In the summer of 2023, OpenAI created a “Superalignment” team whose goal was to steer and control future AI systems that could be so powerful they could lead to human extinction. Less than a year later, that team is dead.

OpenAI told Bloomberg that the company was “integrating the group more deeply across its research efforts to help the company achieve its safety goals.” But a series of tweets from Jan Leike, one of the team’s leaders who recently quit revealed internal tensions between the safety team and the larger company.

In a statement posted on X on Friday, Leike said that the Superalignment team had been fighting for resources to get research done. “Building smarter-than-human machines is an inherently dangerous endeavor,” Leike wrote. “OpenAI is shouldering an enormous responsibility on behalf of all of humanity. But over the past years, safety culture and processes have taken a backseat to shiny products.” OpenAI did not immediately respond to a request for comment from Engadget.

X

Leike’s departure earlier this week came hours after OpenAI chief scientist Sutskevar announced that he was leaving the company. Sutskevar was not only one of the leads on the Superalignment team, but helped co-found the company as well. Sutskevar’s move came six months after he was involved in a decision to fire CEO Sam Altman over concerns that Altman hadn’t been “consistently candid” with the board. Altman's all-too-brief ouster sparked an internal revolt within the company with nearly 800 employees signing a letter in which they threatened to quit if Altman wasn’t reinstated. Five days later, Altman was back as OpenAI’s CEO after Sutskevar had signed a letter stating that he regretted his actions.

When it announced the creation of the Superalignment team, OpenAI said that it would dedicate 20 percent of its computer power over the next four years to solving the problem of controlling powerful AI systems of the future. “[Getting] this right is critical to achieve our mission,” the company wrote at the time. On X, Leike wrote that the Superalignment team was “struggling for compute and it was getting harder and harder” to get crucial research around AI safety done. “Over the past few months my team has been sailing against the wind,” he wrote and added that he had reached “a breaking point” with OpenAI’s leadership over disagreements about the company’s core priorities.

Over the last few months, there have been more departures from the Superalignment team. In April, OpenAI reportedly fired two researchers, Leopold Aschenbrenner and Pavel Izmailov, for allegedly leaking information.

OpenAI told Bloomberg that its future safety efforts will be led by John Schulman, another co-founder, whose research focuses on large language models. Jakub Pachocki, a director who led the development of GPT-4 — one of OpenAI’s flagship large language models — would replace Sutskevar as chief scientist.

Superalignment wasn’t the only team at OpenAI focused on AI safety. In October, the company started a brand new “preparedness” team to stem potential “catastrophic risks” from AI systems including cybersecurity issues and chemical, nuclear and biological threats.

OpenAI and Reddit announced a partnership on Thursday that will allow OpenAI to surface Reddit discussions in ChatGPT and for Reddit to bring AI-powered features to its users. The partnership will “enable OpenAI’s tools to better understand and showcase Reddit content, especially on recent topics,” both companies said in a joint statement. As part of the agreement, OpenAI will also become an advertising partner on Reddit, which means that it will run ads on the platform.

The deal is similar to the one that Reddit signed with Google in February, and which is reportedly worth $60 million. A Reddit spokesperson declined to disclose the terms of the OpenAI deal to Engadget and OpenAI did not respond to a request for comment.

OpenAI has been increasingly striking partnerships with publishers to get data to continue training its AI models. In the last few weeks alone, the company has signed deals with the Financial Times and Dotdash Meredith. Last year, it also partnered with German publisher Axel Springer to train its models on news from Politico and Business Insider in the US and Bild and Die Welt in Germany.

Under the new arrangement, OpenAI will get access to Reddit’s Data API, which, the company said, will provide it with “real time, structured, and unique content from Reddit.” It’s not clear what AI-powered features Reddit will build into its platform as a result of the partnership. A Reddit spokesperson declined to comment.

Last year, getting access to Reddit’s data, a rich source of real time, human generated, and often high-quality information, became a contentious issue after the company announced that it would start charging developers to use its API. As a result, dozens of third-party Reddit clients were forced to shut down and thousands of subreddits went dark in protest. At the time, Reddit stood its ground and said that large AI companies were scraping its data with no payment. Since then, Reddit has been monetizing its data by striking such deals with Google and OpenAI, whose progress in training their AI models depends on having access to it.

At the end of I/O, Google’s annual developer conference at the Shoreline Amphitheater in Mountain View, Google CEO Sundar Pichai revealed that the company had said “AI” 121 times. That, essentially, was the crux of Google’s two-hour keynote — stuffing AI into every Google app and service used by more than two billion people around the world. Here are all the major updates from Google's big event, along with some additional announcements that came after the keynote.

Gemini 1.5 Flash and updates to Gemini 1.5 Pro

Google

Google announced a brand new AI model called Gemini 1.5 Flash, which it says is optimised for speed and efficiency. Flash sits between Gemini 1.5 Pro and Gemini 1.5 Nano, which its the company’s smallest model that runs locally on device. Google said that it created Flash because developers wanted a lighter and less expensive model than Gemini Pro to build AI-powered apps and services while keeping some of the things like a long context window of one million tokens that differentiates Gemini Pro from competing models. Later this year, Google will double Gemini’s context window to two million tokens, which means that it will be able to process two hours of video, 22 hours of audio, more than 60,000 lines of code or more than 1.4 million words at the same time.

Project Astra

Google

Google showed off Project Astra, an early version of a universal assistant powered by AI that Google’s DeepMind CEO Demis Hassabis said was Google’s version of an AI agent “that can be helpful in everyday life.”

In a video that Google says was shot in a single take, an Astra user moves around Google’s London office holding up their phone and pointing the camera at various things — a speaker, some code on a whiteboard, and out a window — and has a natural conversation with the app about what it seems. In one of the video’s most impressive moments, the correctly tells the user where she left her glasses before without the user ever having brought up the glasses.

The video ends with a twist — when the user finds and wears the missing glasses, we learn that they have an onboard camera system and are capable of using Project Astra to seamlessly carry on a conversation with the user, perhaps indicating that Google might be working on a competitor to Meta’s Ray Ban smart glasses.

Ask Google Photos

Google

Google Photos was already intelligent when it came to searching for specific images or videos, but with AI, Google is taking things to the next level. If you’re a Google One subscriber in the US, you will be able to ask Google Photos a complex question like “show me the best photo from each national park I’ve visited" when the feature rolls out over the next few months. Google Photos will use GPS information as well as its own judgement of what is “best” to present you with options. You can also ask Google Photos to generate captions to post the photos to social media.

Veo and Imagen 3

Google

Google’s new AI-powered media creation engines are called Veo and Imagen 3. Veo is Google’s answer to OpenAI’s Sora. It can produce “high-quality” 1080p videos that can last “beyond a minute”, Google said, and can understand cinematic concepts like a timelapse.

Imagen 3, meanwhile, is a text-to-image generator that Google claims handles text better than its previous version, Imagen 2. The result is the company’s highest quality” text-to-image model with “incredible level of detail” for “photorealistic, lifelike images” and fewer artifacts — essentially pitting it against OpenAI’s DALLE-3.

Big updates to Google Search

Google

Google is making big changes to how Search fundamentally works. Most of the updates announced today like the ability to ask really complex questions (“Find the best yoga or pilates studios in Boston and show details on their intro offers and walking time from Beacon Hill.”) and using Search to plan meals and vacations won’t be available unless you opt in to Search Labs, the company’s platform that lets people try out experimental features.

But a big new feature that Google is calling AI Overviews and which the company has been testing for a year now, is finally rolling out to millions of people in the US. Google Search will now present AI-generated answers on top of the results by default, and the company says that it will bring the feature to more than a billion users around the world by the end of the year.

Gemini on Android

Google

Google is integrating Gemini directly into Android. When Android 15 releases later this year, Gemini will be aware of the app, image or video that you’re running, and you’ll be able to pull it up as an overlay and ask it context-specific questions. Where does that leave Google Assistant that already does this? Who knows! Google didn’t bring it up at all during today’s keynote.

WearOS 5 battery life improvements

Google isn't quite ready to roll out the latest version of it smartwatch OS, but it is promising some major battery life improvements when it comes. The company said that Wear OS 5 will consume 20 percent less power than Wear OS 4 if a user runs a marathon. Wear OS 4 already brought battery life improvements to smartwatches that support it, but it could still be a lot better at managing a device's power. Google also provided developers with a new guide on how to conserve power and battery, so that they can create more efficient apps.

Android 15 anti-theft features

Android 15's developer preview may have been rolling for months, but there are still features to come. Theft Detection Lock is a new Android 15 feature that will use AI (there it is again) to predict phone thefts and lock things up accordingly. Google says its algorithms can detect motions associated with theft, like those associated with grabbing the phone and bolting, biking or driving away. If an Android 15 handset pinpoints one of these situations, the phone’s screen will quickly lock, making it much harder for the phone snatcher to access your data.

There were a bunch of other updates too. Google said it would add digital watermarks to AI-generated video and text, make Gemini accessible in the side panel in Gmail and Docs, power a virtual AI teammate in Workspace, listen in on phone calls and detect if you’re being scammed in real time, and a lot more.

Catch up on all the news from Google I/O 2024 right here!

Update May 15, 2:45PM ET: This story was updated after being published to include details on new Android 15 and WearOS 5 announcements made following the I/O 2024 keynote.

At the end of I/O, Google’s annual developer conference at the Shoreline Amphitheater in Mountain View, Google CEO Sundar Pichai revealed that the company had said “AI” 121 times. That, essentially, was the crux of Google’s two-hour keynote — stuffing AI into every Google app and service used by more than two billion people around the world. Here are all the major updates that Google announced at the event.

Gemini 1.5 Flash and updates to Gemini 1.5 Pro

Google

Google announced a brand new AI model called Gemini 1.5 Flash, which it says is optimised for speed and efficiency. Flash sits between Gemini 1.5 Pro and Gemini 1.5 Nano, which its the company’s smallest model that runs locally on device. Google said that it created Flash because developers wanted a lighter and less expensive model than Gemini Pro to build AI-powered apps and services while keeping some of the things like a long context window of one million tokens that differentiates Gemini Pro from competing models. Later this year, Google will double Gemini’s context window to two million tokens, which means that it will be able to process two hours of video, 22 hours of audio, more than 60,000 lines of code or more than 1.4 million words at the same time.

Project Astra

Google

Google showed off Project Astra, an early version of a universal assistant powered by AI that Google’s DeepMind CEO Demis Hassabis said was Google’s version of an AI agent “that can be helpful in everyday life.”

In a video that Google says was shot in a single take, an Astra user moves around Google’s London office holding up their phone and pointing the camera at various things — a speaker, some code on a whiteboard, and out a window — and has a natural conversation with the app about what it seems. In one of the video’s most impressive moments, the correctly tells the user where she left her glasses before without the user ever having brought up the glasses.

The video ends with a twist — when the user finds and wears the missing glasses, we learn that they have an onboard camera system and are capable of using Project Astra to seamlessly carry on a conversation with the user, perhaps indicating that Google might be working on a competitor to Meta’s Ray Ban smart glasses.

Ask Google Photos

Google

Google Photos was already intelligent when it came to searching for specific images or videos, but with AI, Google is taking things to the next level. If you’re a Google One subscriber in the US, you will be able to ask Google Photos a complex question like “show me the best photo from each national park I’ve visited" when the feature rolls out over the next few months. Google Photos will use GPS information as well as its own judgement of what is “best” to present you with options. You can also ask Google Photos to generate captions to post the photos to social media.

Veo and Imagen 3

Google

Google’s new AI-powered media creation engines are called Veo and Imagen 3. Veo is Google’s answer to OpenAI’s Sora. It can produce “high-quality” 1080p videos that can last “beyond a minute”, Google said, and can understand cinematic concepts like a timelapse.

Imagen 3, meanwhile, is a text-to-image generator that Google claims handles text better than its previous version, Imagen 2. The result is the company’s highest quality” text-to-image model with “incredible level of detail” for “photorealistic, lifelike images” and fewer artifacts — essentially pitting it against OpenAI’s DALLE-3.

Big updates to Google Search

Google

Google is making big changes to how Search fundamentally works. Most of the updates announced today like the ability to ask really complex questions (“Find the best yoga or pilates studios in Boston and show details on their intro offers and walking time from Beacon Hill.”) and using Search to plan meals and vacations won’t be available unless you opt in to Search Labs, the company’s platform that lets people try out experimental features.

But a big new feature that Google is calling AI Overviews and which the company has been testing for a year now, is finally rolling out to millions of people in the US. Google Search will now present AI-generated answers on top of the results by default, and the company says that it will bring the feature to more than a billion users around the world by the end of the year.

Gemini on Android

Google

Google is integrating Gemini directly into Android. When Android 15 releases later this year, Gemini will be aware of the app, image or video that you’re running, and you’ll be able to pull it up as an overlay and ask it context-specific questions. Where does that leave Google Assistant that already does this? Who knows! Google didn’t bring it up at all during today’s keynote.

There were a bunch of other updates too. Google said it would add digital watermarks to AI-generated video and text, make Gemini accessible in the side panel in Gmail and Docs, power a virtual AI teammate in Workspace, listen in on phone calls and detect if you’re being scammed in real time, and a lot more.

Catch up on all the news from Google I/O 2024 right here!

Google is shaking up Search. On Tuesday, the company announced big new AI-powered changes to the world’s dominant search engine at I/O, Google’s annual conference for developers. With the new features, Google is positioning Search as more than a way to simply find websites. Instead, the company wants people to use its search engine to directly get answers and help them with planning events and brainstorming ideas.

“[With] generative AI, Search can do more than you ever imagined,” wrote Liz Reid, vice president and head of Google Search, in a blog post. “So you can ask whatever’s on your mind or whatever you need to get done — from researching to planning to brainstorming — and Google will take care of the legwork.”

Google’s changes to Search, the primary way that the company makes money, are a response to the explosion of generative AI ever since OpenAI’s ChatGPT released at the end of 2022. Since then, a handful of AI-powered apps and services including ChatGPT, Anthropic, Perplexity, and Microsoft’s Bing, which is powered by OpenAI’s GPT-4, have challenged Google’s flagship service by directly providing answers to questions instead of simply presenting people a list of links. This is the gap that Google is racing to bridge with its new features in Search.

Starting today, Google will show complete AI-generated answers in response to most search queries at the top of the results page in the US. Google first unveiled the feature a year ago at Google I/O in 2023, but so far, anyone who wanted to use the feature had to sign up for it as part of the company’s Search Labs platform that lets people try out upcoming features ahead of their general release. Google is now making AI Overviews available to hundreds of millions of Americans, and says that it expects it to be available in more countries to over a billion people by the end of the year. Reid wrote that people who opted to try the feature through Search Labs have used it “billions of times” so far, and said that any links included as part of the AI-generated answers get more clicks than if the page had appeared as a traditional web listing, something that publishers have been concerned about. “As we expand this experience, we’ll continue to focus on sending valuable traffic to publishers and creators,” Reid wrote. 

In addition to AI Overviews, searching for certain queries around dining and recipes, and later with movies, music, books, hotels, shopping and more in English in the US will show a new search page where results are organized using AI. “[When] you’re looking for ideas, Search will use generate AI to brainstorm with you and create an AI-organized results page that makes it easy to explore,” Reid said in the blog post.

Google

If you opt in to Search Labs, you’ll be able to access even more features powered by generative AI in Google Search. You’ll be able to get AI Overview to simplify the language or break down a complex topic in more detail. Here’s an example of a query asking Google to explain, for instance, the connection between lightning and thunder.

Google

Search Labs testers will also be able to ask Google really complex questions in a single query to get answers on a single page instead of having to do multiple searches. The example that Google’s blog post gives: “Find the best yoga or pilates studios in Boston and show details on their intro offers and walking time from Beacon Hill.” In response, Google shows the highest-rated yoga and pilates studios near Boston’s Beacon Hill neighborhood and even puts them on a map for easy navigation.

Google

Google also wants to become a meal and vacation planner by letting people who sign up for Search Labs ask queries like “create a 3 day meal plan for a group that’s easy to prepare” and letting you swap out individual results in its AI-generated plan with something else (swapping a meat-based dish in a meal plan for a vegetarian one, for instance).

Google

Finally, Google will eventually let anyone who signs up for Search Labs use a video as a search query instead of text or images. “Maybe you bought a record player at a thriftshop, but it’s not working when you turn it on and the metal piece with the needle is drifting unexpectedly,” wrote Reid in Google’s blog post. “Searching with video saves you the time and trouble of finding the right words to describe this issue, and you’ll get an AI Overview with steps and resources to troubleshoot.”

Google said that all these new capabilities are powered by a brand new Gemini model customized for Search that combines Gemini’s advanced multi-step reasoning and multimodal abilities with Google’s traditional search systems.

Catch up on all the news from Google I/O 2024 right here!

Google announced updates to its Gemini family of AI models at I/O, the company’s annual conference for developers, on Tuesday. It’s rolling out a new model called Gemini 1.5 Flash, which it says is optimized for speed and efficiency.

“[Gemini] 1.5 Flash excels at summarization, chat applications, image and video captioning, data extraction from long documents and tables, and more,” wrote Demis Hassabis, CEO of Google DeepMind, in a blog post. Hassabis added that Google created Gemini 1.5 Flash because developers needed a model that was lighter and less expensive than the Pro version, which Google announced in February. Gemini 1.5 Pro is more efficient and powerful than the company’s original Gemini model announced late last year.

Gemini 1.5 Flash sits between Gemini 1.5 Pro and Gemini 1.5 Nano, Google’s smallest model that runs locally on devices. Despite being lighter weight then Gemini Pro, however, it is just as powerful. Google said that this was achieved through a process called “distillation”, where the most essential knowledge and skills from Gemini 1.5 Pro were transferred to the smaller model. This means that Gemini 1.5 Flash will get the same multimodal capabilities of Pro, as well as its long context window – the amount of data that an AI model can ingest at once – of one million tokens. This, according to Google, means that Gemini 1.5 Flash will be capable of analyzing a 1,500-page document or a codebase with more than 30,000 lines at once. 

Gemini 1.5 Flash (or any of these models) aren’t really meant for consumers. Instead, it’s a faster and less expensive way for developers building their own AI products and services using tech designed by Google.

In addition to launching Gemini 1.5 Flash, Google is also upgrading Gemini 1.5 Pro. The company said that it had “enhanced” the model’s abilities to write code, reason and parse audio and images. But the biggest update is yet to come – Google announced it will double the model’s existing context window to two million tokens later this year. That would make it capable of processing two hours of video, 22 hours of audio, more than 60,000 lines of code or more than 1.4 million words at the same time.

Both Gemini 1.5 Flash and Pro are now available in public preview in Google’s AI Studio and Vertex AI. The company also announced today a new version of its Gemma open model, called Gemma 2. But unless you’re a developer or someone who likes to tinker around with building AI apps and services, these updates aren’t really meant for the average consumer.

Catch up on all the news from Google I/O 2024 right here!

OpenAI on Monday announced GPT-4o, a brand new AI model that that the company says is one step closer to “much more natural human-computer interaction.” The new model accepts any combination of text, audio and images as input and can generate an output in all three formats. It’s also capable of recognizing emotion, lets you interrupt it mid-speech, and responds nearly as fast as a human being during conversations.

“The special thing about GPT-4o is it beings GPT-4 level intelligence to everyone, including our free users,” said OpenAI CTO Mira Murati during a live-streamed presentation. “This is the first time we’re making a huge step forward when it comes to ease of use.”

During the presentation, OpenAI showed off GPT-4o translating live between English and Italian, helping a researcher solve a linear equation in real time on paper, and providing guidance on deep breathing to another OpenAI executive simply by listening to his breaths.

The “o” in GPT-4o stands for “omni”, a reference to the model’s multimodal capabilities. OpenAI said that GPT-4o was trained across text, vision and audio, which means that all inputs and outputs are processed by the same neural network. This is different from the company’s previous models, GPT-3.5 and GPT-4, which did let users ask questions simply by speaking, but then transcribing the speech into text. This stripped out tone and emotion and made interactions slower.

OpenAI is making the new model available to everyone, including free ChatGPT users, over the next few weeks and also releasing a desktop version of ChatGPT, initially for the Mac, which paid users will have access to starting today.

OpenAI’s announcement comes a day before Google I/O, the company’s annual developer conference. Shortly after OpenAI revealed GPT-4o, Google teased a version of Gemini, its own AI chatbot, with similar capabilties.

The Financial Times has become the latest news organization to strike a deal with OpenAI. In a joint announcement on Monday, the Financial Times and OpenAI said that maker of ChatGPT will use the Financial Times’ journalism to train its AI models and collaborate on developing new AI products and features for the publication’s readers. ChatGPT will also attribute and and link back to the Financial Times when it includes information from the publication in its responses

“It is right, of course, that AI platforms pay publishers for the use of their material,” said Financial Times CEO John Ridding in a statement and added that the Times is “committed to human journalism.” Neither company disclosed the financial terms of the agreement. Earlier this year, The Information reported that OpenAI offers publishers between $1 million and $5 million a year to license their content to train its AI models.

Generative AI is only as good as the training data used to train the models that power it. So far, AI companies have scraped everything they can from the public internet often without the consent of creators, and are constantly on the hunt for new data sources to keep the outputs generated by these models current. Training AI models on news is one way to achieve that, but some publishers are wary of giving up their content to AI companies for free. The New York Times and the BBC, for instance, have OpenAI from scraping their websites.

As a result, OpenAI has been striking financial deals with leading publishers to keep its models trained. Last year, the company partnered with German publisher Axel Springer to train its models on new from Politico and Business Insider in the US and Bild and Die Welt in Germany. The company also has deals with the Associated Press, France’s Le Monde, and Spain’s Prisa Media.

Subscribing to the Financial Times costs at least $39 a month for. But, as some pointed out, its partnership with OpenAI effectively means a dismantling of its own paywall for general readers through generative AI.

Apple has resumed conversations with OpenAI, the maker of ChatGPT, to power some AI features coming to iOS 18, according to a new report in Bloomberg. Apple is also building its own large language models to power some iOS 18 features, but its talks with OpenAI are centered around a “chatbot/search component,” according to Bloomberg reporter Mark Gurman. 

Apple is also reportedly in talks with Google to license Gemini, Google’s own AI-powered chatbot, for iOS 18. Bloomberg reports that those talks are still on, and things could still go either way because Apple hasn’t made a final decision on which company’s technology to use. It’s conceivable, Gurman says, that Apple could ultimately end up licensing AI tech from both companies or none of them.

So far, Apple has been notably quiet about its AI efforts even as the rest of Silicon Valley has descended into an AI arms race. But it has dropped enough hints to indicate that it’s cooking up something. When the company announced its earnings in February, CEO Tim Cook said that Apple is continuing to work and invest in artificial intelligence and is “excited to share the details of our ongoing work in that space later this year.” It claimed that the brand new M3 MacBook Air that it launched last month was the “world’s best consumer laptop for AI,” and will reportedly start releasing AI-centric laptops and desktops later this year. And earlier this week, Apple also released a handful of open-source large language models that are designed to run locally on devices rather than in the cloud.

It’s still unclear what Apple’s AI features in iPhones and other devices will look like. Generative AI is still notoriously unreliable and prone to making up answers. Recent AI-powered gadgets like the Humane Ai Pin released to disastrous reviews, while others like the Rabbit R1 have yet to prove themselves valuable.

We’ll find out more at WWDC on June 10.

Leading artificial intelligence companies including OpenAI, Microsoft, Google, Meta and others have jointly pledged to prevent their AI tools from being used to exploit children and generate child sexual abuse material (CSAM). The initiative was led by child-safety group Thorn and All Tech Is Human, a non-profit focused on responsible tech.

The pledges from AI companies, Thorn said, “set a groundbreaking precedent for the industry and represent a significant leap in efforts to defend children from sexual abuse as a feature with generative AI unfolds.” The goal of the initiative is to prevent the creation of sexually explicit material involving children and take it off social media platforms and search engines. More than 104 million files of suspected child sexual abuse material were reported in the US in 2023 alone, Thorn says. In the absence of collective action, generative AI is poised to make this problem worse and overwhelm law enforcement agencies that are already struggling to identify genuine victims.

On Tuesday, Thorn and All Tech Is Human released a new paper titled “Safety by Design for Generative AI: Preventing Child Sexual Abuse” that outlines strategies and lays out recommendations for companies that build AI tools, search engines, social media platforms, hosting companies and developers to take steps to prevent generative AI from being used to harm children.

One of the recommendations, for instance, asks companies to choose data sets used to train AI models carefully and avoid ones only only containing instances of CSAM but also adult sexual content altogether because of generative AI’s propensity to combine the two concepts. Thorn is also asking social media platforms and search engines to remove links to websites and apps that let people “nudity” images of children, thus creating new AI-generated child sexual abuse material online. A flood of AI-generated CSAM, according to the paper, will make identifying genuine victims of child sexual abuse more difficult by increasing the “haystack problem” — an reference to the amount of content that law enforcement agencies must current sift through.

“This project was intended to make abundantly clear that you don’t need to throw up your hands,” Thorn’s vice president of data science Rebecca Portnoff told the Wall Street Journal. “We want to be able to change the course of this technology to where the existing harms of this technology get cut off at the knees.”

Some companies, Portnoff said, had already agreed to separate images, video and audio that involved children from data sets containing adult content to prevent their models from combining the two. Others also add watermarks to identify AI-generated content, but the method isn’t foolproof — watermarks and metadata can be easily removed.