GeoGPT marks a breakthrough for geoscience. But do concerns over transparency and Chinese restrictions outweigh the application’s potential?

AI tools in geoscience are rapidly emerging, assisting with workloads from seismic interpretation to mineralogical analysis. This interest was on display at the recent Digital Geoscience conference in October.

International Geoethics Day in October had the theme “valuing geosciences for the public good”. The aim was to focus on geoscience to tackle societal challenges and promote community wellbeing.

AI could have a part to play, by delivering scientific knowledge more robustly and efficiently.

What is GeoGPT?

Deep-time Digital Earth (DDE) is developing the first geoscience chatbot, GeoGPT. This tool, like ChatGPT, allows users to enter prompts and receive AI-provided answers.

This may involve questions about petroleum deposits, geological data, horizons, maps, or abstracts relevant to the deposit. It can filter through geoscience documents to make charts.

Yet, there is looming scepticism from many geoscience professionals. A viewpoint, published in the Geoscientist, outlined concerns around ethics, censorship and security.

Ethical and open source

GeoGPT is not technically in the public domain. DDE has described GeoGPT as open source. However, it has not released the training data or models in open repositories, co-authors said.

This is a point of contention and there have been requests for DDE to release GeoGPT training data.

Dr Natarajan Ishwaran, head of international relations at DDE, suggested that when made public, GeoGPT will release its training database. International Union of Geological Sciences (IUGS) president John Ludden said the database will be public once appropriate governance is ensured.

The industry is sceptical about how much it will reveal.

Open access policies usually come with sound copyright agreements. There is mistrust of GPTs scraping data without acknowledgment. GeoGPT’s vague T&Cs have done little to assuage concerns over copyright.

The China question

AI tools in China face local legal restrictions. Chinese laws govern GeoGPT. They prohibit generating content that “incites subversion of state power”, which includes jeopardising state security, interests or image, as The Guardian reported.

Tests on Qwen (an underlying part of GeoGPT’s AI) reportedly resulted in GeoGPT answers matching narratives set by the Chinese Communist Party.

AI law professor Jon Truby raised concerns in The Geoscientist. He warned that a scientific chatbot that prohibited certain questions “because they may tarnish the image of a government, poses a serious risk to academic freedom of expression.”

User data

GeoGPT harvests data from international users. But there are no clear explanations on how it handles this information.

Geoscientists and institutions can upload their own maps and documentation to the tool. Some may be sensitive, providing company and state-specific information.

Many nations would be interested in such information, in order to seize an advantage and maintain energy security.

Back and forth

DDE responded to Cleverley’s letter of concerns published in July. The group found no suggestion of state censorship. It said such intervention seemed “unlikely in a system based entirely in geoscience information”.

DDE said it had solved many GeoGPT problems, but the team was also working on communication and governance challenges.

The debate continues. In October, DDE professionals gave their perspective.

One point raised was the choices for the underlying model, between Qwen, LlaMa and Mistral. This allows the user to compare answers from Chinese, American and French AIs.

Users may still fear that the AI is selective in what it shows, particularly after the news of China’s censorship of Bing. While LlaMa and Mistral are made in other countries, uncertainties remain whether those same systems used in China would appear the same, since they must obey local law.

Transparent

The creators of GeoGPT want to continue the conversation and have invited more users to test it.

Transparency is crucial, as IUGS has said. The group, in July, issued a “strong recommendation that the geoscience corpus used to further train GeoGPT from its base model is transparent and be made available to the community”.

There should be transparency on its funding sources and potential government control of GeoGPT. The group planned to reconvene and provide a progress report for GeoGPT developments.

The debate has rolled on into November. Geoscientist published a letter this month that raised concerns on the publication of the “Advancing transparent and ethical AI” article. It repeated arguments, IUGS’ Ludden said, and transgresses the ongoing review process.

The Geoscientist team rejected the charges, saying it was an ongoing discussion and there was a need for constructive debate in “the timeliest manner possible”.

Power at what cost?

AI’s potential for efficiency and optimisation gains could solve many big data challenges. The GeoGPT debate highlights the complex issues of generative AI in science.

Science mentioned potential links of GeoGPT to unlicensed geoscience literature. Phoebe McMellon, CEO of GeoscienceWorld, told E-FWD that her concern is not for the future of digital geoscience, but science itself.

“Not all AI applications are created equal. I think that we should be cautious with using applications that are not built upon reputable and trusted content.”

This is especially for those systems “that do not uphold the fundamentals of research integrity, that violate legal and ethical standards, and whose answers and conclusions can’t be explained or reproduced”.

She called for AI applications to be “developed with transparency, attribution, accountability, privacy protections and safeguards, and with quality assurance and quality controls”.

This premise also appeared in Cleverley’s June letter in The Geoscientist. The scientist wrote that for transparency and open science, “it is critical that the training data underlying AI models are released to facilitate reproducibility”.

*Mandatory Credit: Photo by Global Warming Images/Shutterstock (1994504a)*

Alleged advantages

Others seem less concerned about the problems and more on the opportunities of AI in geoscience. Zhejiang Lab defended progress to Science. It said the “AI-driven transition for geoscientists is just getting started”.

Many geoscientists in China have written on the advantages of the tool. In a study funded by Chinese institutions, Zhang et al (2024) described its GIS decision-making as “robust with its outstanding language comprehension and reasoning capability.”

Hangzhou’s Zhejiang Lab had a sponsored feature in Science. It called GeoGPT a “toolset to shatter archaic practices and propel geology into a new era”. This AI application will synthesise worldwide information and provide new geoscientific insights.

Conversations and questions

If GeoGPT is to become a “system built for geologists by geologists” in earth sciences and energy transition, it must address those big questions. How will it ensure transparency? Are Chinese developed models constrained? How will it navigate copyright and intellectual property?

What goes for GeoGPT applies to other geoscience AI tools.

Implementing AI across government boundaries might pose challenges, but transparency and freedom of information are crucial to scientific research. The creators of GeoGPT are motivated to alleviate concerns. However, questions linger.

To support open science claims and avoid censorship and data handling fears, we need trust and transparency. Discussing early criticisms in the scientific community is a strong starting point to shape these tools’ future.

GeoGPT is an early mover, but it will not be the last. The time to thrash out these points is now, when there is still scope for change.

Related Content

very long title post. very long title post. very long title post. very long title post. very long title post. very long title post.

“Just get moving” on Tees Valley’s path to energy leadership

No supply chain investment? No CfD!

Wham, bam, thank you CBAM