Jerusalem, 10 November, 2026 (TPS-IL) — As millions of people increasingly turn to artificial intelligence (AI) to understand wars, political disputes, and breaking news, a new Israeli study raises questions about whose views these systems reflect when interpreting news.

Researchers at Ben-Gurion University of the Negev found that leading AI models differ considerably in how they interpret the emotional signals conveyed by news headlines. Some models more closely resembled the responses of human readers, but none consistently reflected the views of all groups.

The findings suggest that AI systems do not necessarily offer a universally shared interpretation of news, particularly when reporting on sensitive political and international conflicts.

The study examined how headline wording can influence readers’ emotional responses. Two headlines describing the same event, for example, might emphasize different aspects of the story, with one focusing on civilian suffering and another on military circumstances. Both could be factually accurate while evoking different levels of sympathy toward the people involved.

Researchers sought to determine whether AI models recognized these differences and whether their interpretations corresponded with those of human readers.

The research team, which included Dr. Haran Shani Narkiss, Dr. Michael Fire, and Dr. Oren Tsur, compared seven leading AI models with responses from 3,011 adults in the United Kingdom.

Participants and AI models were presented with news headlines concerning the Russia-Ukraine war, the war in Gaza, and the 2024 US presidential campaign. They were asked whether each headline evoked sympathy toward a specified side. A nationally representative YouGov survey generated more than 200,000 individual responses.

“LLMs are increasingly becoming intermediaries between people and information,” said Dr. Tsur, an assistant professor at the university’s Institute for Applied AI Research.

“The impact of a news headline goes beyond the facts it reports. Its wording can convey sympathy, implicitly assign responsibility, and subtly shape a reader’s understanding,” he said.

Measuring AI Bias

The researchers found that the extent to which AI models agreed with human responses varied according to participants’ age, gender, education, prior knowledge of geopolitical issues, and existing views of the conflicts.

In other words, a model that interpreted headlines similarly to one group of people did not necessarily reflect the views of another.

Dr. Michael Fire, a senior lecturer at the university, said the size of the survey enabled researchers to examine which groups’ judgments AI models most closely reflected, rather than relying solely on average results.

“With more than 200,000 responses from over 3,000 participants, we can look beyond whether AI agrees with people on average,” Fire said.

The findings carry broader implications as AI tools become increasingly common sources of news summaries and explanations. When an AI system interprets a headline or summarizes a report, it may overlook emotional cues that influence how readers understand an event.

The researchers said future AI development may need to account for demographic and cultural differences. However, they cautioned that deciding whose perspectives should guide AI systems raises difficult ethical questions. Without transparency, efforts to align models with particular audiences could introduce new biases.

They also warned that AI systems that routinely agree with users and reinforce their existing opinions could contribute to echo chambers, increased polarization, and eroding public trust in information sources.