What's in Your AI? California Thinks You Should Know

While California Governor Gavin Newsom drew attention in the past month for saying no to a comprehensive AI safety law that would have called for an AI kill switch, the state did pass 17 gen AI-related bills covering deepfakes, AI watermarking, child safety, election misinformation and performers' rights regarding their digital likeness.

For those interested in what goes into the training data – the billions of bits of content and data being fed into large language models (LLMs) – let me point you to AB 2013, which Newsom did sign. The bill requires that AI companies, on or before Jan. 1, 2026, provide a "high-level summary" of the datasets used to train their systems, including the sources and owners of the data and whether the data was licensed or purchased by the AI company.

That's kind of a big deal since publishers, including The New York Times, are suing OpenAI and Microsoft, arguing that their AI model is built from copyrighted content essentially stolen from their websites. OpenAI and Microsoft counter that if their data does contain copyrighted material, it falls under the fair use provision. Even so, OpenAI has recognized that it may need to pay content creators, given that it's signed licensing deals with publishers, including The Atlantic, News Corp. and Vox Media.

AB 2013, which applies to systems released in or after January 2022, also requires AI developers to say whether the training data includes personal information, the dates when the data became part of the system and "the time period during which the data in the datasets were collected, including a notice if the data collection is ongoing."

Crazy thought: What if AI companies, hoping to gain the trust of consumers and publishers, voluntarily released information about their training data ahead of the deadline? I'm not holding my breath.

Here are some other doings in AI worth your attention.

As election day approaches, constant vigilance required

With US elections just two weeks away, remember to pause before believing or resharing content about candidates, issues and even the integrity of the election process. Because just as Russian disinformation groups used social media to influence voters and spread disinformation in the 2016 and 2020 elections, bad actors, including Russia, are using generative AI tools such as OpenAI's ChatGPT to mislead US voters. 

Who says? OpenAI for one. In a 54-page report released this month, the company said it's "disrupted more than 20 operations and deceptive networks from around the world that attempted to use our models" since the start of 2024 that were intended to trick and confuse voters. Those threats ranged from writing articles for websites to creating social media posts shared by "fake personas."

"Activities ranged in complexity from simple requests for content generation, to complex, multi-stage efforts to analyze and reply to social media posts," OpenAI said. 

There was some good news. While "threat actors continue to evolve and experiment with our models, but we have not seen evidence of this leading to meaningful breakthroughs in their ability to create substantially new malware or build viral audiences," the company wrote. "It is noteworthy that ... the deceptive activity that achieved the greatest social media reach and media interest was a hoax about the use of AI, not the use of AI itself." Details of the hoax can be found on page 20.

How big a deal are the concerns about AI and the elections? Americans – both Democrats and Republicans – said they're worried that AI will be used mostly for "bad purposes," the Pew Research Center noted in a September report. A majority of adults (57%) said they're extremely or very concerned that groups will try to influence the election using AI "to create and distribute fake or misleading information about the candidates and campaigns."

And while more than three-fourths of US adults believe tech and social media companies "have a responsibility to prevent the misuse of their platforms to influence the 2024 presidential election," only 20% say they are very or somewhat confident that these companies will actually be able to prevent their products from being misused, Pew added. 

Sigh.

So quoting Mad-Eye Moody from the Harry Potter universe, remember that constant vigilance will be required on your part as you sift through voter information (that may be written by AI bots), videos and audios (that may turn out to be AI-generated deepfakes) and even seemingly trivial posts that may try to plant a seed of doubt in your mind about vote counts.

The Wired AI Elections Project is tracking AI use in this election cycle, noting that "2024 is already an unprecedented year for democracy: More than 2 billion people – the largest number ever – will vote in national, local, and regional elections in more than 60 countries."  And several other reputable groups, including the Brennan Center for Justice, the News Literacy Project and its RumorGuard alert program, FactCheck.org  and NewsGuard, can help you suss out what's real from fake about this year's elections.

AI video tools gearing up for prime time, starting with Adobe

As I wrote last week, companies including OpenAI, Meta and Google have announced they're developing AI tools that let you create hiqh-quality videos from simple text prompts or from a single photo or image. 

Count Adobe first out of the gate with a public beta of its Adobe Firefly Video Model after having a set of creatives experiment with its text-to-video and image-to-video systems in September. You can watch demos of what the video looks like and how it's being used by those early users. 

To help temper the concerns of filmmakers and creators worried that such tools will undermine their work, Adobe notes that it has licenses to all the video content its AI models have been fed, or trained on. 

"Just like our other Firefly generative AI models," the company wrote in a blog post, "Adobe Firefly Video model is designed to be commercially safe and is only trained on content we have permission to use – never on Adobe users' content."

The public beta comes ahead of OpenAI's Sora, which is due sometime this year, and Meta's recently announced Movie Gen text-to-video creator, which can create 16-second videos. Meta CEO Mark Zuckerberg, in a post created with the tool, wrote that it will be available to the public in Instagram next year. To also show it's taking concerns by creatives into account as it develops its powerful video engine, Meta last week announced that it's collaborating with the Blumhouse production company – "a driving force in  horror" – and a small group of filmmakers, including award-winning actor Casey Affleck, to get "creative industry feedback" on these tools.

"The thing that has surprised me the most about it, and that has been the most enjoyable, is how it's more like a collaborator than it is like a tool," Affleck says in a Meta promo video.

Meanwhile, Microsoft, which in July said it wants the government to pass a law to protect against deepfake fraud, filed a patent application for "providing live image generation based on audio transcription." 

"Microsoft mentions in the document that this AI-powered system converts live audio (during meetings or conferences, for example), and then a language model summarizes that. And then, with this summary, the system creates an AI-generated image," reports MSPowerUser, which spotted the Oct. 10 patent application first.     

ChatGPT tops the AI charts with 3 billion visits

After watching the number of visits to its site rise and fall throughout 2023, ChatGPT is now leader of the pack: It's seen traffic to its site climb steadily over the past five months. Visits more than doubled in September to a record 3.1 billion visits, according to researcher Similarweb.

"That makes ChatGPT (chatgpt.com) the #11 website by worldwide traffic – not in the same league as google.com (82 billion visits) or youtube.com (28 billion), but ahead of amazon.com (2.6 billion, although that doesn't count Amazon's international domains)," Similarweb said.

In the chatbot market, ChatGPT's closest rival in terms of traffic in the US is Microsoft's Bing.com, which had under 2 billion visits in September. Google's Gemini was the third most-visited chatbot, with 274.7 million visits, the researcher said.

The news comes just weeks after OpenAI made history by raising an eye-popping $6.6 billion in investor funding, which valued the company at a robust $157 billion.

Also worth knowing…

Is AI the answer to your money problems? CNET's Kelly Ernst set out to find out in her report on how budget-conscious folks are using chatbots to help make better spending decisions. She includes a guide to popular AI apps and use cases based on insight from financial experts. 

A month after Microsoft said it had signed a deal to power its data centers with the help of the nuclear reactor at Three Mile Island, Amazon and Google both announced nuclear energy plans to help feed expected demand for all the AI compute power they'll need. Amazon said it's investing $500 million in developing nuclear technologies. Google said it signed an agreement to buy nuclear energy from small modular reactors being developed by Kairos Power.

The government's adoption of AI is paying off, seriously. The US government said its use of AI – not gen AI, but machine learning – helped "recover $1 billion worth of check fraud in 2024 alone," reported CNN, which got the scoop from the Treasury Department. CNN added that "US officials quietly started using AI to detect financial crime in late 2022, taking a page out of what many banks and credit card companies already do to stop bad guys."