Study Finds AI Writing Assistants Cause 'Blandification' — Neutral, Impersonal, Less Human

Peer-reviewed study from West Coast universities examined AI's effect on writing style and substance across 100 participants

Participants responded to 'Does money lead to happiness?' with and without AI assistance. The study used 2021 pre-LLM essays as a baseline to measure how AI editing diverges from human self-editing. Accepted to a workshop at a leading AI conference.

Heavy AI users produced essays 69% more likely to give neutral answers versus passionate responses from non/light AI users

The shift toward neutrality is a directional loss of argumentative conviction. Non-AI users and light users tended to take stronger, more personally grounded stances, while heavy users' essays converged toward balanced, non-committal positions.

AI-assisted essays contained 50% fewer pronouns, reflecting a measurable shift to impersonal, formal language

Pronoun reduction is a proxy for depersonalization — fewer first-person references means fewer anecdotes, personal experiences, and individual voice. This linguistic shift aligns with the broader blandification finding.

Lead author coined 'blandification' — AI pushing essays away from anything a human would have written

Natasha Jaques (CS professor at University of Washington, senior research scientist at Google DeepMind) argued that an ideal LLM should produce the essay the user would have written themselves, only faster. Current behavior substitutes a statistically averaged voice for a personal one.

Users rated AI-assisted essays as less creative and less their own voice, yet reported similar satisfaction with final outputs

This satisfaction gap is the study's most concerning long-term signal. If users habituate to lower-quality, homogenized writing while feeling equally satisfied, they may progressively lose both motivation and skill to write distinctively on their own.

Half of initial participants refused to use an LLM at all or used it only for information retrieval

This self-selection dynamic means the heavy-use findings apply to a subset willing to delegate writing to AI. Substantial resistance to AI writing assistance remains, affecting how broadly these effects will propagate.

Study tested Claude 3.5 Haiku, GPT-5 Mini, and Gemini 2.5 Flash — covering all three major commercial AI providers

Using models from Anthropic, OpenAI, and Google means the blandification effect is not attributable to any single vendor's model choices. The convergence across three major providers suggests the cause is structural rather than vendor-specific.

Paper also examines how AI use affects the criteria scientists apply when judging conference paper acceptance

This secondary finding extends the study's scope beyond consumer writing to scientific peer review — raising the question of whether AI is homogenizing how researchers evaluate and communicate ideas, not just how individuals write.