We have updated our Privacy Policy and Terms of Use for Eurasia Group and its affiliates, including GZERO Media, to clarify the types of data we collect, how we collect it, how we use data and with whom we share data. By using our website you consent to our Terms and Conditions and Privacy Policy, including the transfer of your personal data to the United States from your country of residence, and our use of cookies described in our Cookie Policy.
AI
There’s a voice on the internet that sounds just like me. In a way, it is me.
ElevenLabs, a startup software company, uses artificial intelligence to simulate natural-sounding language. If you need a realistic voiceover for a television commercial, short film, or an audiobook, ElevenLabs lets you use one of the human-ish voices in its library. While this technology isn’t ready to put voice actors out of business just yet — there are some definite hiccups — the technology is surprisingly effective.
In January, New Hampshire voters received a robocall from Joe Biden. Except it wasn’t Joe Biden. It was an AI-generated fake made by a political ally of then-Democratic challenger Dean Phillips. It was made with ElevenLabs.
In response, ElevenLabs banned the account responsible and in February changed its policy around impersonating public figures. While the company already banned impersonation of private figures without their consent, or anything intended to harm others, ElevenLabs added a list of “no-go” voices, specifically outlawing impersonation of presidential or prime minister candidates in the US and UK with the intention of adding additional “languages and election cycles.”
With elections in 64 countries this year, and OpenAI planning its own ElevenLabs-like tool too, the threat of mass confusion around elections feels palpable. AI has helped former Pakistani Prime Minister Imran Khandeliver a campaign address from prison, but it also caused problems for Indonesian political candidate Anies Baswedan, the victim of audio pretending to be a political backer chastising him. In addition to spreading disinformation, these tools could also be used to break voice authentication measures used by banks and other financial institutions.
Naturally, I had to give it a try.
Scott's AI-generated voice, courtesy of ElevenLabs.roar-assets-auto.rbl.ms
I paid $22 for the mid-tier version of ElevenLabs, which got me “professional voice cloning,” about two hours of downloadable AI-generated text-to-speech per month, and high-quality audio. To generate my voice clone, I had to upload 10-plus minutes of myself speaking, though it recommended 30 for the best results. After 12 minutes of gabbing into my USB microphone about the results of the NFL Draft, I hit upload.
The software took about two hours to do its magic before I was alerted that my cloned voice was ready to chat. Even just typing in, “Hey, this is Scott Nover,” I was amazed how much it sounded like me. I typed some more. The more text I included, the more the software struggled to sound human. There were unnatural pauses, odd diction, and my detached voice seemed upset about something.
Scott's AI-generated voice, speaking Mandarin, courtesy of ElevenLabs.roar-assets-auto.rbl.ms
ElevenLabs lets you choose from a few different models, and it offers sliders to adjust the voice’s tone, similarity to the uploaded recording, and sassiness. I tinkered with the settings until I found the best ratios and walked away pretty impressed with what I created. It even let me change the language — my polyglot editor Matthew Kendrick says the Spanish sounds good, with a mild Mexican accent and a few odd word choices. The Mandarin is a little less convincing — mostly because white dudes usually can’t hit their tones so accurately.
Scott's AI-generated voice, speaking Spanish, courtesy of ElevenLabs.roar-assets-auto.rbl.ms
If you’ve been following this newsletter, you’ll know we’ve been covering the rise of generative AI and have warned, in particular, about the rise of AI-generated audio.
Soon, AI voices will be everywhere. Some, like my homemade ScottBot, won't matter much. But some—say, a homemade clone of the Ayatollah Khamenei issuing a fatwa or Vladimir Putin inciting violence in Ukraine—could cause major problems on the world stage. And every government should be prepared.Eight major newspapers owned by Alden Global Capital sued ChatGPT maker OpenAI on Tuesday in federal court, alleging copyright infringement. The group includes major regional newspapers, such as the New York Daily News, the Chicago Tribune, and the Orlando Sentinel.
While many publishers have struck lucrative licensing deals with AI companies, allowing them to train their large language models by scraping their websites, others have opted for litigation, most notably the New York Times. The Grey Lady sued OpenAI in December, alleging the Sam Altman-led startup violated federal copyright law by illegally training its model on Times journalism and spitting out responses indistinguishable from what people can find on their website or in their print newspaper. OpenAI has said the suit is “without merit.”
The Alden group also followed the Times' lead in suing Microsoft, too. Microsoft is OpenAI's biggest investor, having poured $13 billion into the startup, and uses the GPT suite of language models in its Copilot assistant and Bing search products.If a large language model proprietor is found to have violated copyright statutes, it could pose an existential threat to that model — meaning it may have to start training a new one from scratch.
Hard Numbers: Unnatural gas needs, Google’s data centers, Homeland Security’s new board, Japan’s new LLM
8.5 billion: Rising energy usage from AI data centers could lead to additional demand for natural gas of up to 8.5 billion cubic feet per day, according to an investment bank estimate. Generative AI requires high energy and water demands to power and cool expansive data centers, which climate advocates have warned could exacerbate climate change.
32 billion: Google is pouring $3 billion into data center projects to power its AI system. That budget includes $2 billion for a new data center in Fort Wayne, Ind., and $1 billion to expand three existing ones in Virginia. In earnings reports this week, Google, Meta, and Microsoft disclosed that they had spent $32 billion on data centers and related capital expenditures in the first quarter alone.
22: The US Department of Homeland Security announced a new Artificial Intelligence Safety and Security Board with 22 members including the CEOs of Alphabet (Sundar Pichai), Anthropic (Dario Amodei), OpenAI (Sam Altman), Microsoft (Satya Nadella), and Nvidia (Jensen Huang). The goal: to advise Secretary Alejandro Mayorkas on “safe and secure development and deployment of AI technology in our nation’s critical infrastructure.”
960 million: SoftBank, the Japanese technology conglomerate, plans to pour $960 million to upgrade its computing facilities in the next two years in order to boost its AI capabilities. The company’s broad ambitions include funding and developing a large language model that’s “world-class” and geared specifically toward the Japanese language.CRISPR, the gene-editing method that won two female scientists the 2020 Nobel Prize in Chemistry, may soon get infused with artificial intelligence. One Northern California startup called Profluent is expected to present its new paper at a gene-editing conference next month, which describes its work using AI to analyze biological data and create new gene-editing systems.
As one professor explained to the New York Times, it’s a departure from how CRISPR typically does gene replacement. Instead of altering genes based on discoveries in nature, the startup instead uses novel methods surfaced by its AI. “They have never existed on Earth,” University of California, San Francisco professor James Fraser said. “The system has learned from nature to create them, but they are new.”
Gene-editing is rife with ethical quandaries, such as questions around modifying human embryos, which could be exacerbated by the rise of AI. Still, CRISPR provides hope: it could provide cures to countless diseases and is already providing innovative new treatments for sickle-cell anemia.
Profluent also chose to open-source one of its gene editors, OpenCRISPR-1, though the underlying AI will stay under wraps, the company said.
Dazhon Darien, a former athletic director at Pikesville High School in Baltimore County, Maryland, was arrested on April 25 and charged with a litany of crimes related to using AI to frame the school's principal. Darien allegedly created a fake AI voice of Principal Eric Eiswert, used it to generate racist and antisemitic statements, and posted the audio on social media in January. Eiswert was temporarily removed from the school after the audio emerged.
The police allege that Darien used the school’s internet to search for AI tools and sent emails about the recording. The audio was then sent to and posted by a popular Baltimore-area Instagram account on Jan. 17. It’s unclear which tool was used to make the recording, but digital forensics experts said it was clearly fake.
At least 10 states have some form of deepfake laws, though some are focused on pornography. Still, AI-specific charges are rare in the US. Darien was charged with disrupting school activities, theft, retaliation against a witness, and stalking.
Deepfake audio has become a major problem in global elections, but this story demonstrates it can also easily weaponize person-to-person disputes.
US President Joe Biden on Monday signed an expansive executive order about artificial intelligence, ordering a bevy of government agencies to set new rules and standards for developers with regard to safety, privacy, and fraud. Under the Defense Production Act, the administration will require AI developers to share safety and testing data for the models they’re training — under the guise of protecting national and economic security. The government will also develop guidelines for watermarking AI-generated content and fresh standards to protect against “chemical, biological, radiological, nuclear, and cybersecurity risks.”
The US order comes the same day that G7 countries agreed to a “code of conduct” for AI companies, an 11-point plan called the “Hiroshima AI Process.” It also came mere days before government officials and tech-industry leaders meet in the UK at a forum hosted by British Prime Minister Rishi Sunak. The event will run tomorrow and Thursday, Nov. 1-2, at Bletchley Park. While several world leaders have passed on attending Sunak’s summit, including Biden and Emmanuel Macron, US Vice President Kamala Harris and European Commission President Ursula von der Leyen plan to participate.
When it comes to AI regulation, the UK is trying to differentiate itself from other global powers. Just last week, Sunak said that “the UK’s answer is not to rush to regulate” artificial intelligence while also announcing the formation of a UK AI Safety Institute to study “all the risks, from social harms like bias and misinformation through to the most extreme risks of all.”
The two-day summit will focus on the risks of AI and its use of large language models trained by huge amounts of text and data.
Unlike von der Leyen’s EU, with its strict AI regulation, the UK seems more interested in attracting AI firms than immediately reining them in. In March, Sunak’s government unveiled its plan for a “pro-innovation” approach to AI regulation. In announcing the summit, the government’s Department for Science, Innovation, and Technology boasted the country’s “strong credentials” in AI: employing 50,000 people, bringing £3.7 billion to the domestic economy, and housing key firms like DeepMind (now owned by Google), while also investing £100 million in AI safety research.
Despite the UK’s light-touch approach so far, the Council on Foreign Relations described the summit as an opportunity for the US and UK, in particular, to align on policy priorities and “move beyond the techno-libertarianism that characterized the early days of AI policymaking in both countries.”- UK AI Safety Summit brings government leaders and AI experts together - GZERO Media ›
- AI agents are here, but is society ready for them? - GZERO Media ›
- Yuval Noah Harari: AI is a “social weapon of mass destruction” to humanity - GZERO Media ›
- Should we regulate generative AI with open or closed models? - GZERO Media ›
- Podcast: Talking AI: Sociologist Zeynep Tufekci explains what's missing in the conversation - GZERO Media ›
- OpenAI is risk-testing Voice Engine, but the risks are clear - GZERO Media ›
Meta is one of the biggest players in generative AI — but, while Wall Street typically loves AI chatter from companies, an episode this week showed that there are limits to this unbridled enthusiasm.
Meta posted impressive first-quarter earnings last week, but the company’s stock plunged 15% on news that it would spend $35-40 billion this year on AI-related investments alone. (That’s up from a previous estimate of $30-37 billion.) The costs were largely for AI infrastructure: data centers, graphics chips, and research and development.
Just mentioning AI has tended to goose valuations on Wall Street. Is this the first indication that’s changing, or is it Meta-specific? The company still makes most of its money from advertising on its social media platforms Facebook and Instagram, and it’s made its Llama line of generative AI models (mostly) open-source, meaning anyone can access and modify the model online for free. Open-source models aren’t impossible to monetize, but for now, it seems Meta’s AI investments are meant to keep more users on its platforms. Meta flooded its apps with AI features last week, adding a chatbot experience in the Instagram search bars, the Facebook news feed, and Messenger conversations.
Meta’s AI spending has made it an unavoidable player in the space. It’s adopting a very different strategy than any other AI company — and one that investors may not value as highly.