Three years ago, "delve" was a word you'd encounter in academic papers and Victorian fiction. By late 2024, researchers tracking language patterns in peer-reviewed journals had documented a sharp increase in its use - concentrated in papers published after GPT-4's March 2023 release. Linguists studying the shift found several AI-associated words appearing in published academic work at rates 5 to 10 times their pre-ChatGPT baselines. The words themselves hadn't become more useful. The people writing the papers had absorbed the default vocabulary of the AI tools they were using daily.
The same drift is showing up in workplace writing, marketing copy, and casual communication.
The Patterns That Give It Away
People who work heavily with AI tools are picking up specific linguistic fingerprints, usually without noticing. The most commonly observed:
- The construction "not just X, it's Y" ("not just a productivity tool, it's a new category")
- Compulsive deployment of "era" ("we're in the reasoning era", "the agent era is here")
- Affirmative one-word openers: "Absolutely." "Certainly." "Great question."
- Bullet-point structuring of thoughts that don't need bullets
- Sentences that open with "Note that..." or "It's important to understand..."
These patterns exist because large language models - AI systems trained on billions of text samples to predict statistically likely next words - learned them from the most common writing found online. Spending hours daily reading AI output means absorbing that rhythm, the same way you'd internalize the speech patterns of people you work closely with.
Detection is getting harder. AI writing detection tools were designed to flag fully AI-generated content, but they struggle with AI-influenced writing - text where a human is the primary author who has absorbed AI patterns. The binary of "AI wrote this / human wrote this" is giving way to a spectrum the current generation of detectors wasn't built to navigate.
The Feedback Loop
This creates a structural problem for the field. Training data for future models includes current human text. If that text increasingly reflects patterns from earlier AI generations, the next model trains on AI-inflected writing. Each generation potentially produces more homogenized output - not by design, but as a statistical artifact of the training pipeline.
For working writers and content creators, the practical question is whether your voice has drifted. The signal is usually in sentence structure and word choice. If your drafts consistently follow a rhythm you didn't have three years ago, and you've spent the past two years working closely with AI tools, the influence is likely measurable.
Some people are running deliberate calibration exercises: drafting one piece per week without AI assistance, reading finished work aloud (AI-generated text often sounds stilted when spoken), or keeping examples from their pre-AI work as a style reference to write against.
The concern isn't that AI tools make writing worse. It's that they make it converge. A statistically average voice has become much easier to produce. For anyone whose work depends on a distinct voice - journalists, marketers, authors, consultants - the question is whether the efficiency trade-off is worth the drift.