Scientists analysed 8 million US speeches – and uncovered a surprising trend

Scientists analysed 8 million US speeches – and uncovered a surprising trend

Machine learning has revealed how US political rhetoric shifted from evidence to emotion over the past 140 years

Getty

Published: April 22, 2025 at 6:30 pm

Scientists have uncovered a stark drop in the evidence-based language used in US political speeches since the 1970s, with significant implications for citizens in the country.

The results, published in the journal Nature Human Behaviour, are born out of an analysis of more than 8 million speeches made in the US Congress over the last 140 years.

“We wanted to know why some politicians who lie quite a lot are often thought of as honest,” Dr Segun Aroyehun, a postdoctoral researcher at the University of Konstanz, Germany, and first author of the study, told BBC Science Focus. “Often it's because they speak about what they believe in, and they sound authentic in that sense.”

Previous work by Aroyehun and his colleagues, which looked at social media posts from members of the US Congress, found that there had been a change in recent years from the use of evidence-based language to language rooted in beliefs or intuition. 

“There is this clear distinction between what we call the evidence-based notion of truth and the intuition-based notion of truth, where feelings are key,” said Aroyehun. 

To expand on this, the team wanted to peer into the past and find out whether this shift is a recent phenomenon and, if not, the effects that such changes to political language might have had.

Using a machine learning model, the team sifted through almost a century and a half of congressional speeches, building a scoring system based on keywords that pointed to either evidence- or intuition-based language.

Evidence-based keywords included things like “analyse”, “data”, “findings” or “investigation”, while intuition-based language included phrases such as “point of view”, “common sense”, “guess” and “believe”. 

The Republican elephant and Democrat Donkey mascots on a red and blue background.
Both Republicans and Democrats showed similar changes in political speech over the last 140 years. - Getty

Each speech was then scored according to the ratio of evidence- to intuition-based language.

What they found was startling. Although far from stable over the past 140 years, the ratio of evidence- to intuition-based language was roughly equal for much of that time. In fact, evidence-based language was actually on the rise through the 1940s, 50s and 60s.

Then it all changed: intuition-based language increasingly began to dominate from the 70s through to the present day. 

Aroyehun said that this pattern applied across both Democratic and Republican politicians, “except for the more recent period where we see a bit of uptick in evidence-based language from the Democrats.” 

He stressed that this ratio in no way reflects whether what is being said in a speech is true or false – only the way in which truths or falsehoods are being expressed.

However, that’s not to say the shift had no impact. The team compared these changes to measures of congressional polarisation and income inequality, both of which rose alongside the share of intuition-based rhetoric.

Meanwhile, congressional productivity – measured through the quantity and quality of enacted laws over time – fell over the same period. 

“The overall message here is to say that rhetoric does matter,” said Aroyehun. “And the nature of how that rhetoric with respect to the pursuit of truth is reflected in the language of US Congress does matter for these measures of societal and political wellbeing.” 

The team now plans to apply a similar methodology to speeches from Germany and Italy to see how well the trends track across other democracies.

About our expert

Segun Aroyehun is a postdoctoral researcher in Prof. Dr. Garcia's Social Data Science Group at the University of Konstanz. He obtained a PhD from the Centro de Investigación en Computación, IPN.  Broadly, his focus is on developing robust approaches for preventing, detecting, and countering occurrences of objectionable content on social media. His research has been published in Nature Human Behaviour and Nature Communications.

Read more: