A 2024 lexical analysis of academic papers shows the usage rate of the word 'delve' doubled in two years — large language models are quietly rewriting human expression habits.
What this is
This study tracked vocabulary changes in English writing before and after the popularization of large language models. Words like 'delve,' 'tapestry,' 'landscape,' and 'moreover' have seen their frequency in academic papers and business reports deviate significantly from historical trends — and they happen to be ChatGPT's preferred expressions. As more people use AI-assisted writing, these 'AI-speak' terms flow back into human language, forming a self-reinforcing loop. RAG (Retrieval-Augmented Generation, the technique that has LLMs search external sources before answering) makes the problem worse — what it retrieves first is often content AI has already generated.
Industry view
Linguists and AI researchers are sharply divided. Supporters argue that large models have lowered the writing barrier for non-native speakers, and a certain convergence in word choice is a reasonable price for technological democratization. But critics point out that homogenization is eroding language's value as both identity marker and thinking tool. More research warns: once AI-generated text exceeds a threshold share, the data used to train the next generation of models will be polluted by their own output — known as 'Model Collapse.' The practical risk is equally glaring: companies invest heavily in brand voice, yet what employees produce with AI reads exactly like their competitors'.
Impact on regular people
For enterprise IT: content moderation must evolve from 'detect AI-generated' to 'detect language homogenization' — brand compliance now has a new dimension. For individual careers: writing text that 'doesn't sound like AI' is becoming a scarce skill; 'AI-speak' in résumés and reports can now be a liability. For consumer markets: user patience for cookie-cutter marketing copy is declining, and 'authentic human' content is starting to command a premium.