برچسب: AI

Along with the brand new Siri AI, Apple is introducing a number of new Apple Intelligence features—including a trio of new AI tools in the Photos app. We’ve been here before: Apple previously released AI-powered image editing features like Clean Up, which didn’t necessarily hit the mark compared to similar tools from competitors like Google or Samsung. But this year appears to be a bit different: Apple’s newest models, including those that work off-device, are improving existing features and powering new tools. For the most part, it seems to be a step in the right direction.

Apple’s Clean Up tool is much better

The new Clean Up tool is perhaps the most important update here. In iOS 26, Clean Up used Apple’s on-device AI models to remove objects, but it was hit-or-miss. Clean Up was okay at basic tasks, but I found it couldn’t remove surrounding shadows, nor could it replace an object with something that looked like it was originally part of the image.

Clean Up now uses a hybrid approach. For simple tweaks, like removing a small object, it uses an on-device model, just like in iOS 26. But, for bigger, more complex tasks (like removing an obstruction around your face), it hands off the task to Apple’s powerful Foundation models hosted on Apple’s own Private Cloud Compute servers. These servers, according to Apple, are completely private and encrypted. Apple says it doesn’t have access to your photos, and it doesn’t use your data for training.

To find these new tools, tap Edit on a photo, then choose Tools at the end of the toolbar. Here, tap Clean Up. By default, the feature is in Auto mode, which is the hybrid approach discussed above. From here, you can also switch to High Quality to force Apple to use the cloud models.

Middle: Cleaning up using only the on-device Fast model (same as iOS 26). Right: Using Apple’s new Cloud models in iOS 27.
Credit: Khamosh Pathak

Then, it’s business as usual. Use your fingers to highlight the object or part of the image that you want to remove. If you’re using Fast, the on-device option, the cleanup process will begin instantly. If you’re using High Quality, you’ll need to tap Clean Up and wait until Apple’s models do their thing. In my experience, the wait time can even stretch to minutes if you’re asking for clear, large objects.

After using this feature for cleaning up multiple images, here’s the best tip I can give you: always use High Quality. Fast is the same as last year’s feature, and while it removes the image, its replacement is lacking, as you can see with the mismatched tabletop in the image above. Even if you are removing a distinct object from a table, High Quality does a better job of replicating the tabletop, as well as shadows falling from other objects.

Left and Middle: Clean up tool using on-device AI on iPhone 16 Pro. Right: Clean Up tool using Apple’s cloud models on iPhone 16 Pro.
Credit: Khamosh Pathak

The improvements continue when dealing with faces. The new iOS 27 feature can use generative AI and your own photos to recreate parts of your face that are obstructed. In my test (which you can see above), Clean Up on iOS 27 got rid of 99% of my coffee mug (though a border somehow still remains). On iOS 26, though, the result is just laughably bad: a soup of surrounding colors.

Extending photos in iOS 27 works like a charm

Credit: Khamosh Pathak

Extend, as the name suggests, expands your photos. Let’s say you have an off-center shot, or just looks unbalanced. Tap Extend from the Tools menu, then pinch in and move the image around. As you do, the surroundings will begin to blur, indicating the areas that iOS will fill in using Apple’s generative AI models. Because Apple uses cloud models, this too might take some time. Tap Extend, and wait.

Overall, Apple’s generative AI for extending images and filling in details is quite good—with some limitations. That’s not necessarily surprising, as it’s trained on Gemini’s own models, which are excellent at image manipulation. I tested the feature by extending the frame in nature, and in indoor settings. It did a good job of guessing what was around me, and even gave me a hand and a leg that weren’t in the original shot. That said, it completely overexposed the image, so while you get more in the frame, you lose the sky entirely. In a photo I took of a coffee shop, the Photos app took the coffee bags that were on the shelf and just repeated them in the extended shot. This is quite a smart way to make the extended image look more realistic.

I took this image at Bookatico Bookstore & Cafe in Vadodara, India.
Credit: Khamosh Pathak

Apple’s Reframe feature needs a bit of work

The Reframe tool lets you change the angle or the perspective of the photo. You might wish you had moved your phone just a bit to the right before taking that snap of your partner, and while the moment is gone, the angle might be saved. With Reframe, you can swipe around on the image to change the perspective, as if you were adjusting the angle when originally taking the photo. The app shows you a live preview of what things will look like (as this is just a preview, it will show some unnatural bending, but that won’t be in the final result). Then, tap Reframe, and let Apple’s cloud models do their thing. After some time, the reframed image will be ready.

Credit: Khamosh Pathak

In my testing, I found this to be the most jarring tool. It does the job, but it struggles with faces quite a lot. It’s best to use it for slight angle changes, and not much else. To stress test, I pushed the angle as far as the Photos app would allow. The result was an image with a slanted face, that looked more 2D than 3D (I will save you the horror of looking at my face with the eyes scrambled). As Apple works on improving the cloud models, though, this can get better.

Remember: All of these features are currently in beta testing. Apple may continue to improve the experience with subsequent betas and with iOS 27’s official release in the fall.

Google’s efforts to transform online search and the ways we find information on the web continue, and there’s now a new feature available in AI Mode: information agents. This new feature will keep tabs on news for you and ping you when needed. (At least, that’s the idea.)

The update was announced at Google I/O 2026 back in May, but it’s now appearing for those with a Google AI Ultra subscription (that’s the $99.99 or $199.99 per month one). Google’s Robby Stein says that more people will be getting access in the summer.

Google hasn’t been any more specific than that, but presumably this will filter down to the other paid plans in the next few months, and eventually to everyone. I’m making use of my Ultra subscription here to get these information agents running.

This Tweet is currently unavailable. It might be loading or has been removed.

Here’s the idea: You look up, say, news on the next James Bond movie. Then AI Mode can keep tabs on the web and ping you whenever there’s news on a casting decision, a release date, or a trailer. It saves you having to run multiple searches because the information comes to you (it’s a bit like the old Google Alerts, if you remember those). These updates appear both in the AI Mode section of Google search and in the Google app on your phone. Or at least they’re supposed to.

How my information agents worked (or didn’t)

To set up your information agents, you can head to Google search on the web, then switch to AI Mode via the button in the search box. You then launch your search like you would if you were chatting to Gemini—something like “tell me who the main stars are in the upcoming Christopher Nolan film The Odyssey.”

After you’ve got your answer, which is hopefully hallucination-free, you might get asked at the end if you want to set up an information agent to keep you updated on whatever it is you’ve searched for. If you don’t get asked, you can put in the request anyway: Tell Google to “keep you updated” or words along those lines.

After some cajoling, I got my World Cup news.
Credit: Lifehacker

You’ll receive a confirmation message, and then whenever new information appears online, you’ll get pinged about it. Updates appear as notifications in the Google app on mobile, and as new entries in the original AI Mode conversation—so if you delete the chat, you stop getting updates (you can access your previous chats via the AI Mode history button on the left.

Google doesn’t say how often you’ll get updates, but in my experience it was… not at all. I set up an information agent to keep me updated on the latest World Cup scores and group standings, which I thought was a fairly straightforward task, but my Google app stayed stubbornly silent and the AI Mode chats stayed static over a whole day as the goals went in and the matches went by.

Looking around the web, it seems that AI Mode information agents are working for other people, so this appears to be an isolated bug that may well get ironed out in a few days—but it’s frustrating to have these features launch and then not work properly. I’ve had the same experience with Gemini Omni too, so maybe it’s just me.

Thank you, AI Mode, for my updates.
Credit: Lifehacker

What did work was telling the information agents to give me a daily summary at a specific time. When I did this, I did indeed get a delivery of the latest World Cup news and everything that had happened in the tournament over the past 24 hours, both in the Google app and as an update to my conversation with AI Mode.

In fact, this might be a better way to use this rather than expecting updates at random times throughout the day. I can definitely see myself using it for big news topics I’m interested in, and to pick up stories I might otherwise miss, though it doesn’t seem to be set up for important breaking news.

Meta is rolling out a new AI search mode on Facebook that will synthesize content from public posts—so instead of a list of links, users will get a summarized response similar to AI-generated results on other platforms. The feature, powered by Meta AI, will also allow users to engage in ongoing conversations and ask follow-up questions in plain language based on the results.

According to Meta’s post announcing the new search function, AI Mode provides “answers grounded in what people are saying publicly across our apps” using information pulled from across its platforms, such as Groups and Reels. As TechCrunch notes, this feature functions similarly to the AI-powered “Ask” tab found in Meta’s recently launched Forum app, which allows users to obtain answers to queries from across groups.

You should still vet Meta’s AI search responses

While it can be useful to glean information from user-generated content containing personal experiences, you should also use Meta AI (and tools like it) with caution. Obviously, AI responses should already be subject to scrutiny, as they often contain incorrect information and hallucinations even when pulled from vetted source material. As reported by 404 Media, it is also surprisingly easy to manipulate AI search results via user-generated content on Reddit and Wikipedia. In fact, nearly a quarter of all citations used by AI tools like Google AI and ChatGPT come from sites like these.

Many posts on Meta platforms contain misinformation and spam, and, like many social platforms, Facebook uses community notes instead of third-party fact-checking. Plus, user-generated content isn’t the most reliable source. At best, the data AI searches pull from may be outdated—for example, a restaurant recommended in a post about travel has actually closed. At worst, the information may be intentionally misleading or malicious.

Among all the other artificial intelligence upgrades Apple is rolling out for us this year, you’ll find that there’s a significant jump forward in Image Playground’s AI image generation abilities. Before now, the app’s outputs were rather limited in terms of size, style, and possible prompts. It was very much AI images for beginners, with the results mostly basic and generic.

With iOS 27, iPadOS 27, and macOS 27, that’s changing. You can try the developer betas now (though you probably shouldn’t; the public betas are arriving in July), and Apple says the full releases will appear later in the year (most likely around September). Image generation is something both ChatGPT and Gemini have been steadily improving, to the point where some generations are difficult to distinguish from real photos. So how does Apple’s new and improved Image Playground compare?

Here’s what Image Playground can do now

Open up Image Playground on the newest versions of iOS, iPadOS, or macOS, and you’ll see there are several new capabilities. First, you can ask for photorealistic images in the prompt box, as well as the sketches and illustrations previously possible: Ask for a photo of an English meadow or a towering temple, and Image Playground will oblige.

There’s more control over your images, too. You can submit a photo as part of a starting prompt, and you can choose between square, portrait, and landscape orientations for your picture—these options weren’t available before. If you want to transport your pet dog to the jungle, that’s possible now.

Image Playground can now work with existing images.
Credit: Lifehacker

Then there are clever editing tools, similar to those in Google’s Nano Banana AI model, that let you change specific parts of a generated image without having to render it entirely again. You can use a follow-up prompt to request changes, and even highlight the part of the picture you want to edit.

You can now change the color of objects, remove objects altogether, change the weather of a scene, whatever you can think of, really. Based on my testing, it all works well, and produces results that look impressive and consistent. I managed to put a cartoon-style cat in a street, and then change the cat’s color without affecting the background.

You can use follow-up prompts to make edits.
Credit: Lifehacker

Apple says that everything created through Image Playground will have the same SynthID watermark as generations from Google Gemini and ChatGPT, and we know because of Private Cloud Compute that no user images will be stored or accessed by Apple, or used to train any of its models.

Finally, you can do more with these images too—setting them as Contact Posters or wallpaper for the lock screen, for example. You can find more AI image tools in the Photos app, where you can apply edits similar to those possible in Image Playground to whatever’s in your photo library.

Image Playground versus the competition

Image Playground is much better—no doubt thanks to a boost from Gemini—but even still, the images aren’t quite up to the very high bar that Gemini and ChatGPT have set now. You can see below how my request for “a photorealistic image of a small, ancient-looking spaceship floating between the stars, with an Earth-like planet behind it” was interpreted by Image Playground (left), Gemini (center), and ChatGPT (right).

The spaceship AI challenge.
Credit: Image Playground / Gemini / ChatGPT

Apple does okay, but to my eye, Gemini and ChatGPT generated results that are more immersive and detailed—like something you’d see in an actual science fiction film. There’s more detail and more imagination, although to Apple’s credit, Image Playground rendered the fastest.

For the next challenge, I tried asking these AI tools to “move my cuddly toy from my floor to a pebbly beach, with the tide lapping at its edges.” Again, all these attempts are good, but Gemini (center) and ChatGPT (right) add extra layers of verisimilitude in terms of color, angle, and texture (though Gemini seems to have created two shorelines).

The toy on the beach challenge.
Credit: Image Playground / Gemini / ChatGPT

I asked all three models to remove the toy and just leave the beach, and they all managed it more or less perfectly. These are Photoshop-level edits that used to take me hours, but can now be completed in seconds. It’s truly impressive. Gemini and ChatGPT still have the edge in terms of quality, but Image Playground comes built into billions of Apple devices. It’s now good enough that many users will likely choose not to switch to something else when they need to generate an image with AI, which might make all the difference.

Apple spent a lot of time talking about the upgraded Apple Intelligence platform and the new Siri AI app at WWDC 2026, and in the days since, a few more details have emerged about how the AI model partnership between Apple and Google will affect the new software—but answering the question of how much of Siri AI is Apple, and how much is Google, is still complicated.

Back in January, we got official news that Apple would be tapping into Google’s Gemini AI models to help power Apple Intelligence, that the deal would last multiple years, and that Apple’s “industry-leading privacy standards” would be maintained.

Neither Apple nor Google explained much at the time about how this partnership would actually play out, but it was clear that this was more significant than Apple’s earlier ChatGPT deal, where Siri simply shunted off prompts it couldn’t reply to.

I expect there was plenty of debate within Apple about whether a technological deal with an arch-rival was worth it, even if it meant catching up more quickly with its AI. Ultimately, CEO Tim Cook and his fellow executives decided that it was—and after WWDC 2026, we have more information on the details.

Siri AI is not the Gemini app…

Over to the WWDC 2026 keynote, where Apple’s senior vice president (SVP) of software engineering Craig Federighi told us the vastly improved Apple Foundational Models (AFM) had been developed through a “deep collaboration” with Google. Apple had been “leveraging” the technology behind the Gemini models, in Federighi’s words, to create the AI that now powers Siri AI and the other new Apple Intelligence features.

And you can certainly see the Gemini influence: Apple’s AI is now truly multi-modal, capable of processing audio, voice, and text, and much better at producing text and images of its own. Image editing is much improved—very similar to Nano Banana 2, you might say—and Apple’s AI now has much better world knowledge too, which is another area where the Gemini models excel.

However, these are still ultimately Apple’s own models. For local models, we have the on-device AFM 3 Core and AFM 3 Core Advanced models, which sit on iPhones, iPads, and Macs—though the latest hardware is needed for for the Advanced version (the one that allows you to tweak Siri’s pace and expressiveness). Per Apple, AFM Core Advanced needs an iPhone Air, iPhone 17 Pro or Pro Max, an M4 iPad or later with at least 12GB of RAM, or an M3 Mac or later with 12GB of RAM.

Two models in one.
Credit: Apple

In follow-up comments (via 9to5Mac), Federighi said, “we don’t have the Gemini app as our app.” That is, Apple Intelligence doesn’t use the Gemini AI models, or the client code Google Gemini app users get, or a knowledge base built from Google Search. All that work has been done by Apple.

There’s no doubt Apple needed the Gemini AI models to get its own models up to par this quickly, but Apple executives are understandably keen to not make too much of the partnership, for the same reasons that they won’t talk about the billions of dollars Google pays each year to remain the default search option in Safari.

Federighi still had time for some barbed comments though: “Some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people—all of us—that it’s ultimately meant to serve,” he said, in a dig at the competitors that have left his company in the dust on AI (while Apple has paid out $250 million in settlements for promising AI features that never showed up).

…but Apple is using some Google servers

That covers the local Apple Foundation Models, which keep all of your queries and data private and protected on whatever device you’re using. It’s with the cloud-based models that the waters get a bit muddier—these are the models that are called in to deal with larger, more complex tasks that can’t be handled solely on-device.

As described by the Apple team (via Ars Technica), the AFM 3 Cloud is for general-purpose use. Then there’s ADM 3 Cloud for image generation, and AFM 3 Cloud Pro for “more sophisticated” queries (and, it sounds like, the beginnings of agentic work). Like the on-device models, they use some Gemini smarts at the most basic levels, but with Apple’s own contributions and tweaks on top.

Those first two models run on Apple servers, but queries sent to AFM 3 Cloud Pro are going somewhere else: They’ll be sent to Google data centers, to be processed by Nvidia GPUs. However, according to Apple, the exact same Private Cloud Compute (PCC) protections will be in place for those data centers as for those run by Apple.

The new Apple Foundation Models are vastly improved.
Credit: Apple

That means no data is stored (it’ll be wiped after the query is processed), no one else can see it (not even Apple or Google), and your identity is masked. Apple lets third-party security auditors check its PCC code, and that’s going to be the case with the AFM 3 Cloud Pro models and Google’s servers, too.

There is one small wrinkle: Apple says “PCC on Google Cloud will be gradually ramping towards the complete set of protections throughout the summer preview period.” So if you’re running one of the developer betas, some of your most complex AI queries might not yet be as fully protected as you would like them to be.

Apple is promising more technical detail on all of this as we get further towards the full launch of the new software updates and Siri AI, but based on the information we have now, it seems to have managed a balance between boosting its AI with Gemini while retaining all of the Apple-ness that its users are going to expect.

Apple spent a lot of time talking about the upgraded Apple Intelligence platform and the new Siri AI app at WWDC 2026, and in the days since, a few more details have emerged about how the AI model partnership between Apple and Google will affect the new software—but answering the question of how much of Siri AI is Apple, and how much is Google, is still complicated.

Back in January, we got official news that Apple would be tapping into Google’s Gemini AI models to help power Apple Intelligence, that the deal would last multiple years, and that Apple’s “industry-leading privacy standards” would be maintained.

Neither Apple nor Google explained much at the time about how this partnership would actually play out, but it was clear that this was more significant than Apple’s earlier ChatGPT deal, where Siri simply shunted off prompts it couldn’t reply to.

I expect there was plenty of debate within Apple about whether a technological deal with an arch-rival was worth it, even if it meant catching up more quickly with its AI. Ultimately, CEO Tim Cook and his fellow executives decided that it was—and after WWDC 2026, we have more information on the details.

Siri AI is not the Gemini app…

Over to the WWDC 2026 keynote, where Apple’s senior vice president (SVP) of software engineering Craig Federighi told us the vastly improved Apple Foundational Models (AFM) had been developed through a “deep collaboration” with Google. Apple had been “leveraging” the technology behind the Gemini models, in Federighi’s words, to create the AI that now powers Siri AI and the other new Apple Intelligence features.

And you can certainly see the Gemini influence: Apple’s AI is now truly multi-modal, capable of processing audio, voice, and text, and much better at producing text and images of its own. Image editing is much improved—very similar to Nano Banana 2, you might say—and Apple’s AI now has much better world knowledge too, which is another area where the Gemini models excel.

However, these are still ultimately Apple’s own models. For local models, we have the on-device AFM 3 Core and AFM 3 Core Advanced models, which sit on iPhones, iPads, and Macs—though the latest hardware is needed for for the Advanced version (the one that allows you to tweak Siri’s pace and expressiveness). Per Apple, AFM Core Advanced needs an iPhone Air, iPhone 17 Pro or Pro Max, an M4 iPad or later with at least 12GB of RAM, or an M3 Mac or later with 12GB of RAM.

Two models in one.
Credit: Apple

In follow-up comments (via 9to5Mac), Federighi said, “we don’t have the Gemini app as our app.” That is, Apple Intelligence doesn’t use the Gemini AI models, or the client code Google Gemini app users get, or a knowledge base built from Google Search. All that work has been done by Apple.

There’s no doubt Apple needed the Gemini AI models to get its own models up to par this quickly, but Apple executives are understandably keen to not make too much of the partnership, for the same reasons that they won’t talk about the billions of dollars Google pays each year to remain the default search option in Safari.

Federighi still had time for some barbed comments though: “Some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people—all of us—that it’s ultimately meant to serve,” he said, in a dig at the competitors that have left his company in the dust on AI (while Apple has paid out $250 million in settlements for promising AI features that never showed up).

…but Apple is using some Google servers

That covers the local Apple Foundation Models, which keep all of your queries and data private and protected on whatever device you’re using. It’s with the cloud-based models that the waters get a bit muddier—these are the models that are called in to deal with larger, more complex tasks that can’t be handled solely on-device.

As described by the Apple team (via Ars Technica), the AFM 3 Cloud is for general-purpose use. Then there’s ADM 3 Cloud for image generation, and AFM 3 Cloud Pro for “more sophisticated” queries (and, it sounds like, the beginnings of agentic work). Like the on-device models, they use some Gemini smarts at the most basic levels, but with Apple’s own contributions and tweaks on top.

Those first two models run on Apple servers, but queries sent to AFM 3 Cloud Pro are going somewhere else: They’ll be sent to Google data centers, to be processed by Nvidia GPUs. However, according to Apple, the exact same Private Cloud Compute (PCC) protections will be in place for those data centers as for those run by Apple.

The new Apple Foundation Models are vastly improved.
Credit: Apple

That means no data is stored (it’ll be wiped after the query is processed), no one else can see it (not even Apple or Google), and your identity is masked. Apple lets third-party security auditors check its PCC code, and that’s going to be the case with the AFM 3 Cloud Pro models and Google’s servers, too.

There is one small wrinkle: Apple says “PCC on Google Cloud will be gradually ramping towards the complete set of protections throughout the summer preview period.” So if you’re running one of the developer betas, some of your most complex AI queries might not yet be as fully protected as you would like them to be.

Apple is promising more technical detail on all of this as we get further towards the full launch of the new software updates and Siri AI, but based on the information we have now, it seems to have managed a balance between boosting its AI with Gemini while retaining all of the Apple-ness that its users are going to expect.

AI is everywhere right now—even in places you don’t expect. You might be jamming to a new song on Spotify or YouTube, only to later learn that the track was “composed” entirely by bots (save for an initial human-generated prompt). Some might argue that AI music has its place, but if you’re like me, you want to devote your attention to art created by real people, who have taken the time to hone a craft and share it with the world. And while I believe AI music can never replace that, the fact is, it’s getting more difficult to identify these tunes when we come across them in the wild.

To be fair, some companies have been working on ways to identify AI content on their platforms, notably Spotify, YouTube, and Apple Music. But while you might come across an AI label here and there, there are still plenty of examples of AI-generated content that aren’t identified as such—in part, because much of the reporting is still based on the honor system. Lifehacker’s David Nield was frustrated by that experience when looking for music to listen to on YouTube, and found the only reliable solution was to do some research before committing to any particular channel. He now has a shortlist of options to choose from, sure, but it’s a lot of work to guarantee your music selection is 100% human-made. It also makes it harder for small creators, who might not have as much “proof” that they aren’t using AI, if you’re this strict with your consumption.

How to use Deezer’s new AI detector

Deezer, a French music streaming service, thinks it has a solution. As reported by MacRumors, the platform now has a new tool it says can identify AI-generated music with nearly 100% accuracy. The company says that it receives over 75,000 AI-generated songs every day, which amounts to 44% of the total uploads to the platform. By looking for artifacts left behind by AI, Deezer claims it can spot whether the track was made by humans or bots. In fact, it seems that the tool is the same Deezer uses to label AI tracks on its own platform.

Of course, this tool works with Deezer itself, so if you’re a user, you already have access. But the company says its AI detector works with as many as 20 different streaming services. That includes the following:

• Spotify

• Apple Music

• YouTube

• YouTube Music

• Tidal

• Amazon Music

• Soundcloud

• Yandex Music

• Qobuz

• Beatport

• iTunes

• Napster

• Pandora

• Anghami

• KKBOX

• Last.fm

• Soundmachine

• Boomplay

• Audiomack

In order to use the tool, you need to connect it to your streaming service of choice. That might rub privacy-minded users the wrong way since you need to give a third-party tool access to your streaming service, but if you’re okay with Deezer accessing your Apple Music or Spotify libraries, you can take advantage of the detection software. Alternatively, you can manually connect Deezer’s detector to playlists if you have the link (but you can’t upload individual tracks). Once you connect Deezer to your platform, it imports your various playlists and looks for any music it thinks was made using AI.

Deezer claims its tool is 99.8% accurate and misses two out of every 1,000 tracks. There’s no real way to test those stats, however, so take them with a grain of salt. I also wish the detector was a bit more flexible. I’d love not to have to connect my entire streaming service to use it, and I would like to have the option to test more than just individual playlists. I think an AI detector would be most useful on a case-by-case basis, rather than when questioning whether part of your playlist contains AI-generated tracks. Still, a tool like this may be a powerful ally in the battle to listen to human-made music—or, at the very least, know for certain that the song you’re enjoying was generated with AI.

Back in April, Anthropic introduced its “Mythos” model to the world. Mythos Preview, reportedly, is such a powerful model that it can find security flaws across all kinds of software. In the wrong hands, bad actors could abuse the model to find vulnerabilities in programs, services, and sites most of us rely on for modern digital life. In effect, Mythos could open up the biggest hacking opportunity in history. What a pitch.

As such, Anthropic pulled the brakes on Mythos. While it maintained that it would eventually release the model to the public, it first needed to trial it with a limited pool of trusted testers, in what it calls “Project Glasswing.” To start, that meant opening up the model to the U.S. and other governments. While Mythos is still not available to the likes of you or me, Anthropic is releasing a new model that promises many of the capabilities of Mythos, without the accompanying cybersecurity risks.

What are Anthropic’s Fable 5 and Mythos 5?

On Tuesday, Anthropic announced its latest model, Claude Fable 5, which it calls a “Mythos-class model” that is “safe for general use.” The company says Fable 5 is supposedly better and more capable than any of its other public models. Anthropic claims Fable 5 scores at the top of most benchmarks, including software engineering, knowledge work, vision tasks, and research. The company goes so far as to say “the longer and more complex the task, the larger Fable 5’s lead over our other models.” There’s also Mythos 5, which seems to be Fable 5 without certain limitations, but isn’t available to the general public.

According to Anthropic’s benchmarking, Fable 5 and Mythos 5 alike outperform Mythos Preview, Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro, in the following categories: agentic coding, knowledge work, spatial reasoning, tool use, legal, multidisciplinary reasoning (without tools), biology, cybersecurity, and health. Mythos Preview ekes out a win in computer use and multidisciplinary reasoning (with tools), but it’s a clean sweep over all other models.

Credit: Anthropic

Anthropic says Fable 5 was able to complete a coding project that would have taken a team over two months to finish in just a day. It can rebuild a web app’s source code from only screenshots. It can beat Pokémon FireRed with a “minimal, vision-only harness,” while other Claude models struggled to play at all. It was able to play Slay the Spire and reached the final act three times more often than Opus 4.8 Mythos 5 builds on its research abilities, with improved stats in drug design, as well as novel hypotheses regarding questions of molecular biology, and the ability to produce novel research in genomics.

How is Anthropic keeping Fable 5 safe?

That’s the big question: If Fable 5 is Mythos-class, how can you ensure that it’s safe to release to the general public? Couldn’t a bad actor take advantage of Fable 5’s capabilities and force it to discover and disclose security vulnerabilities?

Anthropic says it has that figured out. While Fable 5 may be Mythos-level in many ways, the company says that its Project Glasswing testing has produced a model with the proper safeguards for a public release. Fable 5 looks out for “classifiers,” or highly sensitive topics, that it knows it should not answer. What that means is this: When Fable 5 receives a request that it thinks has to do with cybersecurity, biology, chemistry, or distillation, it doesn’t answer the question itself. Instead, it passes the query off to Opus 4.8, Anthropic’s “next-most-capable” model. The model should still be powerful enough to provide accurate answers, but not capable of providing malicious users with the tools necessary to exploit others.

Anthropic says its new guardrails are cautious and conservative, and may be overkill. Benign requests may accidentally trip Fable 5’s security alarms, but that supposedly happens around 5% of the time. As such, Anthropic says Fable 5 is able to handle requests itself roughly 95% of the time. In addition, the company found that after a bug bounty program, no white hat hacker could find a universal jailbreak (or an exploit to bypass safety protocols) after 1,000 hours of testing. While one organization has made progress in finding one jailbreak, Anthropic says it’s confident that its protocols make it impractical for hackers to discover jailbreaks before the company does.

Why drop requests for biology and chemistry? Anthropic says that Mythos is also too good at aiding gene therapy research and development, which can be a benefit to scientists, but a major risk in the wrong hands. In addition, Anthropic knows that there are actors out there trying to “distill” Claude models’ abilities to train their own models to do whatever they want. As such, any of these requests is booted to a lower-performing model.

Anthropic is also making a change to its data retention policy for Fable 5 and Mythos 5. With these models, the company will keep your data for 30 days—not for training, but to help protect against future cyberattacks and jailbreaks. Fable 5 and Mythos 5 are both priced the same: $10 per million input tokens, and $50 per million output tokens, which Anthropic says is less than half the price of Mythos Preview.

Much of my day is spent sitting at a computer writing, and in recent years I’ve most often accompanied this by long (or live) YouTube videos that offer background sound without being too distracting. There are all kinds of options: scenic railway journeys, TV show tunes, piano instrumentals of songs I like, the sounds of forest rain, movie soundtracks, walks across game worlds, and more.

In recent months though, AI-generated mixes with AI-generated thumbnails have become much more prevalent. Run a quick search for study and chill-out music and you’ll find plenty of videos where the artwork looks suspiciously like something ChatGPT would make and the audio track is what you’d expect to get out of an app such as Suno.

It’s harder than ever to spot AI-made content, especially when it comes to simpler, more minimal creations—like illustration-style images or lo-fi chill-out music. I don’t want to listen to AI music, so at the start of each day I’m now clicking around warily on YouTube trying to find something that has been composed and packaged by actual people. It’s not easy anymore, but it’s still possible.

The problems with AI music

AI content strikes again
Credit: Lifehacker

I’m not completely against the idea of AI, though I think there are some major problems with it that we’re not properly reckoning with. Gemini AI might give you a better search result for “the best restaurants in San Francisco for young kids” than a list of 10 blue links, but it still relies on human experience and writing. The AI has never had kids or been to San Francisco, so what happens to those results when actual people stop writing and publishing on the internet?

When it comes to music, I don’t want to listen to tunes put together by machines, based on algorithms and the mashing together of real work done by real artists. You could argue that it doesn’t really matter so much for background electronica that’s being put on while working or studying, but the principle is the same.

There’s a line in the Westworld TV show where one humanoid robot, virtually indistinguishable from a real person, asks the question: “If you can’t tell, does it matter?” We’re now at the stage where we often can’t tell the difference between AI and human content, but I’d argue that the difference does still matter—and matters a lot.

Aside from all the considerations about energy use, environmental damage, and copyright infringement that come along with AI (and which would all take an entire article to cover), I think there are numerous ways that the tech can be helpful. When it comes to art and music, however, I want my clicks and listening time to support actual artists.

It’s something that YouTube is aware of. On some videos you’ll now see a How this content was made section, disclosing the use of AI. The problem is, this relies on either the content creator owning up to it, YouTube’s own AI tools being used, or AI watermarks being included in the files. Based on what I see on the platform, I don’t think much of the AI content is being flagged.

Finding music made by real artists

Welcome to Coulou’s Vinyl Cafe
Credit: Coulou / YouTube

So I’m left in a situation where I’d rather not listen to channels where the artwork or the music is AI-generated, but it’s difficult to spot something that’s made by AI. What I’ve started doing is looking for the channels that are definitely curated and produced by human beings, rather than trying to identify subtle signs of AI.

You’ll actually see it a lot in channel titles and video descriptions now, so you could just search for “no AI” or “AI free.” It’s also worth digging into the descriptions to look for links to the actual music used and the artists who are being supported. Check up on the history of the channel, too—what other videos does it offer? How are they made? If there are real flesh-and-blood human performers in the video, then that’s ideal.

As already mentioned, YouTube has its AI labels, but I wouldn’t rely on them to any great extent. If you can’t find evidence of how the music is being made or who’s behind it, and there are no links to actual recordings or artists (or footage of the music being made), then at this point I think it’s safer to assume it is AI rather than not.

My favorite non-AI Youtube music channels

One of the best and longest-running channels in the business in this category is Lofi Girl, which has been around since 2017, way before the generative AI boom. It was founded by a real record producer (Dimitri Somoguy), with an iconic character drawn by a real person (Juan Pablo Machado)—you can read about it on Wikipedia.

There are also now a growing number of channels that position themselves as containing no AI. One of my favorites is Yellow Cherry Jam: The videos here feature a man, a woman, a dog, and plenty of scenic backdrops. It’s all very relaxing—and real.

I also like Coulou’s Vinyl Cafe, where our man Coulou wanders around his apartment putting on one great record after another. The music is all listed and shown off in the video, and as good as AI video generation has become now, there’s no way it could create an hour of this without a chair leg disappearing or a jumper changing color.

Judging by many of the comments under those videos I’ve linked, AI-free music and AI-free YouTube channels are something a lot of other people are looking out for too: They are there, if you look for them. I’ve now built up a long enough playlist that I’m confident of not running into AI anytime soon, and it sounds great.