hello everyone. Welcome to TechCrunch's regular AI newsletter.
In the AI space this week, record labels accused Udio and Suno, two startups developing AI-based song generators, of copyright infringement.
The RIAA, a trade group representing the U.S. recording industry, announced Monday the lawsuit filed by Sony Music Entertainment, Universal Music Group, Warner Records and others. The lawsuit alleges that Udio and Suno trained the generative AI models that underpin the platform on the labels' music without compensation to the labels and seeks compensation of $150,000 per allegedly infringed work.
“Synthetic music output could saturate the market with machine-generated content that directly competes with, undercuts, and ultimately drowns out the actual sound recordings upon which the services are built,” the labels said in the complaint.
The lawsuit adds to the growing body of lawsuits against generative AI vendors, including giants like OpenAI, and makes nearly identical claims. This means that companies that train on copyrighted works must pay or at least credit rights holders and allow them to opt out of training. If they want. Vendors have long argued for fair use protections, arguing that the copyrighted data they learn from is public and that their models produce transformative works rather than plagiarism.
So how will the court rule? That, dear reader, is the billion-dollar question, and it will take a long time to resolve.
With mounting evidence showing that generative AI models can almost regress, you'd think this would be a slam dunk for rights holders (emphasis mine). almost) Represent copyrighted artwork, books, songs, etc. exactly as they were trained. But generative AI vendors are free from criticism, and good luck to Google for ultimately setting a precedent.
More than a decade ago, Google began scanning millions of books to build an archive for Google Books, a sort of search engine for literary content. Authors and publishers have sued Google over this practice, claiming that copying their IP online constitutes infringement. But they lost. On appeal, the court ruled that Google Books' copy had a “very compelling transformative purpose.”
If the plaintiff can't prove that the vendor's model actually plagiarized at scale, the court may decide that the generative AI has a “very compelling transformative purpose.” Or, as The Atlantic's Alex Reisner suggests, there may be no single ruling on whether generative AI technology as a whole is infringing. Judges can decide on a winner on a model-by-model, case-by-case basis, considering each output produced.
My colleague Devin Coldewey put it succinctly this week: “Not all AI companies are quite free to leave their fingerprints at crime scenes.” As the lawsuit progresses, you can be sure that AI vendors whose business models depend on the results are taking detailed notes.
News
Advanced Voice Mode Delay: OpenAI has postponed Advanced Voice Mode, an eerily realistic, near-real-time conversational experience for ChatGPT, an AI-based chatbot platform. But OpenAI, which this week acquired remote collaboration startup Multi and launched a macOS client for all ChatGPT users, has no idle staff.
Stability provides a lifeline. Stability AI, maker of the financially distressed open image generation model Stable Diffusion, was rescued by a group of investors including Napster founder Sean Parker and former Google CEO Eric Schmidt. The company also appointed former Weta Digital head Prem Akkaraju as its new CEO after debt relief, as part of a broader effort to regain its footing in the highly competitive AI landscape.
Gemini comes to Gmail. Google has launched a new Gemini-powered AI side panel to Gmail to help you compose emails and summarize threads. The same side panel applies to the rest of the search giant's suite of productivity apps, such as Docs, Sheets, Slides, and Drive.
Great Curators: Otis Chandler, co-founder of Goodreads, launched Smashing, an AI- and community-driven content recommendation app that aims to connect users with their interests by exposing the internet's hidden gems. Smashing automatically identifies topics and threads of interest to individual users by providing news summaries, key excerpts, and interesting quotes, and encourages users to like, save, and comment on articles.
Apple rejects Meta's AI. A few days after the Wall Street Journal reported that Apple and Meta were in talks to integrate the latter’s AI model, Bloomberg’s Mark Gurman said the iPhone maker is not planning such a move. Bloomberg said Apple has shelved the idea of putting Meta AI on iPhones, citing privacy concerns and partnerships with social networks whose privacy policies are often criticized.
Research paper of the week
Beware of Russian-influenced chatbots. It could be right under your nose.
Earlier this month, Axios highlighted research by NewsGuard, an organization that combats misinformation. The study found that leading AI chatbots were regurgitating parts of Russian propaganda campaigns.
NewsGuard fed 10 leading chatbots, including OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini, and dozens of prompts asking about stories purportedly created by Russian propagandists, particularly American fugitive John Mark Dougan. According to the company, the chatbot responded with false information 32% of the time, misrepresenting the Russian-authored report.
The study shows that scrutiny of AI vendors is intensifying as election season approaches in the United States. Microsoft, OpenAI, Google and other major AI companies agreed at the Munich Security Conference in February to take steps to curb the spread of deepfakes and election-related misinformation. However, abuse of the platform is still rampant.
“This report concretely demonstrates why the industry needs to pay special attention to news and information,” Steven Brill, co-CEO of NewsGuard, told Axios. “At this time, do not trust the answers provided by most chatbots on news-related issues, especially controversial ones.”
model of the week
Researchers at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) claim to have developed DenseAV, a model that can learn language by predicting what it hears and sees.
Researchers led by Mark Hamilton, a doctoral student in electrical engineering and computer science at MIT, were inspired to create DenseAV by the nonverbal ways in which animals communicate. “We thought we should use audio and video to learn languages,” he told MIT CSAIL’s press office. “Is there a way to have an algorithm watch TV all day and figure out what we’re talking about?”
DenseAV only processes two types of data – audio and visual – and “learns” by comparing pairs of audio and visual signals to find which signals match and which do not. Trained on a dataset of 2 million YouTube videos, DenseAV can identify objects from their names and sounds by searching and then aggregating all possible matches between audio clips and image pixels.
For example, when DenseAV listens to a dog barking, one part of the model focuses on language, such as the word “dog,” and another part focuses on the barking sound. The researchers say this shows that DenseAV can not only learn the meaning of words and the location of sounds, but can also learn to distinguish these “cross-modal” connections.
Going forward, the team aims to create a system that can learn from massive amounts of video- or audio-only data, scale the work to larger models, and integrate with knowledge from language understanding models to improve performance.
grab the bag
No one can accuse OpenAI CTO Mira Murati of being consistently dishonest.
In a speech at the Dartmouth Institute of Technology, Murati acknowledged that generative AI would eliminate some creative jobs, but suggested that those jobs “should never have existed in the first place.”
“I definitely expect a lot of jobs to change. Some jobs will disappear and some jobs will be created,” she continued. “The truth is, we don’t really understand the impact of AI on jobs yet.”
The creators did not take kindly to Murati's comments. Of course. Callous language aside, OpenAI, like the aforementioned Udio and Suno, is facing lawsuits, critics, and regulators who claim it is profiting from artists' work without compensation.
OpenAI recently pledged to release tools to give creators more control over how their works are used in its products, and continues to negotiate licensing agreements with copyright holders and publishers. However, the company is not lobbying for universal basic income or leading any meaningful efforts to reskill or upskill the workforce it impacts.
According to a recent article in The Wall Street Journal, contract jobs that require basic writing, coding, and translation are disappearing. And a study published last November found that freelancers have fewer jobs and are earning significantly less since the launch of OpenAI's ChatGPT.
OpenAI’s stated mission is to “ensure that AGI (artificial general intelligence)—AI systems that are generally smarter than humans—can benefit all of humanity,” at least until it becomes a for-profit company. It hasn’t achieved AGI. But wouldn’t it be commendable if OpenAI, true to the “benefit all of humanity” part, set aside a tiny percentage of its revenues (over $3.4 billion) to pay creators so they don’t get swept up in the generative AI deluge?
I can dream, can't I?
