On 8 October 2026, ESA introduced Earth Virtual Expert, or EVE, an artificial intelligence assistant designed to make Earth observation knowledge easier to query. The promise may be of interest to a team without a dedicated library: ask a question in plain language and receive an initial search through articles, documentation, and studies. But it is important to be precise about the verb "observe". In its version announced on that date, EVE works primarily with textual sources. It does not yet automatically look at a satellite image of Douala to confirm a flood.
A real problem of access to information
Earth observation data is accompanied by mission reports, instrument specifications, publications, software, and product descriptions. Identifying the correct document and its level of validity can be time-consuming. A simple question, such as "What measurements distinguish surface temperature from air temperature?" may require comparing multiple sources. An assistant that flags relevant references can save time, especially for a reader who is new to the discipline. However, the time savings are only useful if the answer allows the reader to return to the original documents.
ESA explains that EVE was developed with Pi School, Mistral and Imperative Space. Its corpus includes sources from ESA, NASA and Copernicus, as well as manuscripts and research published by Wiley. Stating that an answer comes from "reliable sources" does not prove that every sentence produced is correct. A document may be outdated, a method relevant for Europe but unsuitable for an African study, or a figure repeated without its units. The reader must be able to verify the date, scope and meaning of each reference.
Augmented search, but no guarantee of truth
EVE uses retrieval-enhanced research: before composing a response, the system searches for passages that might support it. This can limit inventions due to simple generation from the knowledge stored by the model. The ESA also describes a "LLM-as-a-judge" control, where another automated system reviews the response and identifies what should be corrected or filtered. This is a safeguard, not an independent scientific review. Two language systems may share the same error or misinterpret a single source.
A reasonable test would be to ask a question for which the answer is known, and then to open each of the suggested references. Are the numbers identical to those in the source? Are the units and dates correct? Does the text separate observation, interpretation, and projection? Does it explicitly state "I don't know" when the documentation does not allow a conclusion? An appealing answer that fails any of these checks should not become a published fact.
The imagery will come later, according to the roadmap.
ESA indicates that the second phase of development will enable EVE to search for and analyze Earth observation data and perform multi-step tasks. Satellite image retrieval and presentation are scheduled for 2027. This timeline is a development intention, not a feature available on October 8, 2026. Until then, asking EVE to "see" coastal erosion from a specific scene exceeds the officially described capabilities.
The repositories for the model, code, selected datasets, and evaluations are publicly announced on GitHub and Hugging Face. This transparency can help researchers reproduce tests; it does not eliminate the biases of the corpus or the computational or language constraints. For French-language use, it would be necessary to examine not only the fluency of French but also the fidelity of references and the translation of technical terms.
A useful public trial before making it an editorial tool
RECA could prepare a set of questions, the answers to which are pre-checked in the agencies' documentation. Some would be about a well-defined capability, others about a limit, for example the difference between surface temperature and air temperature. Others would ask for a local result that is not provided by available sources. After each EVE response, two readers would open the links, check the units, and indicate whether the conclusion exceeds the documents. Errors would be published along with successes. This is the most honest way to evaluate a tool that is advertised as being accessible to everyone.
It would also be useful to test the same topic in French and English. A fluent translation may add a nuance that was not present or remove a validity condition. If the system cites only a general page while advancing a local figure, the author must search for the primary source. An assistant that helps find readings can be useful even if it is not reliable as an independent author. Conversely, a seemingly brilliant answer without a verifiable trace has no place in a scientific article.
A traceable response is better than a spectacular one
To judge EVE, more is required than a plausible statement. A useful answer should indicate the sources it consulted, the date of those sources, and the points where the documents do not agree. If it cites a study on Europe to answer a question about Cameroon, it should explain the transfer limit. If it transforms an estimate from a model into a direct measure, the user must be able to detect this by opening the reference. Speed of generation is an advantage only if it does not make such verification more difficult.
The questions in French pose a particular challenge. A technical term may have several translations; a system that chooses the most common word may produce a fluent but scientifically ambiguous answer. A translation might retain the names of variables and products in the original language, then explain them, rather than replacing them entirely. It may also note when the service finds no African sources. This absence is information about the tool's corpus, not proof that the phenomenon does not exist in the field.
EVE announces that its code, templates, and evaluation games are being released. Independent teams can use them to build tests that are better suited to their languages and topics. This does not allow for automatically copying an answer into a scientific journal: the review protocol must remain attached to each specific use. A later version of the tool may gain additional capabilities, but its answers from October 2026 should be judged according to the functions actually available at that time.
What this changes – or does not change – for Central Africa
A Cameroonian editor, class, or lab could use EVE as a starting point to find documents on optical imaging, rainfall, or data quality. The best exercise would be to note the question asked, the answer, the links provided, and the human corrections needed. This error log learns more than a few successful demonstrations. It is also important to test the service's accessibility with a limited connection and to see if the documents relevant to Central Africa are actually in its corpus.
A first experience report could be published with the exact questions and a list of controlled references. If an answer changes after an update of the service, keep both versions and explain why. This traceability would make EVE an open object of study, rather than a black box whose only demonstrations are the most successful ones.
EVE does not replace data acquisition, geospatial processing, field observations, or the expertise of a meteorological service. To publish a risk map, the entire chain remains in place: selection of dated data, explicit methodology, quality control, uncertainty, and editorial responsibility. AI can shed light on access to knowledge; it does not sign the result on behalf of the researcher.