Gen AI has pretty much entered every place where human language is found, whether written or spoken. I was curious about one context: political speech in the American Congress. Was it really used? If so, how was it used?
To answer these questions, I first thought about using AI-detectors, but the number of false positives was too high. So, I've built a corpus of words that are overrepresented in AI writing. To select the appropriate, I’ve based myself on previous literature (more details on the methodology here). I then tried to track the rise of those markers throughout Congressional speeches over the last 10 years.
This cannot prove whether AI usage increased, only that the markers overused by AI increased.
To calculate the increase of the markers, I computed z-scores by taking the distance from the mean divided by the standard deviation. A z-score above 3.0 would indicates a deviation that is statistically significant. The mean and the standard deviation are constructed based on the baseline period (after 2017). I retrospectively applied this same z-score formula to the pre-baseline years (2014–2017) to obtain test years.
Here is the graph for the House of Representatives and the Senate.
This big graph combines the deviations of all the markers. This means that even if an individual word did not have a statistically significant increase, the aggregate of all markers is still statistically significant.
Since I've mostly seen the increase in the House of Representatives, I've decided to focus on it in the rest of the study. Here are the results for individual words in the House of Representatives.
The prevalence of the increase in the house of representative might be due to the greater usage of Extensions of Remarks and one-minute speeches in the House of Representatives. Those are the two places where AI seemed to be the most widely used. It might also come from the fact that Senators have been in power for longer and therefore already have established ways of working.
Interestingly, the usage of markers decreased in 2025 while still being statistically significant. It might be because politicians used less AI in 2025 than in 2024. What seems more probable is that the language of Large Language Models (LLMs) has started to change, making the markers of previous literature less relevant.
After finishing the quantitative part, I read the PDFs of the 20 days where the concentration of markers per thousand words was the highest. Conveniently, the two most frequent usages fit the two political communication styles established by Fenno.
In the 1970s, the congress scholar Richard Fenno observed that the communication style of congressmen could be divided into two categories: the home style and the legislative style.
The legislative style is basically what you imagine a politician doing in Washington: debating policy, voting for or against laws. The home style corresponds to how politicians act and speak to their local constituency , holding babies at agricultural fairs, going to the local baseball game, often with the goal to garner votes.
There are many more home-style speeches in the House of Representatives, because this kind of small-scale communication would have a negligible impact on the electoral results for a senator.
Inside the congress, the home style is mostly found in one-minute speeches. One-minute speeches, as their name suggests, are short speeches any politician can deliver before Congress for a minute. They are very often used to pay homage to a retiree, a local artist or anything local.
This is where the usage of AI seems to be the highest, as can be seen with the adjective "unwavering," which is used for a lot of different context within the congress. Here are a few examples from different politicians of both aisles:
“I rise to honor an individual whose service, character, and dedication have left an indelible mark on this institution and the people it serves. [...] I could not have known then that 4 years later, Veleter would become not only my chief of staff but also a trusted adviser, confidante, and unwavering partner in serving the people of Ohio’s 11th Congressional District.”
“It underscores not only Elivia’s remarkable talent and dedication, but also the recognition of her potential by one of the top programs in the Nation.”
“Rista’s life was a testament to what it means to live with purpose, passion, and an *** unwavering** commitment to the well-being of others.”*
“Their unwavering support is a *** testament** to the strength and generosity of the American spirit.”*
“Her many talents are a testament to her *** unwavering** dedication, tireless commitment, and incredible passion.”*
“ Didier William’s profound impact on the art world and his unwavering commitment to community engagement exemplify the power of art to inspire, connect, and transform.”
“Her life is a testament to resilience, dedication, and service to her family, church, and community.”
The interesting part is that all those different homages have been written by different political, from different political aisle, and each one of them talks about different person or group, from retiree, to firefighter, to local artist, to businessman. Yet the wording is always
the same to the detail.
The legislative-style can be found when politician used AI to defend a policy.
Here are a few examples :
In an increasingly interconnected world where digital technologies touch every aspect of our lives, safeguarding personal privacy has become a critical concern for all Americans, especially during a time when vast amounts of personal data are collected online. As we navigate through a landscape of evolving cyber threats, data breaches, and the development of artificial intelligence, the need for cutting edge, privacy enhancing technologies has never been more pressing. Recognizing the significance of these challenges and the threats we face online, this legislation directs the National Science Foundation to support competitive, fundamental research on privacy enhancing technologies. [...] *** This bill will not only facilitate crucial research efforts but also [not just X, but Y] *** contribute to the development of a skilled workforce and foster effective government coordination to ensure an impactful implementation of these technologies. [...] *** Through cutting-edge research and technologies, we will develop innovative solutions to not only shield sensitive data from malicious actors but [not just X, but Y] also establish robust standards for data collection and sharing practices, *** fostering a more transparent and secure online environment.
By expanding NOAA’s authority to acquire commercial weather data, we are not only improving the efficiency of weather data acquisition but also fostering innovation in the private sector [not just X, but Y]. These measures *** underscore** our commitment to protecting the safety and well-being of the people of New Jersey, ensuring that they have the information and resources needed to withstand and recover from weather-related disasters.*
Politician appear to use more often AI to write bipartisan policy with low stake. Those laws had already a high chance of passing the house and the speeches n were more filling speeches rather than determining one. Overall, those speeches seemed to have very small impact and were very consensual. The political speeches written by AI also tend to be less personal, overusing the “we”.
Those things might be due to the context in which AI is preferably used (low stake speeches) but it can also be due to what GAI allows, as model are trained to be consensual and to refuse violent rhetoric.
Those speeches individually would probably have felt mechanical before GAI. Alone, they might not have a huge impact. Combined, they could potentially link neutral and "boring" policy to GAI-written speeches in the eyes of the public. At the same time, the more human speech, the kind that uses a lot of "I," and the insults that AI usually can't produce because of Reinforcement Learning with Human Feedback (RLHF), will remain the domain of extreme.
It would be an overstatement to say that the recent rise of populism is due to moderate politicians using AI to write their speeches. However, it's an interesting tendency to study and to theorize about for the near future of politics.
Fore more detail, you can check the paper here.