The relationship between human language and AI models has evolved into a complex feedback loop. While models are trained on human-generated data, recent observations suggest that human writing styles are increasingly being influenced by the patterns used by AI.
The Feedback Loop of AI and Human Writing: How Models are Shaping Our Language
Research highlights how specific linguistic traits become entrenched through model training. For example, researchers at Florida State found that the frequent use of certain words, such as "delve," was driven by human reviewers during the reinforcement learning stage. This effect is echoed by studies from the Max Planck Institute, which linked the rise of specific vocabulary to the release of ChatGPT.
This influence works in both directions. As certain words become associated with AI-generated content, humans have begun to self-censor or edit their prose to avoid being flagged by AI detectors. This phenomenon contributes to what researchers call "model collapse," where the diversity of language is sanded down, leaving only the most average and ordinary versions of expression.
The consequences extend to professional domains. In design and technical writing, the push for machine-readable and standardized structures can lead to a loss of nuance. The ongoing tension lies in the fact that the very writing skills humans spent years perfecting—clarity, structure, and meticulous word choice—are the same patterns that modern AI models replicate, making it increasingly difficult to distinguish between human craftsmanship and machine output.
Sources
- I can't tell who's teaching who anymore （Hacker News Frontpage, 2026-09-30）