Peace Speech: Do Peaceful Societies Communicate Differently?

Research shows that news in violent and stable societies uses different words.

Hate speech and its role in inciting violence have long been a source of public concern. In 2019, the United Nations launched a strategy against hate speech, with Secretary-General António Guterres pointing out that hate speech had often been seen as a precursor of atrocities and even genocide in Rwanda, Bosnia, and Cambodia. 

In late 2023, the European Commission called attention to growing islamophobia and antisemitism in the context of the ongoing war in Gaza, and proposed new measures to address it. 

Given the spread and threat of hate speech, some researchers have asked whether a “peace speech” with the opposite effects can be identified. It is this question that led Columbia University’s Advanced Consortium on Cooperation, Conflict, and Complexity (AC4) to investigate whether they could identify different speech patterns and word choices in societies with high and low levels of peace.

In 2008, Dr. Peter Coleman, AC4’s executive director, was invited to organize a day-long session on the study of peace involving UN staff, academics, and philanthropists. Coleman described the session as “an absolute failure, because no one could change the channel and not talk about atrocity prevention and violence.”

“Of course it’s relevant to peace, but it definitely misses half of the story,” he added. After co-writing a book on the underlying components of long-lasting peace, Coleman became interested in investigating what AC4 refers to as “sustainable peace.” He described communities with sustainable peace as “societies that are able to maintain peace internally, peace in their foreign affairs, peace systems – which are clusters of societies – and able to do that for 50, 100, 200 years.” It was out of this desire to take sustainable peace seriously that AC4 developed as a multidisciplinary team driven to carry out different projects researching peace.

Coleman and his research partner Dr. Larry Liebovitch were concerned that existing measures of peace were methodologically flawed and limited by their narrow focus on the absence of violence. The team was also interested in an alternative that could provide real-time data and analysis, as opposed to indexes relying on annual metrics.

In 2019, they decided to deploy data science in their effort to better understand peace. After a variety of approaches were tested, a sub-project delving into the relationship between news speech and peace proved particularly promising. Coleman and Liebovitch have since spent five years exploring how language can affect and be affected by peace and conflict.

“We know that international organizations, USIP, other groups, do track in some locations, hate speech,” Coleman said. The research team tried to turn this idea on its head to try to determine if an opposing “peace speech” could protect societies against conflict. 

While an initial attempt to develop a “peace lexicon” – a list of terms associated with peace –  did not bear results, Liebovitch then proposed that they use a machine learning model to identify linguistic differences between peaceful and conflict societies: “We really went from a top-down theory of what we thought we’d see, which didn’t pan out, to a more bottom-up exploration of the data itself.” Essentially, rather than imposing a specific list of pre-existing “peaceful” terms on the data, the researchers asked AI to comb through news articles from countries experiencing varying degrees of peace and conflict and then identify which words were more common in each.

Coleman and Liebovitch then tested their model by trying it on data beyond its original training set. “We were very surprised that worked,” Liebovitch said. The model was originally trained on extreme cases, with data from countries experiencing either recurring conflict or long-lasting peace, but “it accurately ranked those between those two extremes. That was very satisfying.”

They found that in countries with lower levels of peace, the news media was dominated by references to the government and politics, which was an expected result. Liebovitch said that they had spoken to journalists in those countries who explained that in authoritarian contexts, media tends to become a mouthpiece for the government.

However, the team was surprised to find that in peaceful countries, the most common terms were associated with common activities in daily life, such as sports, work, or family. “That was the tell in the data,” Liebovitch said. “When things are going really well, those sorts of daily activities dominate the news media.” Based on these results, the team concluded that they could measure the number of words dedicated to daily activities as opposed to references to politics to create a new quantitative peace index.

The researchers emphasized that the relationship between language and peacefulness is almost certainly a circular one. “This is a mutual dynamic,” Coleman said. Linguistic differences are an outcome of different societal circumstances, “but we also suspect that what a society chooses or is inclined to focus on […] is also a driver of the priorities of that society, of the concerns of that society. So it definitely is a two-way dynamic.”

Coleman and Liebovitch led a series of workshops with experts and journalists from the countries they analyzed, which helped them make sense of their quantitative data thanks to their understanding of the local industry and political and legal constraints.

The academics remain aware of their research’s challenges and limitations, with Liebovitch highlighting that the model had only used English language media, which could bias the findings. He also said that they were aware of the problems with the datasets they had used, with one in particular being dominated by financial news data which had to be controlled for. The team also had to adapt to the breakneck pace of developments in AI and data science. Additionally, such technology is expensive, even if AC4 received financial sponsorship from the Toyota Research Institute.

AC4 is now focused on gaining a better understanding of the complex social dynamics behind their results, with the researchers expressing interest in exploring whether different topics are covered in particularly hostile or emotional ways.

Coleman raised the possibility of using their results to produce an app to rate the relative peacefulness of readers’ media consumption. Coleman and him also hope that this would be helpful to journalists and editors making sure that their work does not feed nor incite violence. By allowing researchers to track how peaceful or violent language becomes in real time, the model developed by AC4 could provide an early warning about atrocities or conflict. Future research could also explore other forms of media beyond news as a source of data. For example, Coleman said, children’s stories or songs might also be examined to study the popular media that people are socialized by. 

However, both Coleman and Liebovitch remain cautious and took care to emphasize that despite the promising nature of the research, it is still in its early stages and such applications are still aspirational.

“This is a very narrow slice of all of the possibilities,” said Liebovitch, “but it still turned out to be a tasty slice.”