AI Writing Tools See Faster Adoption in Less-Educated US Areas, Study Finds

The Widespread Adoption of Large Language Model-Assisted Writing Across Society

Since the launch of ChatGPT in late 2022, experts have debated the impact of AI language models on the world. New research from Stanford University, the University of Washington, and Emory University, published on the arXiv preprint server, analyzes over 300 million text samples to understand these trends. The study, titled "The Widespread Adoption of Large Language Model-Assisted Writing Across Society," reveals that AI language models are now assisting in writing up to a quarter of professional communications, with a notable impact on less-educated areas of the United States.

Key Findings:

Significant Reliance on Generative AI: The research indicates a new reality where firms, consumers, and international organizations substantially rely on generative AI for communications.
Broad Adoption Across Sectors: AI language models are assisting in writing a significant portion of professional communications, with adoption rates varying across different sectors.
Contradictory Adoption Patterns: Contrary to typical technology adoption trends, areas with lower educational attainment are using AI writing tools more frequently than more educated populations.

Methodology:

The researchers tracked Large Language Model (LLM) adoption from January 2022 to September 2024. Their dataset included:

687,241 consumer complaints submitted to the US Consumer Financial Protection Bureau (CFPB).
537,413 corporate press releases.
304.3 million job postings.
15,919 United Nations press releases.

They employed a statistical detection system that analyzed word usage patterns to identify signs of AI assistance. This approach, based on a previously released framework, tracks shifts in word frequencies and linguistic patterns before and after ChatGPT's release.

AI Assistance Rates:

The study found the following approximate rates of AI assistance:

Financial Consumer Complaints: Roughly 18 percent, with a notable 30 percent in Arkansas.
Corporate Press Releases: 24 percent.
Job Postings: Up to 15 percent.
UN Press Releases: 14 percent.

Urban vs. Rural and Education Levels:

Urban vs. Rural Divide: While urban areas initially showed similar adoption rates to rural areas, urban adoption surpassed rural adoption by mid-2023 (18.2 percent in urban vs. 10.9 percent in rural).
Educational Attainment: Areas with lower educational attainment exhibited higher AI writing tool usage (19.9 percent) compared to areas with higher educational attainment (17.4 percent). This pattern held true even within urban areas.

AI as "Equalizing Tools":

The researchers suggest that AI writing tools may be acting as "equalizing tools," providing a boost to individuals who may lack extensive educational experience, particularly in consumer advocacy.

Corporate and Diplomatic Trends:

Sectoral Adoption: Science and technology companies led in AI integration (16.8 percent), followed by business/financial news (14-15.6 percent) and people/culture topics (13.6-14.3 percent).
Organizational Age: Companies founded after 2015 showed adoption rates up to three times higher than those established before 1980.
Company Size: Smaller companies adopted AI more readily than larger ones.
International Adoption: Latin American and Caribbean UN teams showed the highest adoption (around 20 percent), with other regions showing moderate increases.

Implications and Limitations:

Detection Challenges: The study acknowledges the unreliability of AI writing detectors on a document-by-document basis. However, aggregate analysis reveals telltale patterns.
Underestimation of AI Use: The detection methods may underestimate true AI usage, as heavily edited or sophisticated AI-generated content can evade detection.
Societal Impact: The increasing reliance on AI-generated content poses challenges for communication authenticity and credibility, potentially leading to public mistrust.
Future of Communication: Distinguishing between human and AI writing is becoming increasingly difficult, with significant implications for all forms of communication.

About the Authors:

The study was led by Weixin Liang and Yaohui Zhang from Stanford, with collaborators Mihai Codreanu, Jiayu Wang, Hancheng Cao, and James Zou from the University of Washington and Emory University.

Benj Edwards, Ars Technica's Senior AI Reporter, authored this article. He is also a tech historian with nearly two decades of experience.