01Twitch Defaults Streamers Into Letting Amazon Train Generative AI on Their Content, Requiring an Opt-Out
Twitch, the Amazon-owned livestreaming platform, announced on Wednesday that streamers can now opt out of having their channel content used to train Amazon’s generative AI models. But the setting is enabled by default, meaning creators must actively turn it off if they do not want their material included in future training.
The change follows earlier experimentation with Twitch content. In 2024, Twitch Chief Monetization Officer Mike Minton said the platform’s material was already being used for generative AI prototyping, though not at production scale. The new setting confirms that Twitch is using streamer content to train Amazon models while giving users a way to refuse future use.
The toggle reads, “Allow your channel content to train generative AI content models at Amazon.” It appears at the bottom of each user’s security settings page, and all users are opted in by default. That design may leave creators enrolled if they miss the announcement or do not find and disable the setting.
Twitch says opting out prevents a broad range of channel material—including livestreams, videos on demand, clips, stream chats, and images and text displayed on a channel—from being used in future training. The restriction applies to Amazon-developed models designed to generate or synthesize text, audio, images or video.
The company’s language limits the promise to future training. Twitch has not said how much content was previously used, whether historical material can be deleted, or whether data collected before a creator opts out may be retained or used in existing models.
Turning off the setting also does not stop every AI or machine-learning use on Twitch. The platform says it will continue operating AutoMod, its automated moderation system, which it says does not retain user data to produce content. Twitch may also continue using data for AI-supported features such as automatic captions for clips, which it says do not retain streams or clips to train models capable of generating new material.
Amazon did not immediately respond to 404 Media’s request for comment, while messages sent to Twitch’s publicly listed press email repeatedly bounced. That leaves unanswered how much past streamer material Amazon has used and what happens to content submitted before an account holder finds the opt-out.
02Google Brings American Sign Language Input to Pixel 11, Starting With Gboard and Live Transcribe
Google has released SL2T 1.0, a technology developed jointly by Google DeepMind and Android to bring American Sign Language input to phones. The release moves the project from research into two consumer applications: Gboard and Live Transcribe.
Users can experience SL2T first on Pixel 11 at no additional cost, Google said. The company plans to support more devices “soon,” but its announcement does not provide a timetable or specify the technology’s recognition accuracy and supported use cases.
The Deaf community participated throughout the project rather than only testing the finished product. Google said Deaf perspectives shaped its initial conception by Sam Sepah, a Deaf Googler, as well as data collection with Deaf partners, evaluation through Deaf user studies and an impact assessment conducted with Deaf experts.
Google also established the AI Sign Language Advisory Committee, or AISLAC, to involve global Deaf organizations and subject-matter experts in decisions about development priorities. The company describes this arrangement as participatory governance, giving communities directly affected by the technology a role in guiding it.
For the SL2T 1.0 release in Gboard and Live Transcribe, Google and Deaf experts jointly wrote an impact report covering the system’s capabilities and current limitations. Google said it plans to continue that collaborative reporting approach for all major sign-language releases. The supplied announcement does not detail the limitations identified in the report.
The initial launch remains narrow: American Sign Language input, two applications and Pixel 11 as the first supported device. Google said that bringing ASL input to phones is only the beginning of a broader effort toward parity with spoken and written languages.
Its longer-term work includes extending the technology to additional sign languages, developing sign-language generation and adding more advanced AI capabilities. No release dates were given for those expansions, while support for additional devices is described only as coming soon.
03Google reveals Gemini usage patterns: 63% of active users use voice, while 20% of Gemini Live users share a camera or screen
Google says Gemini has reached 1 billion monthly active users faster than any of its other products, and it is now detailing how those people interact with the chatbot. The figures show substantial use of voice, real-time visual assistance and attachments, while pointing to new study tools planned for the coming weeks.
The billion-user count covers people who opened the Gemini app or web interface during the past month to enter a prompt or access Gemini Live, the product’s live conversational experience. A person who used Gemini only once still qualifies. The metric excludes people who encountered Gemini solely through Google Search, Gmail or Drive, even though the AI is integrated into those services.
Josh Woodward, the Google vice president overseeing Gemini, said 63% of Gemini’s active users use voice input. Google also sees a growing number of “voice-only” users, although it did not disclose their share or explain how it measured the behavior.
Visual interaction has a narrower denominator. Among people who use Gemini Live—not among all Gemini active users—20% share a camera feed or their screen to get help. Google did not provide a geographic breakdown for either the voice or visual-use figures.
Attachments are particularly common in education-related use. Woodward said 38% of school-related requests include an attachment. Google is responding by preparing a new batch of study tools that it plans to release in the coming weeks, but it has not described what those tools will do.
The company also said users generate 150 million images with its AI each day. Woodward said businesses use images produced with Nano Banana for marketing materials. Google applies its SynthID watermark to the generated images.
Taken together, the figures move the focus beyond Gemini’s billion-user milestone and toward the inputs people choose: speaking instead of typing, showing the assistant their surroundings or screens, and supplying files with school-related requests. The next concrete step is Google’s release of the still-undetailed study tools in the coming weeks.
SpaceXAI Launches Grok Bot as an Always-On Workplace Agent Grok Bot can use a shared cloud computer to sign into services and complete multistep assignments, escalating when approval is required. The beta is available to select SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers. theverge.com
OpenAI Executive Brad Lightcap Announces Departure Brad Lightcap, OpenAI’s special projects lead and former chief operating officer, is leaving after eight years and said he plans to start “something new.” He expects to remain at the company for several more weeks. theverge.com
Thrive Holdings Raises $2 Billion to Deploy AI Across Traditional Businesses OpenAI-backed Thrive Holdings raised $2 billion at a $12 billion valuation to expand its model of acquiring businesses and introducing AI into their workflows. The company, which currently focuses on accounting and IT, plans a third platform for regulatory work involving physical infrastructure. techcrunch.com
Lovable Raises $400 Million at a $13.3 Billion Valuation Lovable, a startup whose AI tools let users build software through natural-language prompting, raised a Series C led by Menlo Ventures and the Scaleup Europe Fund. The company said it reached $500 million in annualized revenue in June and now hosts 60 million projects. techcrunch.com
Cognition Reportedly Seeks a Valuation of at Least $40 Billion Cognition, maker of the Devin coding agent, is reportedly discussing another funding round only three months after raising $1 billion at a $26 billion valuation. Bloomberg’s sources said the new valuation would depend on the company reaching $1 billion in annualized revenue. techcrunch.com
Needle 2 Targets Offline Tool Calling on Low-Cost Devices Cactus Compute introduced Needle 2, a 45-million-parameter model designed to translate user requests into function calls and structured outputs on phones, wearables, and other constrained hardware. The company says its 2-bit model occupies 14MB, uses 28MB of session memory, and already runs locally in Pebble’s Index 01 app. cactuscompute.com
Pixel Watch 5 Adds Offline Gemini and AI-Based Health Trends Google’s $399 Pixel Watch 5 can run simple Gemini commands offline and proactively surface relevant actions and information. New blood-pressure and insulin-resistance trend features are scheduled for September and are described as nondiagnostic, while Breathing Emergency Detection remains pending FDA clearance in the US. theverge.com
OpenAI’s Head of Ethics Leaves Within a Year of Joining OpenAI’s head of ethics has departed less than a year after joining the company. ft.com
D’Addario Admits Suno Was Used to Regenerate Promotional Music Guitar-accessories company D’Addario acknowledged that Suno Studio was used to regenerate a track in a product video after initially denying generative AI involvement. The company apologized and said employees and creative partners will have to disclose future generative AI use. theverge.com
AI-Assisted Canine Cancer Project Becomes a Y Combinator Startup Entrepreneur Paul Conyngham launched Gamgee to develop personalized mRNA cancer vaccines for dogs, with recruitment underway for initial clinical trials in Australia. The company’s pitch stems from one dog treated with the vaccine alongside another therapy, making it impossible to attribute the reported tumor reduction specifically to the vaccine. theverge.com