Salesforce CEO Marc Benioff says AI models have ‘stolen’ copyrighted content from media companies

• Salesforce CEO Marc Benioff spent the annual Dreamforce conference promoting its AI products.
• Benioff issued warnings about the perils of generative AI while touting Salesforce’s AI as ethical.
• Benioff, who owns Time, said media companies are getting their information “stolen” by AI.

Salesforce CEO Marc Benioff’s message to companies interested in generative AI is simple: the technology isn’t ready for prime time. Unless his company is looking to sell it.

AI was a recurring theme at the company’s annual flagship conference, Dreamforce, which I attended this week in San Francisco. Salesforce, which makes cloud-based customer relationship management software, billed this year’s Dreamforce as “the world’s largest AI event” and kicked off the week with the announcement of Einstein 1, its new generative AI product.

Benioff spent much of the conference issuing dire warnings about the dangers of generative AI while also praising his own company’s new technology as ethical and secure.

Onstage at the Yerba Buena Center for the Arts, he prodded celebrities, politicians, and business leaders, including OpenAI CEO Sam Altman, to express their fears about the existential risks of feeding intelligent machines vast amounts of personal data, before reminding the audience that Salesforce was setting the standard for ethical and responsible data handling.

After two days of listening to this, I began to wonder how Benioff could be so certain that his large language models aren’t behaving as badly as everyone else’s and aren’t being trained on copyright-protected data.

“We’re not scraping the internet with our models, if that’s your question,” Benioff explained Wednesday afternoon.

Salesforce has “an open philosophy” when it comes to developing large language models, or LLMs, according to Benioff, building some on top of preexisting models and others from scratch. For example, Salesforce’s CodeGen LLM, which was released in 2022, was trained from the ground up using Apex, an internal programming language.

“It’s inside our core system,” Benioff explained. “It’s not being generated by the computer.”

Benioff, who purchased Time magazine in 2018, has a strong dislike for copyright infringement.

“As a media owner, it’s a major issue because I go to the models and I’ll find Time magazine material in there and go, ‘Wait a minute, that’s my content,'” Benioff explained.

“Media companies are starting to wake up and realize a lot of their information has been stolen — probably some of yours, too,” he added, a pointed look in my direction.

Kathy Baxter, Salesforce’s principal architect of responsible AI and technology, interjected before I could determine whether the look was made in jest.

Salesforce, according to Baxter, only uses customer data with consent and conducts legal reviews before using open-source data for AI model training.

“I’m very proud of all of the models we’re developing,” Baxter said. “We can be extremely transparent about that data, and you can stand by it knowing it is not copyrighted.” Never.”