Salesforce has called for government action on a number of initiatives at the intersection of sustainability and the development of artificial intelligence.
The technology behemoth published its sustainable AI policy priorities Monday, as concerns mount over the resources needed to develop and run massive AI systems.
Researchers have already pinpointed the mass amounts of energy and carbon needed to train generative AI systems, like OpenAI’s GPT-3, which it uses to run systems like ChatGPT and Dall-E.
Academics at Cornell University published research that shows that the training of GPT-3 alone—without any public-facing use—resulted in 552 metric tons of CO2 emissions. For frame of reference, the average household in the U.S. creates about 7.5 metric tons of CO2 in one year. The CO2 emitted by training GPT-3 for 14 days is equivalent to the average U.S. home’s emissions for 73 years.
But data has been a bit more limited on the carbon emissions that come as a consequence of using the systems once they hit the market. Research by Hugging Face, an AI developer, and Carnegie Mellon University shows that text generation, image captioning and image generation are the most energy- and carbon-intensive tasks where generative AI system usage is concerned.
Related Stories
But image generation, because of its iterative nature, is the most carbon-thirsty generative AI process of the three, the researchers noted in their paper.
“The most carbon-intensive image generation model generates 1,594 grams of CO2 for 1,000 inferences, which is roughly the equivalent to 4.1 miles driven by an average gasoline-powered passenger vehicle, whereas the least carbon-intensive text generation model generates as much carbon as 0.0006 miles driven by a similar vehicle, i.e. 6,833 times less,” they wrote.
Despite those researchers’ early findings, the question marks over the long-term environmental ramifications of consistent generative AI usage—especially as companies begin ramping up their implementation—remain in place.
Salesforce wants regulators to change that.
Megan Lorenzen, director of climate and energy at Salesforce, said the company knows collaboration between government and industry will be crucial as AI development continues.
“Today, AI is more ingrained in our everyday lives and the way we work than ever before, and its impact will only grow from here. As we start to understand the implications of this technology, we need to make sure we get it right from the start,” she said. “With the backdrop of an escalating climate crisis, we are looking to enable a more sustainable and equitable future through the use of AI—both minimizing environmental impact and sourcing climate innovation.”
Salesforce’s new policy priorities suggest regulators should require providers of general-purpose AI models to disclose both the carbon footprint and the energy ramifications associated with training and operating those models. It also wants to see the government stipulate standardized metrics for the proposed public-facing reporting efforts.
Additionally, it recommends that governments “consider environmental impact when determining the risk of AI systems and classifying high-risk models.” At the moment, this could prove most applicable to the EU member states, since the EU’s landmark AI Act separates models by their risk levels based on a number of factors. Including environmental risk could be on the docket for other countries without sprawling AI legislation in place, like the U.S., were it also to adopt a risk-based legal model for the technology.
Finally, Salesforce recommends that if or when the government classifies a model as high risk, it should stipulate targets for energy efficiency, like clean energy or low-carbon grids, for running the models appropriately.
Lorenzen did not say whether the company will use capital to lobby legislators on the policy priorities it has outlined. In 2023, it spent $3.26 million on U.S. lobbying efforts, all under the “Internet” category, according to Open Secrets data.
While Lorenzen said Salesforce was the first corporation to advocate for the Transformational AI to Modernize the Economy (TAME) Against Extreme Weather Act, which would leverage AI to predict climate events and model wildfire spreading, to date, it has not supported other active pieces of legislation.
For instance, Senator Ed Markey (D-MA) introduced the Artificial Intelligence Environmental Impacts Act of 2024 in February, which would stipulate that the National Institute of Standards and Technology (NIST) develop a rubric by which to measure and share the scope of AI’s environmental impacts, while also creating an opt-in framework to help AI developers share their environmental impacts. At present, Salesforce has not backed that bill.
Lorenzen said the company is looking to advocate for both state and federal-level policy and noted that “Salesforce is actively evaluating and engaging with lawmakers on several other pieces of legislation.” She did not indicate whether Markey’s proposal was included in the San Francisco tech giant’s legislative interests.
It remains to be seen how a number of governments will regulate AI—and how those regulations will iterate as companies continue to develop the technology. In the meantime, Salesforce said it has already prioritized ensuring it develops “right-sized” models, which effectively means that, rather than creating monstrous, general-purpose AI models, it will create models intended for specific use cases, which it said are “smaller and more sustainable, cost effective and easier to fine-tune, improving the user experience.”
It also works to use efficient hardware types, site its data centers in lower-carbon regions and disclose its own impact.
Lorenzen said Salesforce being proactive in its approach could help encourage other companies to understand the value of transparency.
“Our role as a market leader is to blaze a trail that others can follow. Salesforce has already done this in a variety of areas, like climate and sustainability disclosures, where we’ve been leading by example in reporting our own impacts and encouraging others to do the same,” she said. “If we don’t speak up, if we don’t work with our peers and governments to be bold and to create a sustainable future for all, then change at the scale and speed the planet needs will be out of reach.”