AI wrote nearly a quarter of corporate press releases in 2024
Cell Press
Since 2022, American companies, consumers, and even the United Nations have used large language models—artificial intelligence (AI) systems such as ChatGPT that are trained to create text that reads like human-generated writing. In a study publishing October 2 in the Cell Press journal Patterns, researchers reveal that AI is used in an average of 17% of analyzed corporate and governmental written content, from job posts to press releases, and this rate will likely continue to increase.
“This is the first comprehensive review of the use of AI-assisted writing across diverse sectors of society,” says corresponding author James Zou of Stanford University. “We were able to look at the adoption patterns across a variety of stakeholders and users, and all of them showed a very consistent increasing trend in the last 2 years.”
Large language models became widely available to the public in late 2022. Today, more than a billion people around the world use them regularly.
Zou and his team decided to use an AI detection program that they’d previously developed to investigate the adoption patterns of these AI tools across four different writing contexts: US consumer complaints, company press releases, UN press releases, and job postings. They collected text published between January 2022 and September 2024 from each of these domains and ran it through the program.
To start, the team analyzed more than 687,000 complaints submitted between 2022 and 2024 to the Consumer Financial Protection Bureau, a US government agency responsible for protecting consumers from banks and other financial companies. They found that about 18% of these complaints were likely written by AI.
For the corporate news releases, the researchers analyzed text published in three major news release platforms in the US: Newswire, PRWeb, and PRNewswire. They found that since the launch of ChatGPT, nearly a quarter of releases on these sites were AI generated. In particular, science and technology releases had the highest AI use rate by the end of 2023.
For job postings, the researchers found that posts from large companies on LinkedIn were less likely to be written by AI. However, after investigating vacancy postings from smaller firms, they found that large language models likely assisted in about 10% of the posts.
The team also looked into UN press releases written in English. They found a significant increase in AI-assisted writing, from 3% in early 2023 to more than 13% by late 2024.
Overall, the researchers found that the portion of content flagged as written mainly by AI increased sharply from 1.5% before the release of ChatGPT in November 2022 to more than 15% by August 2023. After that, growth slowed down, and the AI adoption rate was about 17% by August 2024.
“These estimates likely reflect a lower bound of the actual adoption rates,” Zou says. The detector they used cannot accurately differentiate texts that are heavily edited by humans, he notes.
Zou also says that the AI detection tool works best with a large collection of text. It would not be able to pinpoint whether a single article used AI.
“I do expect that in the future, the adoption rates will continue to increase but probably not as rapidly as those in the first year,” Zou says.
“Like all new technologies, it’s difficult to say if these AI models are simply ‘good’ or ‘bad.’ They still make mistakes, so if people completely outsource their job to these tools and don’t bother to check the accuracy, that could lead to errors in their writing.”
###
This work was supported by the National Science Foundation, the US National Institutes of Health, the Silicon Valley Foundation, and the Chan-Zuckerberg Initiative.
Patterns, Liang et al., “The widespread adoption of large language model-assisted writing across society” https://www.cell.com/patterns/fulltext/S2666-3899(25)00214-4
Patterns (@Patterns_CP), published by Cell Press, is a data science journal publishing original research focusing on solutions to the cross-disciplinary problems that all researchers face when dealing with data, as well as articles about datasets, software code, algorithms, infrastructures, etc., with permanent links to these research outputs. Visit: https://www.cell.com/patterns. To receive Cell Press media alerts, please contact press@cell.com.
Journal
Patterns
Method of Research
Data/statistical analysis
Subject of Research
Not applicable
Article Title
The Widespread Adoption of Large Language Model-Assisted Writing Across Society
Article Publication Date
2-Oct-2025
Use of ambient AI scribes to reduce administrative burden and professional burnout
JAMA Network Open
About The Study:
This multicenter quality improvement study found that use of an ambient artificial intelligence (AI) scribe platform was associated with a significant reduction in burnout, cognitive task load, and time spent documenting, as well as the perception that it could improve patient access to care and increase attention on patient concerns in an ambulatory environment. These findings suggest that AI may help reduce administrative burdens for clinicians and allow more time for meaningful work and professional well-being.
Corresponding Author: To contact the corresponding author, Kristine D. Olson, MD, MSc, email kristine.olson@yale.edu.
To access the embargoed study: Visit our For The Media website at this link https://media.jamanetwork.com/
(doi:10.1001/jamanetworkopen.2025.34976)
Editor’s Note: Please see the article for additional information, including other authors, author contributions and affiliations, conflict of interest and financial disclosures, and funding and support.
# # #
Embed this link to provide your readers free access to the full-text article
About JAMA Network Open: JAMA Network Open is an online-only open access general medical journal from the JAMA Network. On weekdays, the journal publishes peer-reviewed clinical research and commentary in more than 40 medical and health subject areas. Every article is free online from the day of publication.
Journal
JAMA Network Open
AI can strengthen pandemic preparedness
Artificial intelligence could be a valuable tool for detecting emerging diseases earlier, researchers from five European universities and research institutes argue in The Lancet Infectious Diseases.
image:
Bird's nest and face mask. Photo: Lene Hundborg Koss.
view moreCredit: Photo: Lene Hundborg Koss.
How to identify the next dangerous virus before it spreads among people is the central question in a new Comment in The Lancet Infectious Diseases. In it, researchers discuss how AI, combined with the One Health approach, can contribute to improved prediction and surveillance.
“Artificial intelligence cannot by itself prevent pandemics, but the technology can be a powerful supplement to the knowledge and methods we already use. The better we become at integrating data from humans, animals, and the environment, the better prepared we will be,” says Professor Frank Møller Aarestrup from the DTU National Food Institute in Denmark, one of the authors of the Comment in the renowned medical journal.
It was co-authored by Professor Marion Koopmans from the Erasmus Medical Centre in the Netherlands. She warns that once a disease starts spreading, it is very hard to bring under control.
“The interventions required are drastic – as we saw during COVID-19. That is why it is crucial to detect new pathogens before they gain a foothold,” says Marion Koopmans, noting that once established, new diseases can become persistent challenges, as COVID-19 has also shown.
The team of authors, which also includes experts from Eötvös Loránd University (ELTE) in Hungary, the University of Bologna in Italy, and the UK Animal and Plant Health Agency, speaks from their experience as collaborators over years, focusing on One Health approaches to emerging disease preparedness in the VEO consortium – a European research initiative developing data-driven tools to detect and track emerging infectious diseases.
Pandemics often originate in animals
The outbreaks of diseases such as SARS-CoV-2, avian influenza, and mpox demonstrate the difficulty of controlling new potential epidemics. Many pathogens originate in animals, but when and where they will spill over into humans is unpredictable. The authors of the Comment highlight how climate change, intensive animal production, and human encroachment into natural habitats increase the risk of so-called spillover events – situations in which pathogens cross from animals to humans and, in the worst case, develop into epidemics. Spillovers have been likened to sparks: most extinguish, but some ignite fires that spread uncontrollably. Being able to detect such spillovers as early as possible is a challenge that the team has been studying using big data approaches.
AI can reveal patterns in complex datasets
Artificial intelligence can help to analyse such datasets from diverse sources – such as climate, land use, animal production, transport, population movements, and socio-economics. When these datasets are combined, AI can reveal patterns that would otherwise be difficult to discern.
“AI can help us identify where in the world surveillance should be intensified geographically, but also in specific animal species, in wastewater, or in humans. In this way, we can prioritise efforts where the risks are greatest, so-called hotspots,” says Frank Møller Aarestrup.
Genetic signals as early warning
Once such hotspots are predicted, metagenomic sequencing can be added as a catch-all approach for detection of pathogens, both known and new ones. Metagenomic sequencing is the analysis of genetic material – in samples from wastewater, air, food, or the environment. It is increasingly used to provide insight into a vast diversity of known and unknown microorganisms. Many of the genetic fragments identified are not yet characterised.
“When we sequence a sample, we may find millions of genetic fragments. Most resemble something familiar and harmless, but we are left with thousands of unknowns. Here, AI can help detect patterns and point to what might be dangerous,” explains Frank Møller Aarestrup.
Once it is clear there is a potential pathogen, questions can arise about how dangerous it is. The potential for viruses from animals to infect humans, spread and cause disease in part is embedded in the genetic code. AI-based tools can be used to predict how mutations might alter viral properties.
“We see huge developments in this area. AI-based protein models can provide an indication of what a mutation does to the structure of viruses, and how that then can be translated to risk of spread, or risk of severe disease. While challenging now, we see great potential for the use of AI to speed up risk assessment,” says Marion Koopmans.
AI as a co-scientist – opportunities and limitations
The comment also describes early prototypes of so-called AI “co-scientists”, capable of conducting an entire research cycle – from hypothesis generation and literature review to data analysis and reporting.
“I envisage AI becoming a recognised competence at the table – on a par with different types of researchers. AI can deliver analyses or suggestions that we as scientists can evaluate. In that way, the technology becomes a supplement that can strengthen our decision-making processes,” says Frank Møller Aarestrup.
“That also implies that we need to learn what our future role is as teachers and supervisors. How do we make sure that the novel ways of working provide trustworthy output? Will we be able to recognise mistakes with advancements of AI models? We also need to go back to the classroom. Really exciting,” says Marion Koopmans.
The authors conclude that artificial intelligence offers intriguing possibilities for enhancing pandemic preparedness. Still, it must be seen as a complement – not a replacement – to the classical surveillance and research approaches already in use.
Read more
The comment “Artificial intelligence and One Health: potential for spillover prediction?” was published in The Lancet Infectious Diseases and authored by Marion Koopmans (Erasmus MC), Istvan Csabai (ELTE), Daniel Remondini (University of Bologna), Emma Snary (Animal and Plant Health Agency), and Frank Møller Aarestrup (DTU).
Journal
The Lancet Infectious Diseases
Article Title
Artificial intelligence and One Health: potential for spillover prediction?
COI Statement
The authors are involved in a research consortium on One Health preparedness through data science (Versatile emerging infectious disease observatory: forecasting, nowcasting and tracking in a changing world [VEO]). VEO has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 874735.
Scalable AI tracks motion from single molecules to wildebeests
U-M researchers build artificial intelligence that discovers unique molecule behaviors in fluorescence microscopy—and could soon follow particles, wild animals or even celestial bodies
University of Michigan researchers have developed a tool powered by artificial intelligence that can help them examine the behavior of a single molecule out of a sea of information in the blink of an eye—or at least overnight.
Understanding the behavior of single molecules is important: it can lead to knowledge of different cellular processes or track how diseases begin and progress. To track the behavior of single molecules, researchers tag the molecules with what's called a fluorophore. They excite these fluorophores with a laser, then use powerful microscopes to follow the behavior of the tagged molecules over time.
But identifying important behaviors of these tagged molecules requires sifting through the vast amounts of data this kind of microscopy often produces. This requires an incredible amount of time, attention and luck—and even then, researchers can miss important information.
To combat this, the U-M research team developed META-SiM. Unlike task-specific AI models which focus on a single problem, such as language translation, the researchers developed META-SiM as an AI foundation model. Foundation models are large-scale AI models trained on many different kinds of experiments and analyses and a massive amount of data. This allows the tool to conduct a wide variety of analyses and scan through entire datasets to identify interesting behaviors that need further study.
The study, supported by the National Institutes of Health, is published in Nature Methods. Jieming Li and Leyou Zhang, former U-M graduate student researchers, led the work.
While currently focused on the evolution of a signal strength over time, reflecting different states, down the road the researchers say META-SiM’s algorithm can move beyond molecules and track other phenomena such as single particle diffusion, animal migration patterns or even the movement of asteroids through our solar system.
"The idea is to grow from single molecules to any larger scale. In principle, data have similarities to one another, and this AI algorithm is able to find out what those similarities are—as well as any deviations—no matter what scale you're working at," said senior study author Nils Walter, co-director of the Center for RNA Biomedicine. "We could also track, say, the movement of wildebeests across Kenya and Tanzania, or even potentially celestial bodies moving across the universe."
The researchers developed META-SiM by training it on millions of simulated traces that imitate many types of behaviors that molecules display in the lab. But one real-world example of what META-SiM could track is a frequent cellular origin of human genetic diseases, Walter said.
Our body produces different types of proteins for different types of cells—skin, muscle, bone or eye and so on—and their function. One way it does this is by splicing pieces of genetic information from our DNA together in different ways. When fused together properly, this information, called exons, becomes a messenger RNA. This mRNA then expresses a protein tailored to a specific organ.
But 60% of human genetic diseases occur because of malfunctions that occur when this genetic information is spliced together. META-SiM could theoretically find sporadic instances where the mis-splicing occurs, and then suggest therapies to combat the mistake.
Co-author and U-M research scientist Alexander Johnson-Buck likens looking for the behavior of a single molecule to a complex game of Where's Waldo?, the children's book series in which the goal is to find one tiny person wearing a red hat, glasses and a red-and-white-striped sweatshirt among huge crowds of people, sometimes wearing similar clothes.
"Doing analysis on large data sets like our single molecule fluorescence microscopy data is like doing a Where's Waldo? puzzle where you're trying to find Waldo," Johnson-Buck said. "Except maybe instead of a single page, it's hidden on dozens of pages or more, and maybe you don't know what Waldo looks like, and there might be multiple Waldos."
While META-SiM still cannot zero in on Waldo, what it can do is show scientists areas where Waldo might be hiding.
"It accelerates analysis and finds the key things that you would normally have to sift through the data for half a year or so to find basically overnight," Walter said.
According to Johnson-Buck, "you will still need an expert to interpret that discovery and to put it into context, but it makes the discovery aspect potentially a lot faster."
Study: Foundation model for efficient biological discovery in single-molecule time traces. (DOI: 10.1038/s41592-025-02839-4, available upon request or when embargo lifts)
Journal
Nature Methods
DOI
Generative AI is more efficient than nature at designing proteins to edit the genome
After studying the diversity of mobile elements in the genome to feed and train generative AI tools, researchers have been able to design and lab validate new synthetic proteins that can edit the genome more efficiently than natural proteins.
image:
Dimitrije Ivančić (co-first author of the paper), Avencia Sanchez-Mejías (CEO of Integra Therapeutics),
Alejandro Agudelo (co-first author of the paper) and Marc Güell (ICREA researcher and CSO of Integra Therapeutics)
Credit: Integra Therapeutics
Researchers at Integra Therapeutics, in collaboration with the Pompeu Fabra University (UPF) Department of Medicine and Life Sciences (MELIS) and the Center for Genomic Regulation (CRG), have designed and experimentally validated new synthetic proteins that can edit the human genome more efficiently than proteins provided by nature. This work, a global pioneer published today in the journal Nature Biotechnology, will be of great use in improving the current gene editing tools used in biotechnology research and personalized medicine by developing cellular (CAR-T) and gene therapies, especially to treat cancer and rare diseases.
The ability to insert large DNA sequences into genomes in a safe, targeted manner has been a revolution in the research and development of advanced therapies in recent years. Among the most promising systems are transposases, such as PiggyBac, which copies and pastes DNA to introduce therapeutic genes into patient cells. However, their potential has been limited by the scarce diversity of known transposases and their lack of precision.
Exploring biodiversity
The researchers used computational bioprospecting to screen more than 31,000 eukaryotic genomes and discovered more than 13,000 new, previously unknown PiggyBac sequences. After performing experimental validation in cultured human cells, 10 active transposases were identified, demonstrating that there is a large functional diversity that has not yet been explored. Two of these new transposases showed activity comparable to versions already optimized for laboratory and patient use, and one of them even exhibited high activity in human primary T cells, a crucial cell type for cancer therapies.
Designing with generative artificial intelligence
In the second phase, researchers went beyond nature and used a protein large language model (pLLM), a form of generative artificial intelligence. They trained the model with the 13,000 PiggyBac sequences discovered to generate completely new sequences with enhanced activity. This approach not only optimized one of the existing transposases, but also demonstrated that AI-engineered variants are compatible with advanced gene editing technologies such as the FiCAT platform.
“Publishing this paper in Nature Biotechnology opens the way to revolutionizing the field of gene editing and advanced therapies and cements Integra Therapeutics’ position at the forefront of gene therapies and the use of innovative tools like AI for protein design in our development,” notes Dr. Avencia Sánchez-Mejías, CEO and co-founder of Integra Therapeutics.
“For the first time, we have used generative AI to create synthetic parts and expand nature. Like the cognitive power of ChatGPT can be used to write a poem, we have used the protein-based large language models to generate new elements that comply with the physical and chemical principles of genes,” explains Dr. Marc Güell, scientific director at Integra Therapeutics and ICREA researcher at MELIS-UPF where he heads up the Translational Synthetic Biology Lab.
“These AI models are trained with all known protein sequences on earth and learn the internal language or ‘grammar’ of proteins. Using this grammar, they are able to speak this language perfectly, generating completely new proteins that maintain structural and functional meaning,” says Dr. Noelia Ferruz, who leads the Artificial Intelligence for Protein Design Group at the CRG.
To accelerate and expand its FiCAT technology and pipeline of therapeutic products, Integra Therapeutics forges strategic partnerships with leading companies and research centers like UPF and CRG.
Journal
Nature Biotechnology
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
Discovery and protein language model-guided design of hyperactive transposases
Article Publication Date
2-Oct-2025
COI Statement
AA, JLV, JJW, Maria G MS,RD, MG, ASM, NF and DI are employed or have consulted for Integra Therapeutics. MG and ASM are shareholders of Integra therapeutics. DI, MG, ASM, AA and RD have filed patents related to this work.
