HORIZON-CL5-2026-07-D1-04 Fighting disinformation and effectively communicating on climate change - Collaboration Offer

Dogus Teknoloji aims to develop a digital tool as part of this call. The tool will collect high-engagement content from the internet, assess them first using large language models and classify misinformation; then send them to human experts for verification. Verified and labelled information will be stored in a platform that offers role-based Access to different stakeholders.

Details of such a platform can be found below:

·        First, it will collect high-engagement content (social media, news articles, etc.) via scalable, multi-source pipelines. At this stage, keyword filtering or semantic filtering (meaning to use pre-trained language models suited for this task) can be used. The optimal strategy will be chosen based on computational cost and performance, while ensuring compliance with data privacy regulations (e.g., GDPR) and content licensing requirements.

·        Initial screening can be done by a smaller, cost-efficient LLM for confidence scoring and basic classification. Complex or low-confidence cases are transferred to a powerful LLM. A RAG-based architecture can be used to ground LLM detection findings by referencing verified sources. Confidence scores and explainable AI outputs will be displayed to improve transparency and trust.

·        The system forwards candidates to human experts for verification (Human-in-the-Loop). To avoid slow evaluation by experts, they only review content falling below a specific AI confidence threshold or those flagged as high-impact/viral. Experts will interact through a dedicated interface with prioritization queues, batch validation tools, and AI-generated reasoning to streamline decision-making.

·        Expert-validated content is tagged (by topic, source, and narrative) and compiled into a secure platform. The platform will support role-based access and views (for public authorities, media, and civil society partners), including customizable dashboards, advanced filtering, and visual analytics (such as trend graphs and narrative evolution heatmaps). Tagged data will be continuously collected as long-term metrics to track narrative evolution and source prevalence. Expert feedback will also feed back into the AI models to enable continuous learning and improve classification accuracy over time.

An example flow showing how this pipeline might be used:

·        A popular post regarding climate/climate change/environment gets circulated

·        The pipeline captures it; flags it as “containing potential disinformation”

·        Lightweight LLM labels “likely misleading” at an intermediate Score, it’s transferred to powerful LLM

·        Powerful LLM with RAG cites official statements and trusted reports; confidence score rises and it’s flagged as “high‑impact.” It’s transferred to experts

·        Expert reviews AI’s rationale, checks sources, and confirms misinformation

·        Item is tagged, published to stakeholder views; authorities receive an alert, media see a ready‑to‑use debunk summary, civil society gets an explainer.

·        The entry is logged for future use; feedback improves future model performance.

This tool aims to satisfy the following outcomes of the call:

• Advanced knowledge and understanding of disinformation dynamics

• Tools and products for detection and mitigation at scale.

• Tailored and effective communication techniques

• Increased acceptance of climate action and strengthened trust