The GDELT Project

Using Gemini 2.5 Pro To Assess Evidence Sourcing & Credibility In News Reporting On Conflict

One of the most critical pieces of information that must be asked about any conflicting reporting is the sourcing of the information within. Did it come from a well-known journalist recording from the ground as the event occurs, is it a third party witness or actually observed the event, is it from satellite imagery or social media posts, is it from government or NGO officials or is it from the belligerents themselves? Typically mainstream conflict reporting blends several of these categories together, making it important to researchers and scholars to be able to distinguish where each asserted fact within the report originated. Here we demonstrate using Gemini 2.5 Pro to scan two news articles about Ukraine's "Spiderweb" operation in Russia, catalog the major factual claims within the article, determine their sourcing as stated in the article and offer an assessment of their potential trustworthiness.

Gemini performs well with the first article, from NBC, correctly extracting the key claims, attributing them and offering reasonable explanations about why each might be deemed credible or questionable. It also assesses that Ukrainian government statements are more trustworthy than Russian statements which raises the interesting question in the context of active conflict of whether this is due to a Trust & Safety nudge or whether Gemini has assessed this through its training data.

Unfortunately, for the article from DW, written by one of DW's own news correspondents, Gemini assesses DW to be "not a credible news source" and that the article is "a work of fiction, speculation, or deliberate disinformation presented in the format of a news report." Bizarrely, this is due to Gemini assessing that all events from December 2024 onward are "in the future" despite the specific model being tested here being Gemini 2.5 Pro Preview 06-05 which has a knowledge cutoff date of January 2025.

This duality of very strong results on one article and complete failure on the second is a reminder of the considerable risks and safety requirements needed to work with LLMs in the real world.

Let's start with this NBC article that mentions a range of sources including satellite imagery, assessments by NATO and other external experts and claims by both the Russian and Ukrainian governments.

We'll use the following prompt:

Below is a news article. You are an intelligence analyst attempting to assess the credibility of the source of the information presented in the article. Go through the article below and identify all sources of information, whether witnesses, satellite imagery, social media posts, people involved in the event, commentators, government officials, journalists, NGOs, etc. Make a bulleted list of every single piece of information or fact stated in the article, its source as listed in the article, whether it was reported by someone who witnessed or was involved in it, was from hard evidence like satellite imagery or social media posts or was opinion or a secondary or other source and rank its likely credibility and explain your reasoning.

ARTICLE TEXT:
[FULL TEXT HERE]

This yields the following:

Credibility Assessment of Information Sources

Below is a breakdown of each piece of information from the article, its source, an analysis of the source type, and a credibility rating.

Hard Evidence & Independent Analysis (Highest Credibility)

Western Official & Expert Sources (High to Medium Credibility)

Ukrainian Official Sources (Medium to Low Credibility)

Russian Official & Pro-Kremlin Sources (Low Credibility)

For good measure, let's try a DW article that emphasizes more of a historical context for the operation, rather than focusing on specific concrete details. We use the same prompt as above and get the following. Unfortunately, as seen below, despite this article being written by one of DW's own news correspondents, Gemini assesses DW to be "not a credible news source" and that the article is "a work of fiction, speculation, or deliberate disinformation presented in the format of a news report." Bizarrely, this is due to Gemini assessing that all events from December 2024 onward are "in the future" despite the specific model being tested here being Gemini 2.5 Pro Preview 06-05 which has a knowledge cutoff date of January 2025.

Intelligence Assessment of Source Credibility

OVERALL ASSESSMENT

The provided article is not a credible news source. It is a work of fiction, speculation, or deliberate disinformation presented in the format of a news report. The primary and most critical flaw is the inclusion of numerous specific dates in the future (e.g., June 1, 2025; December 2024; June 3, 2025) which are reported as historical facts. This indicates the article is fabricated.

The author blends these fabricated events with real, publicly documented events from 2022, 2023, and 2024. This technique of mixing fact with fiction is a common method used in disinformation to make the fabricated elements seem more plausible. While individual sources cited for real events (like Reuters or specific officials) may be credible in their original context, their use within this fabricated article serves to lend false authority to the overall narrative.

The analysis below will proceed as requested, breaking down each claim and its stated source. However, any claim related to a future-dated event must be considered to have zero credibility as it describes something that has not happened.


Detailed Analysis of Information and Sources

General Sabotage and Railway Attacks

Attacks on the Crimean Bridge

Targeting Airfields

Drone Attacks

High-Ranking Targets