Last year Meta identified 135 materials that could potentially be used to draw down carbon dioxide, work it described as “groundbreaking.” But when scientists tried to reproduce the results, they found that none of the materials could perform as promised and that some did not even exist.

For the research, Meta’s team performed 40 million calculations to render a list of materials that could attract carbon dioxide without drawing in other compounds, like water vapor, and it used its findings to train free-to-use A.I. tools. Meta told the Financial Times that its research sought only to identify materials that were “promising and deserving of more thorough inspection” and that it had made clear that some materials were “implausible or highly unstable.”

But when European researchers tried to replicate the research, they concluded that Meta had oversold its findings. Berend Smit, of the Swiss Federal Institute of Technology in Lausanne, told the Financial Times that some of its results were “nonsense.” He and his colleagues said that Meta had based its work on faulty data and that its A.I. tools were not fit for purpose.

Elsewhere, experts have recently posited that A.I. could prove instrumental in slashing emissions. A new paper in Nature, led by the influential British economist Nicholas Stern, finds that A.I. could boost the use of renewable power on the grid; aid the development of lab-grown meat; and help design cheaper electric cars. Advances in these areas, it concludes, could cut emissions enough to offset the impact of new, energy-hungry data centers.