Salesforce expands disclosures about AI environmental impact

Software firm Salesforce has added metrics about energy consumption and carbon emissions to the model cards it publishes for several of its widely used artificial intelligence models.

Model cards, sometimes called AI “nutritional labels,” are fact sheets that are published alongside machine learning models to provide developers with information about how the AI was trained, how it performs, potential applications and demographic factors.

Many enterprise software companies publish them, but Salesforce is unique in its decision to include carbon emissions and energy consumption, as of June.

“It’s still early, but we’ve received positive feedback so far on the usefulness of bringing more transparency and structure to AI sustainability, with many customers expressing hope that other model providers will follow a similar path,” said Sunya Norman, senior vice president of impact at Salesforce. 

Salesforce has published model cards since at least 2020. The first to add the electricity and carbon metrics are ones trained directly by Salesforce — including first name match and account match models. 

“As organizations scale AI, they’re asking broader questions about how to deploy it responsibly, including how to better understand its environmental impact alongside performance, cost and business value,” she said.

Long-time commitment

Salesforce has published AI environmental data for at least two years as part of its “Sustainable AI Policy,” which calls for tighter governance and transparency about AI’s social and environmental impacts. 

The strategy was championed by Boris Gamazaychikov, who left Salesforce this spring to co-found a consulting and research firm dedicated to managing AI’s environmental impacts.  

The information for the AI model cards was calculated using the AI Energy Score, an initiative started by Gamazaychikov and computer scientist Sasha Luccioni. 

The resource is published by open source software firm Hugging Face. It covers common tasks such as AI image generation or speech recognition and analyzes the hardware used, along with the location of the data center, run times and other metrics. 

While measures for AI training models are fairly straightforward, because they involve a discrete process and computing environment, it’s more challenging to estimate the impacts for inference — when the AI is applied to a specific task — because that impact varies depending on where it runs. The Salesforce calculations include that phase.

“More broadly, we see AI sustainability as an ecosystem challenge that will require shared methodologies, collaboration and continued innovation,” Norman said.

The Salesforce sustainability team collaborated with the company’s office of ethical and humane AI use to create the model card data. Salesforce plans to add this information to additional models as measurement techniques become more straightforward, she said.