SCOPE uses a team of AI agents to rate news outlets for bias and credibility

What the researchers built

As arxiv.org reported in a paper posted on 18 October 2026, a team has built SCOPE, short for Source Credibility and Outlet Profiling Engine. You give it the web address of a news outlet. It returns a profile of that outlet's bias, its factual reliability and its overall credibility.

The notable part is that SCOPE is not one AI model answering a question. It is a multi-agent system, meaning several AI components, each with a narrow job, that pass work to one another. The paper lists five jobs: collecting articles, analyzing their content, searching for outside evidence, scoring, and writing the final report.

How the pipeline works

Think of it as a small newsroom desk. One agent pulls articles from the outlet. Another reads them and assesses the content. A third goes looking for external evidence about the outlet, such as what others have found about its record. A scoring step then combines those inputs, and a last agent writes it up.

The authors say the system follows an MBFC-style methodology. The abstract does not spell out the acronym, but it means the scoring approach is modeled on an existing style of outlet rating rather than invented from scratch. The aim is a profile built on a recognizable framework.

The paper also stresses a demonstration interface. It lets a user look at the intermediate evidence and the output of each agent, not just the final rating. That design choice is the core of the paper's pitch about transparency.

What the test showed

The team tested SCOPE on a benchmark of 200 media outlets. They compared it with four simpler setups:

• an approach that relies only on a large language model (the kind of AI behind chatbots);
• one that looks only at articles;
• one that relies only on search;
• a media-background-check baseline.

According to the abstract, SCOPE came out ahead on most metrics. It had the lowest MAE for both bias and factuality. MAE stands for mean absolute error, which is the average size of the gap between a predicted score and the reference score, so lower is better. It also detected failed fact-checks perfectly and had the lowest factual error rate.

Note the wording "most metrics." That means SCOPE did not win everything. The abstract does not say which metrics it lost on, and it gives no actual error figures, so the size of the improvement is not clear from this summary alone.

Why splitting the job helps

The authors' conclusion is that breaking the task into steps and grounding each step in evidence can make media profiling more transparent and more reliable. The comparison is the interesting part. A model asked to rate an outlet from memory has nothing to point to. A system that fetches articles and searches for evidence can show where its answer came from.

That matters because a credibility score is only useful if someone can check it. If a rating looks wrong, a visible evidence trail lets a person find which step went astray: bad articles, weak search results or a flawed scoring rule.

Limits and what comes next

This is a niche research paper, and it is worth reading with some caution. The abstract reports results on one benchmark of 200 outlets. It does not describe how the reference ratings were made or how SCOPE would handle outlets that are new, small or deliberately deceptive. Those questions may be answered in the full paper, but they are not in the abstract.

Rating news outlets is also contested ground. Any scoring system for bias carries the choices of whoever designed its method. Being able to inspect the evidence does not remove that, but it does let others argue with specific steps instead of a black-box verdict.

The code and a live demonstration are both published, so people outside the team can test the claims. How the system performs on outlets its authors did not choose will say more than any abstract.