Meta has come under fire from scientists accusing the tech giant of using flawed data to train an artificial intelligence tool aimed at identifying materials for capturing carbon dioxide from the atmosphere. Researchers claim that Meta’s project raised unrealistic expectations about the feasibility of large-scale CO₂ removal.
Last year, Meta announced it had published a “groundbreaking” dataset and trained free-to-use machine learning models designed to help researchers tackle the challenging task of finding materials that efficiently bind carbon dioxide. However, scientists from Heriot-Watt University and the Swiss Federal Institute of Technology in Lausanne (EPFL) found that none of the 135 materials Meta identified as strongly binding CO₂ actually possessed that property, and some materials did not exist at all.
“I wish they had computed a bit less and thought a bit more,” said Berend Smit, a chemical engineering professor at EPFL, calling some of Meta’s findings “nonsense.” He criticized the “Big Tech mentality” of acting quickly without sufficient reflection.
Meta responded by defending the validity of its dataset, stating it was based on “valid calculations useful for training machine learning models” and emphasized its commitment to open sourcing technology to foster collaboration and innovation.
Tech companies, including Meta and Microsoft, have invested heavily in reducing the costs of carbon removal techniques such as direct air capture (DAC) to offset their growing carbon footprints amid rapid development of generative AI tools. While DAC is not yet operational at scale, Meta has committed to purchasing credits from startups working on the technology.
The contested findings were published in a peer-reviewed journal of the American Chemical Society, with contributions from Meta researchers and Georgia Institute of Technology. A.J. Medford, associate professor at Georgia Tech and a co-author, said the research aimed not to conclusively identify new materials but to experiment with advanced screening techniques and uncover new challenges for the field.
Meta’s team performed approximately 40 million quantum mechanics calculations to create a database of metal-organic frameworks predicted to selectively bind CO₂ over other air components like water vapor. The project required vastly greater computing power than typical academic labs can access, and the resulting data trained an AI model reportedly much faster than existing chemistry simulations.
However, attempts to replicate Meta’s results revealed that the materials’ CO₂ binding capacities had been overestimated and that the open-source AI tools were not fit for their intended purpose. One major issue was reliance on an outdated academic chemical database containing erroneous descriptions of elements, later corrected.
Meta acknowledged that the materials identified were merely “promising and deserving of more thorough inspection” and noted that some calculations involved “implausible or highly unstable structures” which had been disclosed by the authors.
The broader carbon removal sector faces challenges beyond Meta’s dataset. US policy shifts under former President Trump reduced subsidies for clean energy projects, causing uncertainty in the field. For instance, Climeworks, a leader in DAC technology that does not rely on Meta’s materials, recently laid off over 100 staff but also announced surpassing $1 billion in equity funding, signaling investor confidence.
Wijnand Stoefs, policy lead on carbon removals at nonprofit Carbon Market Watch, said DAC has been “enormously overhyped” by tech companies including Meta. He highlighted the technology’s high cost and energy consumption compared to transitioning from fossil fuels to clean energy, warning of a “deep crisis” for carbon removal advocates.
Nevertheless, some scientists praised Meta’s open approach. “By publishing everything openly, Meta enabled the research community to dig in and build better tools,” said Susana Garcia, professor of chemical and process engineering at Heriot-Watt University.
