The experiment, known as the Spinoza project, was launched in November 2023 by Reporters Without Borders (RSF) and the General Information Press Alliance, and brought together more than 120 partner media outlets from 12 publishing companies, including Actu.fr, EBRA, L’Équipe, La Nouvelle République, La Provence, Le Télégramme, Libération, L’Union, Nice-Matin, PMSO, Sogemedia, and Sud Ouest.
The initiative was born from a desire among journalists and publishers to familiarise themselves with generative AI. The project aimed to enrich journalism by creating an AI tool with reliable databases, focused on helping journalists research climate change. The current version includes scientific reports and legislative texts. This first experiment proved that AI tools that respect journalistic ethics are not only possible but essential.
The project identified four key pillars necessary for designing a trustworthy AI tool for journalists:
- assist journalists rather than replace them;
- produce content that serves the public interest;
- develop the tool based on concrete use cases;
- contain the generative algorithm
Journalists see AI tools as a way to save time on specific tasks, such as translating documents, generating interview transcripts, summarising information, rewording text, and producing article summaries. 45% of the French journalists surveyed by RSF in 2023 already use generative AI in their professional work, and 93% plan to use it in the future.
How the Spinoza tools work?
Spinoza integrated six databases that include scientific data, legislative data, data from public institutions, data from the French Agency for Ecological Transition (Ademe), journalistic content, and data from Agence France-Presse (AFP).
By incorporating press articles into its database, Spinoza filled the gaps left by other sources, such as scientific reports, which are often limited to specific regions or, conversely, describe phenomena on a very broad scale. The collaborative work carried out during the experiment demonstrated how valuable journalistic content is to AI-driven tools.
The Spinoza assistant is built on a search and synthesis system designed to address journalists’ concerns regarding environmental topics. The tool is used by following these steps:
- The user enters a question or keywords into the search bar.
- The tool reformulates the question according to the different databases to be explored. For example, if the user asks, “How can we limit ice melt?”, the tool may rephrase it as, “What regulatory measures have been implemented to limit ice melt?” to obtain relevant results from a database of legal texts. Alternatively, it may be reworded as, “How can ice melt be limited? Which articles discuss this?” for a more precise search in a database of press articles.
- An embedding algorithm explores four distinct databases, each specialised in a specific domain (e.g. scientific reports or press articles). This engine selects relevant excerpts from each database, forming four separate corpora. A relevance score shows the similarity between the terms used in the question and the words found in the database.
- Each corpus is processed by a specialised agent, using a language model fine-tuned for its domain through specific instructions (prompts). For example, the “science agent” synthesises data from IPCC reports, while the “press agent” extracts and processes journalistic articles. Each of these agents produces a summary based on its respective corpus.
- The prompts are designed to prevent the agents from generating content when relevant data is missing. If no information in the database answers the user’s query, the system will display a message indicating that no relevant information is available. The instructions also prioritise numerical data and include clickable links, directing users to the original source document.
- The main agent, Spinoza, then consolidates all the summaries generated by the specialised mini-agents. It aggregates the information, highlighting differences and discrepancies between sources.
The founding principle of the Spinoza project was the transparency of sources: the user has access not only to the final summary but also to the intermediate summaries generated by each mini-assistant, as well as the sources used (via clickable links), and some original documents. This transparency allows journalists to verify and deepen their understanding of the information at every stage of the process.
Furthermore, the tool is not a chatbot. It does not retain previous interactions — it functions as a research and synthesis tool rather than a conversational text generator. This design choice addresses journalists’ and publishers’ concerns regarding reliability, transparency, and editorial control, while still delivering a powerful tool that can navigate the complexity of dense scientific documents.
The project was closed at the end of 2025 – if you want to try the tool, click.
