When the News Doesn’t See You

THE boy was twelve years old, holding a toy gun. But that’s not what some news outlets called him when Cleveland police shot Tamir Rice in 2014. Instead, they described him as a “Black male with gun.” Other coverage mentioned his parents’ criminal records, details that seemed to have nothing to do with a child playing in a park.

Rob Voigt still remembers reading those articles. The University of California, Davis sociologist wondered whether this was an isolated case or something more systematic. The disparities felt more generalised than what had been documented before, he thought. So Voigt and his colleagues decided to find out.

What they discovered paints a troubling picture of how America talks about violence. Over nearly a decade, the team analysed 36,000 news articles about shootings across the United States, linking each story to specific incidents and the neighbourhoods where they occurred. The patterns they found were stark, and they held even when controlling for everything from the type of shooting to local crime rates.

Mass shootings in white-majority neighbourhoods received roughly twice the news coverage of identical incidents in communities where most residents were people of colour. A school shooting in a predominantly white area would typically get coverage from around fifteen articles. The same type of tragedy in a majority-minority neighbourhood? About seven.

Police shootings showed the opposite pattern. When officers fired their weapons in neighbourhoods of colour, the media attention was disproportionately high compared to officer-involved incidents in white areas. It’s as if the news operates with two different standards for when violence matters.

But the disparities run deeper than just how many articles get written. The team, which included researchers from Northwestern University and the University of Washington, developed computational methods to analyse not just coverage volume but the actual language used in these stories. They trained large language models to predict whether an article described an incident in a majority-white or majority-minority neighbourhood based solely on the writing. The model’s accuracy? Nearly 76 percent, far above what chance would predict.

That number tells you something has gone fundamentally wrong. “Media coverage of such incidents can perpetuate harmful biases,” says Voigt, “extending the impact beyond the immediate trauma of those involved.”

The researchers created what they call an “ontology” of features to understand exactly how the coverage differs. They looked at everything from which participants get named to how those people are described, from the formality of the writing to which authority figures get quoted. Each feature was then tested across thousands of articles, controlling for incident characteristics and neighbourhood demographics.

In white neighbourhoods, articles focus heavily on the shooter, often describing them in terms of their social roles beyond the incident. A typical passage might note that the shooter was a student, someone’s family member, a former Marine who suffered from PTSD. The language emphasises complexity and personhood. One article the researchers analysed described a shooter this way: “He is a former Marine who suffered from a traumatic brain injury and post-traumatic stress disorder, his stepfather said.”

Coverage of shootings in communities of colour rarely includes such humanising detail. Instead, these articles more frequently frame incidents in terms of race and crime, use more formal language reminiscent of police reports, and focus on mortality and the broader narrative of gun violence in America.

Victims receive particularly disparate treatment. When victims’ names appear in articles about majority-white neighbourhoods, they tend to be mentioned repeatedly throughout the piece, suggesting deeper engagement with their stories. In coverage of majority-minority areas, victims’ names might appear once but without the same sustained attention. The framing differs too, with victims in communities of colour more likely to be portrayed as criminals even when they were the ones who were shot.

The research team found that 62 percent of racial mentions in gun violence articles refer exclusively to people of colour, whilst the same is true for only 13 percent of mentions of white people. This suggests what researchers call a “white as default” perspective, where race becomes noteworthy primarily when someone isn’t white.

Articles about incidents in white neighbourhoods also quote authority figures more frequently. Local politicians and police officials show up to comment, lending their voices to the narrative. Whether this reflects journalists seeking out these sources or authorities choosing to engage, the result is the same: media representations depict greater institutional involvement when violence occurs in predominantly white areas.

The patterns hold across different population densities, though with some interesting variations. In low-density areas, the disparities in quoted authority figures are most pronounced. As neighbourhoods become more urban, some gaps narrow whilst others persist. Racialization, for instance, increases as density increases for white-majority neighbourhoods, suggesting heightened attention to race in relatively urban, white communities.

Ruth Bagley, the Northwestern University linguist who led the data analysis, emphasises that these findings represent systematic differences at scale. They’re not about individual journalists making biased decisions, but about larger patterns that accumulate across the media landscape. When you read enough of these articles, certain features start appearing together in ways that correlate with neighbourhood demographics, even when the incidents themselves are similar.

Previous research had documented some of these biases in limited contexts, particular cities or specific types of shootings. Mass shootings by white perpetrators, for instance, have been shown to receive different coverage than those by Black or Muslim shooters, who are more often labelled as “thugs” or “terrorists.” But this study is thought to be the first to document such disparities systematically across the entire country.

The research team couldn’t make causal claims about why these patterns exist. News coverage emerges from complex interactions between editorial decisions, source availability, audience expectations and institutional practices. But the descriptive findings alone reveal what they call “disparate impact”, echoing patterns found in other domains where racial bias has been documented at scale.

Some victims of gun violence have spoken about feeling dehumanised by news coverage of their experiences. They report being portrayed as criminals or somehow responsible for what happened to them, or feeling that coverage perpetuated fears rather than fostering understanding. The computational analysis suggests these harms accumulate more heavily in communities of colour.

The magnitude of these effects might seem small when you look at any single linguistic feature in isolation. An odds ratio of 1.5 for whether an article mentions someone’s social role doesn’t sound dramatic. But the researchers argue that these effects compound across the constellation of features they identified. You’re not just seeing one or two differences, you’re seeing dozens of systematic patterns that all point in similar directions.

To test this, they assembled a balanced dataset of articles and removed all names and explicit race mentions. Even with that information stripped away, their models could still predict neighbourhood racial composition with high accuracy just from the linguistic patterns. That suggests the differences run through the fabric of how these stories get told.

What gets covered, and how it’s covered, shapes public understanding. Previous research has shown that people dramatically overestimate the role of mass shootings in gun deaths (which they believe cause 25 percent of gun deaths when the actual figure is 3 percent) whilst underestimating suicides (which account for 60 percent). Media attention distorts perception, and when that attention is distributed unequally by race, it compounds existing disparities.

The findings arrive at a moment when newsrooms are increasingly conscious of bias in their coverage. But awareness doesn’t automatically translate into change, particularly when biases are embedded in language patterns rather than explicit editorial decisions. Writing style guides that address subtle linguistic patterns proves far more challenging than addressing overt bias.

Some of the solutions might involve the same computational tools used to identify the problems. If language models can detect these patterns, they might eventually help editors spot and correct them. But that raises its own questions about how much we want to rely on algorithms to police our storytelling.

For now, the research offers something more fundamental: evidence that the biases people have long suspected and occasionally documented in case studies actually exist at scale. Gun violence disproportionately affects communities of colour, and the new work suggests media coverage may compound these inequities through systematic differences in how incidents are represented.

Voigt wants to see more research on what causes these patterns and how they affect public perception. Do different types of coverage influence support for policy interventions? Can newsroom training change these patterns? Are there models from other domains that might help?

The questions multiply quickly, as they tend to when you shine a light on something that’s been hiding in plain sight. For decades, we’ve been reading these articles without necessarily noticing their patterns. Now the patterns are visible, quantified across tens of thousands of stories. What we do with that knowledge remains to be written.