Geoscience has an extensive, rich and diverse history of open-source software and platforms, including QGIS for spatial analysis, GPlates from the EarthByte group for plate reconstruction, GemPy for geological modelling, OneGeology for geological maps, and the OSDU Data Platform for energy data. But where do AI models, such as large language models (LLMs), and their applications fit into these definitions? And should proprietary subsurface data be uploaded to them?
Meta describes its Llama models as open-source, although their licences impose restrictions that do not meet the Open Source Initiative's (OSI) definition of open-source AI. Their training data are not available as a complete corpus for independent reproduction of training. In the geosciences, the Deep-time Digital Earth (DDE) initiative and Zhejiang Lab's GeoGPT models have also been described as ‘open-source’, although the relevant conversational model licences do not meet the OSI definition. These examples illustrate why public availability of model weights (‘open-weight’) should not be confused with ‘open-source’ licensing.
These examples illustrate why public availability of model weights (‘open-weight’) should not be confused with ‘open-source’ licensing.
The Model Openness Framework (MOF), endorsed by the UN University, has a three-tier classification. Its top tier, Class I ‘open science’, requires end-to-end reproducibility, including the actual training data used; AllenAI’s OLMo family is an example. In Class II ‘open tooling’, the European Space Agency (ESA) Earth Virtual Expert (EVE) would be an example. The lowest tier, Class III ‘open model’, requires the core model components and supporting documentation to be publicly released under appropriate open licences; French-origin Mistral AI’s Mixtral 8x7B, released under the Apache 2.0 licence, is a candidate for this class. Conversational models from Meta and DDE/GeoGPT would not qualify for any MOF class because their licences are not OSI-approved. An OSI-approved licence is necessary but not sufficient: models must also meet the MOF’s other requirements for a given class.
Moving to online platforms, LLMs are used to power applications (including agents) which come with their own source code, prompting and knowledge engineering. Commercially hosted US-based services such as Anthropic's Claude and OpenAI's ChatGPT provide access to proprietary AI systems. Although some access is available for free, this does not make the underlying models or services open-source. The chatbot GeologyOracle, which uses the ChatGPT API, illustrates the distinction between an application built on a hosted model and an independently deployable, open-source application. The same principles apply to DDE/GeoGPT, the Chinese state-funded initiative whose online application is self-described as ‘free’ and ‘open-source’. In practice, however, its model licences are not OSI-approved, and its full application-layer source code has not been released, so the platform cannot be independently deployed and is not open-source.
Claims of openness should be evaluated against the materials actually released. Where licences restrict use, training data or processing pipelines are unavailable in full, or the LLM-based application cannot be independently deployed, these limitations should be stated explicitly. Such distinctions matter regardless of whether the system is developed by a commercial company, a public research institution or a government-funded initiative.
“Claims of openness should be evaluated against the materials actually released.”
This matters for data sovereignty and commercial confidentiality when uploading proprietary data on natural resources such as oil & gas and critical minerals, and where actors wish to keep their online query transactions private. For sensitive subsurface data, organisations can deploy open-weight models within their own infrastructure, potentially retaining greater control over data and queries. That control, however, depends on the security and configuration of the entire deployment, not just the model. The LLM is only part of the picture: much of a geoscience LLM-based application's value comes from the code, prompts, retrieval mechanisms and knowledge engineering that surround it. Where these are not openly available, self-hosting a model alone will not reproduce the hosted application.
“Incorrect or ambiguous claims risk misleading colleagues and fracturing trust in the international geoscience community.”
Beyond sovereignty, there is a question of trust. AI undoubtedly brings benefits, whether it is open-source or not, and a degree of hype is to be expected. But geoscientists and providers have an ethical obligation to communicate about AI accurately, with full disclosure and clarity. Before describing a model or platform simply as ‘open’, we should ask a few simple questions: Is its licence OSI-approved? Are the model weights, training code and data, and application-layer software code available? Can the application be independently reproduced and deployed in an environment of our choosing, under the control of an organisation we choose? Incorrect or ambiguous claims risk misleading colleagues and fracturing trust in the international geoscience community.
Free to use, is not the same as open.
Paul H Cleverley is Professor at Robert Gordon University.
Views expressed in this column are solely those of the author.
