Geologists have raised concerns about possible Chinese censorship and bias in a chatbot being developed with the support of the International Union of Geological Sciences (IUGS), one of the world’s largest scientific organizations and a Unesco partner.
The GeoGPT chatbot is aimed at geoscientists and researchers, particularly in the global south, to help them develop their understanding of earth sciences using pieces of data and research spanning billions of years of the planet’s history.
It is an initiative of Deep-time Digital Earth (DDE), a largely Chinese-funded program established in 2019 to enhance international scientific cooperation and help countries achieve the UN’s Sustainable Development Goals.
Part of the underlying AI for GeoGPT is Qwen, a large language model built by the Chinese technology company Alibaba. One of those who tested a pre-release version of the chatbot, Prof Paul Cleverley, a geologist and computer scientist, said in an article recently published in the Geoscientist, the journal of the Geological Society, the UK’s professional association for geologists, that GeoGPT had “serious issues around a lack of transparency, government censorship and possible copyright infringement”.
In response to the article, DDE representatives Michael Stephenson, Hans Thybo, Chengshan Wang and Ishwaran Natarajan said that the chatbot also used Meta’s Llama, another major language model, and that they had not noticed any government censorship during testing, which according to them was “unlikely”. given that the system was “based entirely on geoscience information”.
The DDE academics said: “Problems with GeoGPT have been largely resolved, but the team will work to improve the system even more. It should be emphasized that GeoGPT is currently unreleased and not in the public domain.”
David Giles, a professional geoscientist, said it was “blatantly untrue” that a system based on geoscience data could be free of sensitive information.
Tests on Qwen, part of GeoGPT’s underlying AI, show that geoscience-related questions can produce answers that appear to be influenced by narratives set up by the Chinese Communist party.
For example, when asked how many people died in a mining operation in Ghana run by the Shaanxi Mining Company, Qwen says: “I am unable to provide current or specific information about events, including mining accidents, as my knowledge is based on data up to 2021 and I do not have real-time access to news updates.”
The same question posed to ChatGPT, the chatbot developed by the American company OpenAI, provides the answer: “The Shaanxi Mining Company in Ghana has experienced several fatal incidents, resulting in a total of 61 deaths since 2013. This including a significant explosion in January 2019 that claimed 16 lives alone.”
It is not clear what kind of answer GeoGPT, which is still in development, would give to this question.
Dr Natarajan Ishwaran, the head of international relations for DDE, said: “The team building GeoGPT has full independence. We can assure you that GeoGPT – currently in an exploratory phase and not yet open to the public – will not be affected by any government censorship.”
He added that users will be able to choose between using Alibaba’s Qwen or Meta’s Llama as the model for GeoGPT.
Geoscientific research and data include commercially and strategically valuable information on deposits of natural resources such as lithium, which are essential for the green transition.
Giles said there was a risk that a Chinese-developed platform could “filter” information to withhold content useful for “mineral exploration”.
He added: “China is very aggressively looking for minerals around the world. There is a strategic advantage and an economic advantage to looking for mineral reserves.”
to newsletter promotion
An article published in 2020 by Chen Jun, an academic at the Chinese Academy of Sciences, said DDE, the scientific program that created GeoGPT, will “help improve China’s detection and security capabilities in global resources and energy”.
Stephenson, Thybo, Wang and Natarajan, of DDE, said the 2020 article aimed “to encourage Chinese scientists to engage in international science programs” and was “purely the opinion of the author”, not of DDE or the Chinese Academy of Sciences.
Mohammad Hoque, a senior lecturer in hydrogeology and environmental geoscience at the University of Portsmouth, said “one danger” of using a Chinese language model for academic research is that “there will be some bias, because they are local laws must obey”.
GeoGPT’s terms of use state that it is prohibited to request the chatbot to generate content that “undermines national security” and “incites subversion of state power”. The Terms of Use also state that they are governed by the laws of China.
Hoque said GeoGPT has a greater obligation of transparency because it was developed under the auspices of an international research collaboration. “The most important thing is to know what data they use to refine and train [GeoGPT]. We have an expectation to know under IUGS.”
John Ludden, the president of the IUGS, said the GeoGPT database would be made public “only if the IUGS is satisfied that the appropriate governance is in place”.
Ishwaran said when GeoGPT is opened to the public, its training database will be made available “to those who want it”.
Geologists interviewed by the Guardian said the extent of DDE’s ties to China was not widely known among professionals. According to s planning document published in 2021, the multimillion-pound project was “almost 99%” funded by sources in China.
The program is part of the IUGS, an international NGO representing more than 1 million geoscientists in 121 countries, including the UK’s Geological Society. Its secretariat is based in Beijing and receives “tremendous” financial and logistical support from the Chinese government, according to the organization’s 2023 annual report.
Ludden said: “The best thing for science is to be open and share data. DDE does this for geological data if openly available [and] will lead to inward investment in any nation… [and] discoveries in research.”
