AI writing tools are no longer a niche experiment. A Stanford University-led study of more than 300 million text samples found signs of large language model assistance across consumer complaints, corporate press releases, job postings, and United Nations press releases.
The research, which tracked writing from January 2022 to September 2024, suggests that generative AI has moved into routine communication. It also found a pattern that runs against a common assumption about new technology: in some cases, places with lower educational attainment appear to be using AI writing tools more heavily than places with higher educational attainment.
A Large Look At Everyday AI Writing
The study, titled The Widespread Adoption of Large Language Model-Assisted Writing Across Society, was led by researchers from Stanford, the University of Washington, and Emory University. Weixin Liang and Yaohui Zhang from Stanford served as lead authors, with collaborators Mihai Codreanu, Jiayu Wang, Hancheng Cao, and James Zou.
The dataset was unusually broad. It included 687,241 consumer complaints submitted to the US Consumer Financial Protection Bureau (CFPB), 537,413 corporate press releases, 304.3 million job postings, and 15,919 United Nations press releases.
Across those sources, the researchers found signs that AI language models were helping write a meaningful share of text. Their estimates included roughly 18 percent of financial consumer complaints, 24 percent of corporate press releases, up to 15 percent of job postings, and 14 percent of UN press releases.
Those figures do not mean every word in those documents was generated by an AI system. The study looked for evidence of AI assistance in writing, not full automation. That distinction matters because many people may use a language model to draft, revise, polish, or organize a message before submitting it.
How The Researchers Detected Patterns
The study does not claim that AI detectors can reliably judge a single document. Instead, it uses a population-level method. The researchers compared large groups of texts from before and after ChatGPT’s release in late 2022, looking for shifts in word frequencies and linguistic patterns.
The basic idea is that large language models tend to produce different word choices, sentence structures, and writing patterns than typical human writing. Those differences may be subtle in any single complaint, press release, or job post. But across millions of documents, the changes can become visible enough to estimate adoption rates.
To test the approach, the researchers created test sets with known percentages of AI content, from zero percent to 25 percent. Their method predicted those percentages with error rates below 3.3 percent, which gave them confidence in the aggregate estimates.
The researchers also warned that their numbers likely represent a minimum level of AI use. Heavily edited AI-assisted writing can be harder to detect, and more advanced AI-generated content may leave fewer obvious signals. As a result, the true level of generative AI use may be higher than the reported adoption rates.
The Education Pattern Was The Surprise
One of the clearest findings came from consumer complaints submitted to the CFPB. The researchers found that urban areas had higher adoption overall, at 18.2 percent, compared with 10.9 percent in rural areas. That part fits a familiar pattern in technology adoption.
The less expected result appeared when the researchers compared regions by educational attainment. Areas with lower educational attainment showed 19.9 percent adoption, while higher-education areas showed 17.4 percent adoption. Within urban areas, the same direction held: less-educated communities showed 21.4 percent adoption versus 17.8 percent in more educated urban areas.
That matters because technology adoption has often been associated with urban areas, higher-income groups, and people with higher levels of educational attainment. In this case, the consumer complaint data pointed to a different possibility: AI writing tools may help people prepare formal communications in situations where writing skill, time, or confidence can affect how clearly a complaint is presented.
The state-level variation was also large. Arkansas showed the highest adoption rate at 29.2 percent, based on 7,376 complaints. Missouri followed at 26.9 percent, based on 16,807 complaints, and North Dakota reached 24.8 percent, based on 1,025 complaints.
At the low end, West Virginia showed 2.6 percent, Idaho 3.8 percent, and Vermont 4.8 percent. Major population centers were closer to the middle, with California at 17.4 percent based on 157,056 complaints and New York at 16.6 percent based on 104,862 complaints.
Business, Jobs, And Institutional Communications
The same research found that AI-assisted writing was not limited to consumer advocacy. Corporate communications and labor-market writing also showed signs of adoption.
Across the sectors analyzed, adoption rose sharply beginning three to four months after ChatGPT’s November 2022 launch, then stabilized in late 2023. That pattern appeared in consumer complaints, corporate communications, and job postings.
In job postings, organization age stood out as the strongest predictor of AI writing usage. Companies founded after 2015 showed adoption rates up to three times higher than firms established before 1980. In certain roles, newer companies reached 10–15 percent AI-modified text, compared with below 5 percent for older organizations.
Company size also appeared to matter. Small companies with fewer employees incorporated AI more readily than larger organizations.
Corporate press releases showed another split. Science and technology companies used AI most extensively, reaching a 16.8 percent adoption rate by late 2023. Business and financial news showed 14–15.6 percent adoption, while people and culture topics showed 13.6–14.3 percent.
What This Means For The Future Of Writing
The study presents AI writing tools as part of the infrastructure of communication. They are being used in consumer complaints, corporate announcements, job postings, and international organization press releases. The adoption is not uniform, but it is broad.
The most important implication is not just that more text may be AI-assisted. It is that the benefits and patterns of use may not follow the usual map of technology diffusion. In the consumer complaint data, AI writing tools appeared especially relevant in places where educational attainment was lower.
That does not erase the urban-rural divide, which remained visible in the study. But it does suggest that writing assistance can play a practical role for people trying to communicate with institutions. If a person needs to make a complaint, describe a problem, or produce formal language, a large language model may lower the barrier to getting that message written.
For companies, public institutions, and platforms that receive written submissions, the research points to a changed communication environment. AI-assisted text is already present in large volumes, and the reported figures may understate the real level of use. The future of professional writing may be less about whether AI is involved, and more about how openly, effectively, and responsibly it is used.