San Francisco:

California legislators are set to vote on a bill as soon as this week that would broadly regulate how artificial intelligence is developed and deployed in California even as a number of tech giants have voiced broad opposition.

Here is background on the bill, known as SB 1047, and why it has faced backlash from Silicon Valley technologists and some lawmakers:

WHAT DOES THE BILL DO?

Advanced by State Senator Scott Wiener, a Democrat, the proposal would mandate safety testing for many of the most advanced AI models that cost more than $100 million to develop or those that require a defined amount of computing power. Developers of AI software operating in the state would also need to outline methods for turning off the AI models if they go awry, effectively a kill switch.

The bill would also give the state attorney general the power to sue if developers are not compliant, particularly in the event of an ongoing threat, such as the AI taking over government systems like the power grid.

As well, the bill would require developers to hire third-party auditors to assess their safety practices and provide additional protections to whistleblowers speaking out against AI abuses.

WHAT HAVE LAWMAKERS SAID?

SB 1047 has already passed the state Senate by a 32-1 vote. Last week it passed the state Assembly appropriations committee, setting up a vote by the full Assembly. If it passes by the end of the legislative session on Aug. 31, it would advance to Governor Gavin Newsom to sign or veto by Sept. 30.

Wiener, who represents San Francisco, home to OpenAI and many of the startups developing the powerful software, has said legislation is necessary to protect the public before advances in AI become either unwieldy or uncontrollable.

However, a group of California Congressional Democrats oppose the bill, including San Francisco’s Nancy Pelosi; Ro Khanna, whose congressional district encompasses much of Silicon Valley; and Zoe Lofgren, from San Jose.

Pelosi this week called SB 1047 ill-informed and said it may cause more harm than good. In an open letter last week, the Democrats said the bill could drive developers from the state and threaten so-called open-source AI models, which rely on code that is freely available for anyone to use or modify.

WHAT DO TECH LEADERS SAY?

Tech companies developing AI – which can respond to prompts with fully formed text, images or audio as well as run repetitive tasks with minimal intervention – have called for stronger guardrails for AI’s deployment. They have cited risks that the software could one day evade human intervention and cause cyberattacks, among other concerns. But they also largely balked at SB 1047.

Wiener revised the bill to appease tech companies, relying in part on input from AI startup Anthropic – backed by Amazon and Alphabet. Among other changes, he eliminated the creation of a government AI oversight committee.

Wiener also took out criminal penalties for perjury, though civil suits may still be brought.

Alphabet’s Google and Meta have expressed concerns in letters to Wiener. Meta said the bill threatens to make the state unfavorable to AI development and deployment. The Facebook parent’s chief scientist, Yann LeCun, in a July X post called the bill potentially harmful to research efforts.

OpenAI, whose ChatGPT is credited with accelerating the frenzy over AI since its broad release in late 2022, has said AI should be regulated by the federal government and that SB 1047 creates an uncertain legal environment.

Of particular concern is the potential for the bill to apply to open-source AI models. Many technologists believe open-source models are important for creating less risky AI applications more quickly, but Meta and others have fretted that they could be held responsible for policing open-source models if the bill passes. Wiener has said he supports open-source models and one of the recent amendments to the bill raised the standard for which open-sourced models are covered under its provisions.

The bill also has its backers in the technology sector. Geoffrey Hinton, widely credited as a “godfather of AI,” former OpenAI employee Daniel Kokotajlo and researcher Yoshua Bengio have said they support the bill.

(Except for the headline, this story has not been edited by NDTV staff and is published from a syndicated feed.)

New Delhi:

A traditional Google search consumes 0.0003-kilowatt hour (KWh) energy on an average, according to a 2009 Google report. That energy can power your household (9 watt) light bulb for about 2 minutes. Google averages about 8.5 billion searches per day as of 2023, translating to 2,550,000 KWh worth of electricity per day, which is about 2000 times more electricity than an average Indian consumes (1255 KWh) for an entire year.

This May, Google announced that the company will integrate AI into its search engine, which will be powered by its most powerful AI model – Gemini.  According to Alex de Vries, a Dutch Data Scientist who spoke to The New Yorker on the subject, a single Google search with AI integrated in it will consume 10 times more energy (3 KWh) than the traditional Google search. That is 20,000 times more than the average Indian’s consumption for a year. 

Alex de Vries, who is also the founder of Digiconomist-the organisation responsible for Bitcoin Energy Consumption Index- has said that Google’s energy consumption will reach about 29 billion terawatt-hours (TWh) per year if they go ahead with the AI integration in its search. This figure is equivalent to the electricity consumption of Ireland, and more than that of Kenya.

 
Why are AI systems so hungry for power? 

AI systems require a lot of computational power for running complex algorithms to process large corpus of ever-increasing data. When you enter a prompt in ChatGPT, it is processed by the chatbot using its servers hosted in data centres. These centres alone account for 1-1.5 percent of the entire global electricity use, according to the International Energy Agency.

“I think we still don’t appreciate the energy needs of this (AI) technology,” Sam Altman, C.E.O of OpenAI, said at a public event in Davos this January. Altman expressed the immediate need for a “breakthrough” technology like nuclear fusion to power the AI operations given the growth projections of the frontier technology. This is a clear indication that the industry leader of the Large Language Model chatbots is seeking avenues to sustain current electricity consumption levels and safeguard its future demand.  

AI’s Carbon Emissions -Setback for UN 2050 Net-Zero Emission Goal? 

Data centres, powering cloud computing as well as AI systems produce 2.5 to 3.5 percent of global greenhouse gas emissions, according to The Shift Project, a French nonprofit working to reduce energy reliance on fossil fuels. This amounts to the same levels of greenhouse gas emissions as the entire aviation industry.

The energy consumption and the subsequent carbon footprint of different AI models vary significantly. For example, BigScience project BLOOM, which is an AI model with 76 billion parameters (internal variables that a model learns during training) consumes 433 megawatt-hour (MWh) of electricity. 

In contrast, OpenAI’s GPT-3 from 2020, which had a comparable number of parameters at 175 billion, consumed 3 times more electricity at 1287 MWh, according to the Artificial Intelligence Index Report 2024 data, published by Stanford Institute of Human-Centered Artificial Intelligence (HCAI).

The CO2 equivalent emissions (in tonnes), that is the total greenhouse gas emissions expressed in terms of carbon dioxide, for BLOOM was 25 tonnes and for GPT-3, it was a whopping 20 times more, at 502 tonnes.

The report also states that there is a serious lack of transparency from the AI developers on the environmental impact of their models, with most not making their carbon footprints public. 

The recently released Google’s Sustainability Report 2024 also reflects the energy hunger of the nascent technology of AI. The company saw a near 50 percent surge in carbon emissions in the last 5 years on the back of powering its new AI technologies. Microsoft’s Sustainability Report 2024 shows similar trends with a spike of 29 percent increase in CO2 emission in 2023, compared to the previous year.   

According to SemiAnalysis, a US-based independent AI research and analysis company, AI will drive the increase of electricity consumption of data centres to 4.5 percent of the global energy generation by 2030. Another estimate from the International Energy Agency suggests that data centres’ total electricity consumption can double from 2022 levels to 1000 TWh (equaling the current electricity consumption of Japan). India has around 138 data centres, with reportedly 45 more set to be functional by the end of 2025. The United States has the highest number of data centres with 2701 of them.  

Lawmakers are starting to take stock of the situation. The European Union has taken cognisance and adopted a new regulation in March this year. Under the scheme, all data centre operators are required to report their energy and water consumption (used for cooling systems). They are also mandated to provide information on efficiency measures being implemented to ensure reduction. 

In February, US Democrats introduced the Artificial Intelligence Environment Impact Act of 2024. The act proposes setting up an AI Environment Impact Consortium of experts, researchers and industry stakeholders to address the environmental impact of AI.