AI-Powered Browsers Are Broadly Accurate News Summarizers That Reduce Political Bias and Negative Affect

While AI summarizers generally reduce anger and fear, they can actually increase fear in specific contexts. A new study from the University of Warsaw and the University of California, Davis, reveals that for threat-centered topics like immigration or international conflict, AI tools may introduce more fear than they remove. This finding challenges the assumption that AI always sanitizes news content.

The emergence of AI-integrated browsers marks a fundamental shift in how humans consume information. Historically, digital intermediaries—such as search engines or social media feeds—determined which information a user saw. Modern large language models (LLMs) have introduced a new layer of mediation. They do not just select content; they rewrite it. By generating summaries and synthetic overviews, these systems transform the actual substance of the news before it reaches the reader.

Until now, our understanding of this transformation has been fragmented. Most research has focused on the reliability of standalone chatbots in controlled settings. We lacked a systematic understanding of how these "black-box" systems behave when embedded directly into web browsers. This study addresses that gap by auditing the commercial summarization features millions of people use every day.

Beyond mere exposure to active transformation

The status quo in algorithmic mediation research treats AI as a discovery tool. However, the authors argue that browser-integrated AI acts as a new class of "editorial intermediary." Unlike a news aggregator that merely presents headlines, an AI summarizer condenses, reinterprets, and stylistically alters the source material.

This creates a significant tension in democratic discourse. If these tools are prone to hallucinations (the generation of factually incorrect or nonsensical information), they could inadvertently spread misinformation. Conversely, if they sanitize news too aggressively, they might strip away the essential nuance that journalists intend to convey. Previous audits have raised alarms about AI chatbots introducing factual errors. However, the broader question of how these tools reshape the emotional and political character of news has remained unanswered.

An automated audit of the browser experience

To investigate this, the researchers developed an automated pipeline to emulate real-world user interaction. Rather than feeding text into a chatbot via manual prompts, the team used virtual machines. They opened live news pages in Google Chrome (powered by Gemini), Microsoft Edge (powered by Copilot), and Perplexity Comet. The pipeline invoked the browser's native summarization function and captured the resulting text.

The study utilized a massive dataset of 13,777 articles from 15 U.S. news outlets. These outlets were selected to cover the political spectrum from left-leaning to right-leaning. The measurement process relied on a multi-stage verification architecture:

1. Factual Decomposition: Researchers used an LLM to break each summary into individual, decontextualized sentences. This is analogous to taking a complex legal contract and breaking it into discrete, independent clauses to verify each one against the original text.
2. Multi-Dimensional Annotation: Once decomposed, these sentences were evaluated for factual support. The summaries were also passed through transformer-based classifiers (machine learning models designed to recognize patterns in text) to measure ideological bias, negative affect (emotional tone), and journalistic quality.
3. Validation: To ensure the automated metrics were reliable, the authors validated the LLM-based labels against human annotators. They achieved high levels of agreement across all measured dimensions.

High accuracy with a smoothing effect

The results suggest that browser-based AI summarizers are more reliable than many critics assume. They act as a powerful "smoothing" filter on the news. The paper reports an average factual accuracy of 82.8% across all browsers [Figure 1A]. This means more than eight out of ten sentences are fully supported by the source. While this is broadly accurate, accuracy varies by tool. Microsoft Edge reported the highest accuracy at 88.3%. Google Chrome reported the lowest at 76.6% [Figure 2A].

Crucially, the study finds that these tools systematically attenuate—or reduce the intensity of—various news characteristics. The authors report that AI summarizers consistently reduce political bias and negative affect. For instance, 56.1% of articles expressing anger received more neutral summaries [Figure 4B]. Similarly, the tools improved journalistic writing quality by increasing clarity and stripping away clickbait tactics [Figure 1D].

However, this "improvement" comes with a trade-off in engagement. The paper finds that while the AI makes news easier to follow, it also flattens the writing style. Specifically, 78.0% of high-engagement articles received summaries that were actually less engaging than the original text [Figure 1D]. The AI essentially trades the "vividness" of human storytelling for a dense, detached, and formulaic information delivery.

Nuance lost in the condensation

While the findings are largely positive regarding accuracy, the authors highlight several critical caveats. First, the reduction in bias is not perfectly symmetrical. The paper finds that pro-Republican articles are significantly more likely to receive neutral summaries than anti-Republican articles. Additionally, Microsoft Edge appears to attenuate right-leaning stances more aggressively than other browsers .

Whether this is a deliberate safety tuning or a byproduct of how partisan language is structured remains an open question.

Second, the "smoothing" effect can lead to a loss of vital context. The authors note that while the AI reduces anger and fear, it can occasionally increase fear in certain contexts. This occurs particularly in threat-centered topics like immigration or international conflict. In these cases, the AI may strip away the "hedging" (cautious language used to avoid overstatement) found in the original reporting. When an AI condenses a nuanced discussion of a potential threat into a blunt statement, the perceived urgency can spike.

Finally, the study is geographically and linguistically bounded. The audit is limited to English-language news within the U.S. political context. Because the underlying models are proprietary and constantly evolving, the observed effects may change as models are revised.

The verdict: A reliable but sterile intermediary

Is AI-powered summarization ready for wide use? The evidence suggests a qualified yes. If the goal is to quickly grasp the factual gist of a story without being swayed by inflammatory headlines, these tools are remarkably effective. They are broadly accurate and clean up the "noise" of modern digital journalism, such as clickbait and excessive personal tone.

However, if the goal is to capture the full weight of a journalist's argument, these tools are not yet a substitute. They act as a high-pass filter that removes the "extremes." This includes both factual errors and emotional peaks. For developers and policymakers, the takeaway is clear. These tools are not just helping people find news; they are actively editing the reality of the news landscape.

Figures from the paper