Dmitry Kobak and colleagues published "Delving into ChatGPT usage in academic writing through excess vocabulary" in Science Advances — one of the most significant empirical contributions to AI detection research to date. The methodology was elegant: analyse the frequency of every word in millions of PubMed biomedical abstracts published before and after November 2022, and look for words that changed in frequency more than expected by natural trend.
What they found was striking. Seventy-four words showed statistically anomalous increases in the post-ChatGPT period. These increases could not be explained by normal vocabulary drift; the timing corresponded precisely with ChatGPT adoption among academic writers. Key findings:
- "delve" — increased approximately 28× its pre-ChatGPT baseline frequency in academic abstracts
- "meticulous" — increased 16× in contexts where it was previously rare
- "commendable" — a word almost never used in scientific writing, suddenly ubiquitous in peer review
- "tapestry" — used as a metaphor at rates that have no precedent in the corpus
- "intricate", "pivotal", "testament" — all showed similar spikes
The research estimated that by early 2024, approximately 10% of PubMed abstracts showed signs of AI involvement — likely an undercount, since it only catches cases where the AI vocabulary was not edited out. The implications for AI detection are profound: these word frequencies serve as statistical evidence that does not depend on any single signal but compounds across a document.
The 5 Core ChatGPT Writing Patterns
Pattern 1: Elevated Academic Vocabulary
ChatGPT consistently reaches for elevated, formal, academic vocabulary even in casual contexts. Ask it to write a friendly email and it will produce something that sounds like a journal abstract. This happens because the model was trained on text that skews heavily toward formal published writing — academic papers, Wikipedia, books, news articles — rather than the casual conversational writing that makes up most human digital communication.
The pattern is self-reinforcing: RLHF raters often preferred responses that sounded sophisticated, so the model was rewarded for elevated vocabulary, even when the context did not call for it.
Pattern 2: Balanced Hedging
ChatGPT constructs sentences that hedge in two directions simultaneously. The structure "while X, it is important to note Y" or "although A, B nevertheless remains" appears constantly. This is not how humans debate or argue; it is a learned behaviour from RLHF training that penalised one-sided responses on contested topics. The result is writing that always sounds reasonable and never takes a real position — detectable by its refusal to commit.
Pattern 3: Formal Transitions as Paragraph Openers
In human writing, transition words like "Furthermore," "Moreover," and "Consequently," are used mid-sentence to connect clauses. In ChatGPT writing, they are used to open paragraphs — turning every paragraph boundary into an explicit logical marker. Reading a ChatGPT document feels like reading a very structured essay because every paragraph is explicitly signposted. Count the paragraph-opening transitions; more than two in a five-paragraph piece is a strong signal.
Pattern 4: Comprehensive Coverage
ChatGPT never misses a standard perspective. Ask it to discuss a topic and it will cover every major angle, objection, and counterpoint in a systematic way. Human writers prioritise, omit, and emphasise based on personal judgment and experience. AI covers the waterfront. This means AI writing is often broader but shallower than good human writing, and lacks the opinionated selections that give human prose personality.
Pattern 5: The Conclusion Ritual
Possibly the most reliable single signal: ChatGPT ends almost every multi-paragraph response with an explicit summary paragraph beginning "In conclusion," "In summary," or "To summarize." This reflects the structure of academic writing that dominated its training data. Human writers in most contexts — blog posts, emails, reports, essays — do not write formal conclusions. They end with an action item, a reflection, a punchline, or they simply stop. The presence of a formal concluding paragraph in a casual context is near-diagnostic for ChatGPT authorship.
How ChatGPT's Patterns Have Changed Across Versions
GPT-3.5, the model underlying the original ChatGPT, showed these patterns in their most obvious form. Sentence structures were highly uniform, vocabulary choices were extremely elevated, and the conclusion ritual was almost universal. GPT-4 improved on naturalness significantly: sentence variety increased, hedging became more nuanced, and some of the most obvious vocabulary tells were toned down.
GPT-4o represents the current frontier. It is considerably better at mimicking human writing styles, especially when prompted to adopt a casual voice. However, the core patterns persist. The vocabulary fingerprint identified by Kobak et al. remains present because it is embedded in the model's weights, not just a surface-level stylistic choice. Even GPT-4o uses "delve," "meticulous," and "tapestry" at rates that exceed typical human baseline frequencies.
The key insight: as models improve, individual signals become weaker, but the combination of signals remains statistically meaningful. A good detector, like our ChatGPT detector, uses multiple signals simultaneously rather than relying on any single word or pattern.
The "Delve" Problem: A Deep Dive
It is worth spending a moment on "delve" specifically, because it has become something of a cultural symbol for the AI detection field. Before November 2022, "delve" was a niche word used primarily in formal academic writing — think archaeology papers, literary criticism, and philosophy. It was not a word that appeared in everyday business communication, student essays, or casual journalism.
After ChatGPT launched, "delve" began appearing everywhere: in student essays, marketing copy, LinkedIn posts, scientific abstracts, customer service emails. The word had been prominent in ChatGPT's training data in elevated academic contexts, and the model learned to use it as a marker of intellectual seriousness. Ask ChatGPT to make something sound more sophisticated, and "delve" is frequently what you get.
The irony is that the word now functions as an anti-sophistication signal. Experienced editors and readers have learned to associate "delve" with AI output, making it counterproductive even as a stylistic choice. It remains the single most reliable one-word indicator of ChatGPT authorship in contexts where it would not naturally appear.
How to Verify: A Practical Workflow
For a quick manual check, open the suspicious document and use Find & Replace (Ctrl+H or Cmd+H) to search for each of these terms and note the count:
- "delve" — more than 0 occurrences in a typical essay: suspicious
- "meticulous" — more than 1: suspicious
- "tapestry" — any occurrence in non-literary writing: very suspicious
- "testament" — more than 1: suspicious
- "furthermore" as a paragraph opener — more than 2: suspicious
- "in conclusion" or "in summary" as a final paragraph opener: very suspicious
Then paste the full text into our free AI detector for a complete statistical analysis. The tool combines vocabulary analysis with sentence-level statistical patterns (burstiness, perplexity) to produce a single probability score. For a technical explanation of how those scores are calculated, see our how it works page.
Understanding ChatGPT's writing patterns isn't just useful for detection — it also makes you a better writer. The patterns that make AI writing identifiable are the same patterns that make it generic and forgettable. Avoiding them makes your own writing more distinctive, more human, and more effective.