Circle to Search is now two years old, and looking back on the AI feature, it's clear that it was a monumental one for Google. At a time when phone brands were struggling to find practical use cases for artificial intelligence, Google struck gold by adding Circle to Search to Android. In the time since its release, Android gained flashier AI features, and I wouldn't be surprised if many of you forgot about Circle to Search completely.

If that is the case, it's time to revisit Circle to Search. Not just for nostalgia, but for the extra functionality Google added over the last few years. This isn't the same Circle to Search that debuted alongside the Samsung Galaxy S24 series — it can do more than just search what's on your screen. These are the three Circle to Search tools that make the feature still worth using in 2026.

Remix anything on your screen with Nano Banana

By now, you've probably heard a lot about Nano Banana, the model that powers generative AI image creation and editing across the Google AI suite. You might not know that you can access Nano Banana straight from Circle to Search, eliminating the need to use another app — like Gemini or Google Messages — to remix your photos. Simply activate Circle to Search and circle a photo. Then, a Create button with a Nano Banana emoji appears.

You can type a Nano Banana request from within the Circle to Search user interface, which then opens a Google AI Mode conversation when entered. It works just like Google's generative editing tools in other apps. You can make changes to a photo, add things to a photo, or request a combination of both.

For instance, if you have a screenshot or photo with sensitive info, you can ask Nano Banana to blur it. Or, you could ask the generator to turn a boring spreadsheet or list into an engaging graphic. Using the feature in Circle to Search saves you the hassle of needing to import photos or files as context in a Gemini conversation. Circle to Search can see your screen, giving Nano Banana all the context it needs, and it's easier to access in any app or site.

It compliments Circle to Search's crop and share feature nicely. Instead of taking a screenshot and cropping it manually, you can Circle to Search and export what you've selected to save you time.

Translate text instantly without a separate app

Another neat Circle to Search tool is the full screen translation feature. This is helpful if you're using an app or website that does not support native translation. Circle to Search simply translates the text on your screen when it is activated, so it'll work anywhere in Android. The feature works with a long list of languages, too, and it'll auto-detect the language on your screen.

Google gave Circle to Search's translation tool a major upgrade late last year. It added support for scrolling, so Circle to Search can continue translating text as you move through an app or site. This eliminates the need to Circle to Search multiple times if all the text won't fit into a single screen. Of course, you could use the Google Translate app, but this feature is all about convenience.

Scan QR codes and barcodes on your screen

Scanning QR codes and barcodes with an Android phone is easy when they're on another device or printed out on paper. It's trickier when the code is presented by an app or website on your device, though. Circle to Search can help, as it works just like the Camera app for detecting codes, but for content on your screen. When you activate Circle to Search with a barcode or QR code on the screen, it'll show a chip that lets you copy the code or follow the link to a website.

It's just one of the many ways Circle to Search intelligently presents different features based on the situation. When there's an image on screen, it will suggest Nano Banana for generation and editing. If there's a QR code or barcode, the system will read it and present an action chip. This could come in handy when interacting with all kinds of QR codes, which are used to complete eSIM activations, join Wi-Fi networks, and share other types of data.

Circle to Search may go down in history as Android's best AI feature. It does all the right things, blending artificial intelligence smarts with practical benefits. Selecting text, searching images, using generative AI, and scanning codes is all easier with Circle to Search. That's without mentioning other perks like song recognition and on-screen translation.

The versatility is what makes Circle to Search great, and if you're still only using the feature to access Google Search or Google Lens quicker, you're missing out. It's the one AI tool I use on every phone, and it just keeps getting better. Somehow, Google managed to add functionality to Circle to Search without the interface feeling cluttered. It's still easy to navigate. Instead of just replacing Google Lens, it could also replace Pixel Studio, Google Translate, and a handful of other apps.

It's no secret that smartphones are incredibly mature, and the days of justifying annual upgrades are over. Processors are fast, displays are crisp, cameras are sharp, and crucially, phone makers are supporting devices for longer than ever with up to seven years of Android OS upgrades. On the flip side, budget devices are winning the price-to-performance ratio, but they often fall short in one major area: memory.

Frankly, it's easy to switch from the most expensive Android phones to the cheapest ones. I've recently moved from the Samsung Galaxy S25 Edge to the TCL 60 XE NXTPAPER, and these are phones priced roughly $900 apart. Sure, one feels snappier and more premium, although both handle all my daily needs perfectly fine. I can live with a budget processor, display, or camera, but doing the same with low RAM is much trickier.

There's no better example of this than the Google Pixel 9a. Despite the Pixel 9a being the newest A-series Pixel available, it continually misses out on new AI features, presumably due to memory constraints.

The Google Pixel 9a started at a major disadvantage

Ideally, you want to buy a phone with future-proofing in mind. To keep a phone for a half decade or longer, it's essential to pick a device based on your needs today and in the future. This is why it's recommended to upgrade the storage configuration of your smartphone if you can afford it, because it'll give you the room to take on photos, videos, media, and apps as you live with a device for years.

As a result, it's hard to justify purchasing a phone that's already behind the curve on day one. Those limitations will only become more apparent as time goes on. However, that's what Google is asking Pixel 9a buyers to do. This phone ships with only 8GB of memory, which is much less than you'll find on the rest of the Pixel 9 and Pixel 10 series.

This is a modest amount of RAM that will be fine for daily use, until you want to use on-device AI. Artificial intelligence demands large amounts of available RAM. Since the Pixel 9a only has 8GB of RAM, it can't utilize the latest Gemini Nano models with multimodality. It has to use the Gemini Nano XXS model, which is a text-only model that only runs when it's needed.

The rest of the Pixel 9 and Pixel 10 series use Gemini Nano models that run all the time and utilize multimodality to handle text, audio, and visuals simultaneously. With only the Gemini Nano XXS model, the Pixel 9a started at a disadvantage in terms of feature set. It missed out on Pixel Screenshots and Call Notes, which are features that every mainline Pixel 9 and Pixel 10 series phone received.

It's possible that you consider the lack of AI features a benefit instead of a drawback. Some would rather not have their phone riddled with unwanted AI tools. However, when the Pixel 9a launched, I suspected this would only be the beginning of the feature omissions for this phone. Google's promise of seven years of Android OS upgrades and regular Pixel Drops won't mean much if the Pixel 9a misses out on the best features.

Not even a year into the Pixel 9a's existence, it's already starting to miss out on more new features. Scam Detection won't work on the Pixel 9a in certain regions. Additionally, the new AI notification summaries aren't available on the Pixel 9a. These are just a few features introduced with recent Pixel Drops that aren't available on Google's current midranger.

Some features simply can't use cloud processing

Theoretically, the Pixel 9a's onboard memory limitations shouldn't be much of an issue. Lots of AI processing can be offloaded to the cloud for additional power. Google Pixel 9a users can still use Gemini, Gemini Live, Nano Banana, and plenty more AI features that leverage the cloud. The handset can even use some on-device AI features, like Google Recorder transcription, since these features use older models that predate Gemini Nano.

The issue is that some AI features you really don't want processed in the cloud to protect your privacy. It's possible Google could bring Call Notes, Pixel Screenshots, and AI notification summaries to the Pixel 9a via cloud processing. But that would mean your calls, screenshots, and notifications being sent to Google servers for privacy. It's an option on paper, but not in reality.

The other reason certain AI features run on the device is for speed. It takes time to transmit data to a cloud server, process it, and return it to your device. In some applications, this rules out cloud processing as a viable option.

To sum up, features like Call Notes and AI notification summaries need to be processed on your device for low-latency responses and security, but they can't be on the Pixel 9a due to its RAM shortcomings.

Every other smartphone part has effectively peaked

Google is behind, even in comparison to other Android midrangers. The Samsung Galaxy A56 is configurable with up to 12GB of RAM, while the OnePlus 15R offers 12GB of memory as standard. To be fair, this is not a problem exclusive to Google phones, or the Pixel 9a specifically. The aforementioned Galaxy A56 gets "Awesome Intelligence" features instead of Galaxy AI due to its lower standard memory than flagship Galaxy phones.

In 2026, whether you use a Snapdragon 8 Elite or a newer Snapdragon 8 Elite Gen 5 chip is inconsequential for most users. The same goes for Apple, Qualcomm, or Google smartphone processors. There are clear differences between them, and we have benchmarks to demonstrate that, but the bottom line is that every modern system-on-a-chip (SoC) can handle the basics for the vast majority of customers.

While almost every other smartphone part has peaked, phones will always require more RAM to accommodate newer, more demanding AI features. The Google Pixel 9a is proof of what can happen if you ignore the RAM specification of your next smartphone. It's perhaps the only smartphone spec that still matters, because without enough of it, you'll be left without the latest and greatest features.

Of all the emerging form factors for AI, smart glasses appear to be the most promising. Many of us already wear some form of glasses daily, whether those are prescription glasses, blue-light blocking glasses, or sunglasses. Embedding smart features and AI into a product that is already a huge part of our lives sounds like a recipe for success. There's just one problem — prescription glasses-wearers, despite being likely to benefit the most from having AI on their face, are usually left out.

To me, Meta makes the best smart glasses on the market, and they're ready for primetime. So, I took a leap. I spent over $500 on third-party prescription lenses for my Oakley Meta HSTN and Vanguard. Here's what it was like customizing Oakley Meta smart glasses with unsupported prescription lenses.

Ordering custom Meta lenses

I've worn prescription Oakley glasses and sunglasses for half a decade now, and I was excited when Meta announced it was bringing its smart glasses to Oakley frames. Unfortunately, the Oakley Meta HSTN didn't officially support my prescription, and the Oakley Meta Vanguard is not offered with prescription lenses.

I wasn't about to give up, so I searched for third-party companies willing to make prescription lenses for Oakley Meta glasses. After months of research, I placed my first order with Lensology, a U.K. lens crafter that specializes in "click and fit" service for Meta glasses. I ordered a pretty specific lens configuration due to my prescription strength, and the color matched my Meta HSTN frames in the "Warm Gray" colorway.

The ordering process with Lensology was easy, and the company takes care of all import charges, so you aren't hit with a bill when they arrive. The wait, however, was a bit lengthy. It took approximately two weeks for Lensology to manufacture my lenses, as the configuration I chose was non-stock. Then, the U.S. FDA spot-checked my lenses during the import process, further delaying them.

I was told this was Lensology's first-ever FDA spot check (what luck!), and the company's customer service was responsive and helpful. Really, the delay was out of Lensology's control, and my lenses arrived roughly within a month of placing my order despite the snag. To make things right, Lensology sent another set of lenses free of charge, which passed customs immediately, confirming that this was just a fluke.

I then searched for a way to wear prescription lenses with the Oakley Meta Vanguard and stumbled upon SVED Optical. The company manufactures prescription inserts for various types of sports glasses and goggles, and indeed, they have one for Vanguard.

The experience ordering from SVED was exceptional for a few reasons. The company is based in the U.S., which simply makes things easier for a stateside buyer. The Vanguard inserts were also significantly cheaper, costing only $164 (even less with an email discount). They slot in behind the Vanguard's built-in Oakley Prizm sunglass lenses, which makes them more affordable.

SVED delivered my Oakley Meta Vanguard inserts 12 days after I placed my order. That turnaround time is good enough to rival official Meta/Oakley handling speeds, if it sold Vanguard prescription lenses.

Installation and fit

I try to order straight from LensCrafters, the EssilorLuxottica-owned company Meta partners with for smart glasses, whenever possible. It costs more, but the return policy and warranties are exceptional. I like being able to walk into a store and get my situation handled immediately, and not having that with Lensology or SVED was a risk. To their credit, both brands accept mail-in returns within 30 days.

When each order arrived, the quality and insertion process left me with no complaints. Putting the SVED inserts into the Meta Vanguard is so simple that anyone can do it. You remove the stock Vanguard nose piece and slip the SVED inserts into place. The prescription lenses are attached to a new nosepiece, which secures them to the Vanguard frame.

I was skeptical about the added weight and how my prescription Vanguard setup would feel on my face. Surely, the lenses would rattle and scratch against the stock sunglass lenses, right? Between daily use and a 5K run, I was pleased to find that none of these were noticeable issues.

The process was trickier for the Meta HSTN glasses, because you have to physically remove the stock lenses and push the prescription ones in their place. It'll be a harder process the thicker the prescription lenses. Lensology has a great instructional video on YouTube, though, and the company even records a test fit from their factory that validates the lenses fit in HSTN frames.

If you have any experience changing out prescription lenses, the process is easy. Even if you don't, you'll get the hang of it quickly. The secret is to push really hard — you might think the glasses will break, but it's a necessary part of the process.

Absolutely worth it

Meta's AI glasses are already pricey, but adding prescription lenses through a third-party company further increases the cost of entry. There's inherent risk, even with reputable companies. You might be left at a loss if you get buyer's remorse and need to pay to ship the custom lenses back through the return process. That's why, after months of indecision, I gave SVED and Lensology a shot so you don't have to take that risk.

I recommend both SVED and Lensology, and I ordered lenses from each with my own cash. If you can only pick one, I'd go with the Vanguard inserts. The cheaper price means you won't be in trouble if your prescription changes and you need a new set. On the other hand, the SVED and Vanguard combo isn't feasible for use indoors or at night. If you need clear or Transitions lenses, the Lensology and HSTN setup is the way to go.

Now that I've gone through the process of putting unsupported prescription lenses in Oakley Meta HSTN and Vanguard, I think anyone on the fence should do the same. It's worth it because Meta glasses are a game-changer, and everyone should get to experience them.

FAQ

Video generation models aren't new. Artificial intelligence has been able to create realistic video content for a while, and consumer tools are readily available. Both Google and OpenAI have released Veo and Sora text-to-video and image-to-video generation models in the past. So, why is the technology only blowing up now with the launch of OpenAI's Sora app?

For those unfamiliar, Sora is the AI video-generation app from OpenAI that combines a generator and social platform in one. It launched for iOS and on the web in September and recently debuted for Android. Chances are, even if you haven't heard of Sora, you've probably seen a video from it shared on another platform, like X or TikTok.

Sora is on its way to becoming a household name for OpenAI in the same way ChatGPT did two years ago. It's succeeding and starting to infiltrate pop culture in a way Google's Veo video-generation models haven't. After trying both Veo 3 and Sora 2, it's clear that OpenAI is winning the generative video race by going all-in on presentation — and removing as many limitations as possible.

Veo 3 and Sora 2 are both hyper-realistic video generation models

AI video generation models like Veo 3 and Sora 2 face an uphill battle. Simply making an excellent, hyper-realistic video generator isn't enough to be successful. Brands like Google and OpenAI need to make AI videos cool, and I wasn't sure that was possible until the Sora app launched.

By wrapping the Sora 2 model in a TikTok-like social feed and adding human cameos, OpenAI seemingly changed the perception of AI content overnight — what was once slop and garbage became funny and engaging. That has nothing to do with the model itself, but rather, the interface.

Sora, like Gemini, is both a model and an interface. Sora 2 is the video-generation model powering the experience, and OpenAI says it is "more physically accurate, realistic, and more controllable than prior systems." The company adds that it thinks Sora 2 "may be the GPT‑3.5 moment for video."

That sounds a lot like Google's claims regarding Veo 3.1, which just rolled out to the Gemini app in October. Google said Veo 3.1 adds "richer audio, more narrative control, and enhanced realism." It's pretty similar to what OpenAI touted about Sora 2 a month earlier.

Why hasn't Veo 3 surged in popularity like Sora 2? It's because Google doesn't have an app to tie it all together.

OpenAI's Sora app turns AI video generation into a social platform

Google's Veo 3 model, while technically as capable as Sora 2, is often positioned more toward developers and businesses than the average consumer. It's accessible in the Gemini app for Google AI Pro or Ultra subscribers, and it powers photo-to-video generation in the Google Photos app. Even those implementations are ones the user needs to seek out — they're not a major part of either app.

By comparison, Sora 2 is available at sora.com, but OpenAI also built an entire app around it. The company noticed the vertical video and infinite scroll formats working for social media platforms like TikTok, Instagram Reels, YouTube Shorts, and others. It responded by making the Sora app a similar hub for AI-generated videos and a place where users can make their own.

A core part of the Sora experience is the cameo feature, which allows users to make realistic AI videos of themselves or others. These can eventually be shared with others via the Sora feed. The idea is that you connect with your friends on Sora, create AI cameos, and share them, so that everyone has a TikTok-like feed of clever AI content.

Comparing Veo and Sora quickly revealed that OpenAI's approach with cameos immediately separates it from the competition. At a time when many video generators are blocking the creation of AI video content about real people, OpenAI made it Sora's entire reason to exist. There are limitations to prevent people from using Sora irresponsibly, but they end up feeling more like detours than outright roadblocks.

Here's what OpenAI says about how people can protect their likeness within Sora:

Still, these are looser restrictions than what you'll find using Veo 3 with Gemini. People are unsurprisingly always trying to find ways around them, too.

While testing Veo 3 and Sora, I struggled to get Gemini to create anything realistic involving actual people or celebrities. Gemini would say, "I can't generate that video," and link to its video policy guidelines. Oddly enough, there's nothing in there that specifically bars generating videos of real people, but Gemini and Veo models are clearly trained against it.

This puts Google and Veo 3 in a precarious position because Sora's popularity proves people crave realistic AI videos of actual people.

Are Google's strict policy guardrails holding its AI features back?

OpenAI's headfirst dive into lifelike AI videos represents a sharp contrast to Google's insistence on AI responsibility. I can't say for certain why OpenAI is starting to bend the lines, but after following the AI industry closely for years, I have a theory — AI development is plateauing, and OpenAI's competitors are catching up.

Shortly after announcing Sora, OpenAI's CEO, Sam Altman, said in October 2025 that ChatGPT would soon be allowed to engage in "erotica for verified adults" (via AP). Altman added that OpenAI is "not the elected moral police of the world."

These are separate decisions, but as a whole, they appear to show OpenAI systematically removing guardrails for its AI products. By letting Sora go all-in on AI videos of real people, OpenAI created an instant hit. Meanwhile, Veo is still relatively unknown to the masses.

To me, AI videos are still slop, regardless of whether Veo or Sora makes them. Who knows if OpenAI's bet on Sora, social, and cameos will pay off? But for now, it's hard to argue with early results — the Sora app has already been downloaded millions of times across iOS and Android.

FAQ

Artificial intelligence wants to be everywhere, including on your wrists and in your workouts. Compared to generative AI "slop" like photo and video creation, there's a case to be made for AI infiltrating your workout and fitness plans. Not everyone has access to a personal trainer, sports physician, or physical therapist easily — in fact, few do — and that's why an accurate and accessible AI feature could fill the gap.

Of course, that puts a lot of pressure on a piece of tech known to be fallible. Anytime your health is involved, you need to be careful what devices and features you trust. Whether we like it or not, AI is coming for the health space — Garmin, Strava, Apple, Samsung, Google, and many other brands are rolling out AI-powered features now.

Suunto upgraded its Suunto Coach, an AI-assisted workout helper, to support AI-generated training plans over the summer. I've been testing them with the newly-announced Vertical 2 adventure watch, and I'm intrigued for two reasons. For one, Suunto Coach and My plan are free. Another is that they're actually using AI, which not every workout feature branded as such can claim.

So, is it any good? Here are my early thoughts after training with the Vertical 2 for two weeks and setting up an AI workout plan.

Setting up a custom My plan with Suunto Coach

After logging a bunch of running and walking workouts with the Suunto Vertical 2, I went to set up the AI-generated training plan with Suunto Coach. The process starts with a brief questionnaire for context as to your training goals and capabilities.

Suunto will ask how many hours you are currently training in a week, and then ask your maximum time commitment. You'll have the chance pick a dedicated day for long runs, or state a preference for strength and cross-training workouts. Finally, you'll share your race goal and terrain.

The questions might vary based on your training plan. I selected the "Get fast for 5km" training plan, as I'm more worried about cutting my 5K time than increasing mileage. I answered the questionnaire accordingly, with my goal being to achieve a personal best on a flat and fast course.

From there, Suunto will do some AI processing behind the scenes, developing your training plan. This will take a while, often between 10 and 30 minutes. The good news is you only have to do this once for each plan. Using the context based on your answers, Suunto will come up with a personalized training regimen.

The initial approach reminded me of Samsung's Running Coach, a similar feature that includes a brief questionnaire to get a picture of your capabilities. There looks to be more AI processing going on here with Suunto Coach. It takes more time to create a training plan and offers more variety than Running Coach, which is more rigid with fixed time and distance-based running levels.

That's what intrigues me about Suunto Coach and the AI-generated training plans, which are called "My plan" and are still in beta. For better or worse, they're really using AI. Meanwhile, AI isn't making a meaningful impact on Running Coach or Apple's Workout Buddy. The latter essentially just recites live stats you're already used to seeing on your Apple Watch.

A workout tool with potential, and some quirks

Once the training plan generated, I saw immediate potential and a few oddities. The plan does seem personalized and thoughtful, with weekly targets, long-term goals, and instructions specific to individual workouts. It does a fantastic job at taking best practices for running and applying them to a specific plan.

For instance, my plan included strategic rest days and an easy, low-mileage week that followed each pair of hard training weeks. The 10-week set prescribed six weeks of intense workouts and three weeks of recovery running, including a taper for the final week. It's all designed to maximize performance while also managing miles and avoiding overuse injuries.

These are well-known running tips and tricks that even experienced runners struggle to implement. It's great to see them in a thorough training plan that takes the strategy off the runner and onto AI.

That said, some things could use a tweak. The time, mileage, and pace targets for each workout didn't seem to align with the body of work I had already tracked on the Vertical 2. This leads me to believe the AI brains of Suunto Coach is using the questionnaire more than your prior stats when making a plan. It is supposed to evolve on the fly, and it is still a beta.

I also found it weird that the AI workout plan randomly inserted a cross-country skiing workout during week eight. I've never cross-country skied, and it's not the easiest sport to try for the first time, so it's an interesting choice.

Suunto Coach feels best as a guide, not a plan

I plan to stick with Suunto Coach, despite the room for improvement. It seems genuinely helpful for training as long as you don't take it too seriously.

By that, I mean following its recommendations generally and not to the letter. In my case, I'll follow its broad plan for weekly mileage, workout mix, and training load — but avoid hitting an exact mileage or pace target.

This is a free tool included with Suunto watches, and it'll definitely be handy for the right runner. You just need to take its tips with a grain of salt, just like any AI feature.

Everyone has their opinion on the Android vs. iOS debate, and it generally comes down to personal preference. However, artificial intelligence features seem like the one area where the best Android phones objectively take the crown. Google is firing on all cylinders in AI development, and the switch from Google Assistant to Gemini resulted in generally positive feedback.

Apple Intelligence, meanwhile, has hit snags and roadblocks at every turn. Core features advertised years ago still haven't arrived, and the ChatGPT integration aims to bail out Apple for stalling AI progress. So, is this as clear-cut as it may seem? Here's how Apple Intelligence on iPhone compares to using Gemini on Android.

Gemini vs. Siri: Does iOS or Android have the best assistant?

Siri, the digital assistant on iPhone, is intertwined with Apple Intelligence features. It's worth noting that the revamped version of Siri teased at a prior Apple Worldwide Developers Conference hasn't actually shipped yet. By comparison, Google is in the process of replacing Google Assistant with Gemini as Android's digital assistant of the future.

Siri and Gemini both have two different modes of interaction: voice and text. iPhone users can call up Siri with a long press of the side button, a double-tap of the gesture bar, or the wake word "Hey Siri." Android users can access Gemini with a long press of the side button or the wake word "Hey Google." As Gemini is a full-fledged chatbot, it has better access to the web and is powered by large language models for context and AI processing.

That enables unique features for Gemini that Siri can't match. While both digital assistants support text commands, only Gemini can accept file uploads for additional context. Gemini is also better at responding to basic questions in natural language, whereas Siri will frequently direct users to a search engine if it can't come up with a response.

Keep in mind that users of Siri and Gemini have reported unreliability in some use cases, like while managing device settings or controlling smart home systems. In fact, Gemini lacks a few core Google Assistant features that were either deprecated or have not yet been added to the AI digital assistant.

Gemini Live vs. Visual Intelligence: Which multimodal AI is best?

Android users gain access to a popular multimodal AI interface called Gemini Live, which is available on most Android phones across brands and generations. However, you may not know that iPhone users have a similar alternative, called Visual Intelligence.

Gemini Live supports live video streaming and screen sharing, while simultaneously accepting voice commands and text input. This multimodal solution allows you to share what's on your screen or in your surroundings and ask questions about it. As the "live" branding suggests, this all happens in real time, with only brief delays to account for cloud processing.

Visual Intelligence might look similar, but it's much different. The feature functions more akin to uploading a photo to Gemini for context than using Gemini Live. With Visual Intelligence, you can snap a photo and converse with ChatGPT about it over text. Or, you can search Google for the photo using Lens. Either way, Visual Intelligence doesn't support the "live" component and is limited compared to Gemini Live.

Gemini Live is the more helpful and versatile option here — I've used it for everything from tending to plants to designing apartments. The good news for iPhone users is that Gemini Live can also be enjoyed on iOS via the Gemini app.

Magic Eraser vs. Clean Up: Who has better image editing?

Apple and Google each offer their own generative editing features that leverage AI, known as Clean Up and Magic Eraser. Samsung has its own, too, but we'll compare Clean Up to Magic Eraser since the latter is available on all Android phones. Google has been doing object removal and generative fill for longer than Apple, so it's better, right?

That actually might not be the case. In our latest test, comparing outputs between Clean Up and Magic Eraser led to mixed results. Clean Up successfully removed certain objects, like a dog's leash, in a way that looked natural and impressive. In another instance, Clean Up and Magic Eraser removed people from a photo that passed the eye test at a distance — but lacked quality up close. And in some cases, Clean Up would fail to correctly remove objects it removed just a few minutes before.

In other words, Clean Up and Magic Eraser are both less-than-stellar ways to remove unwanted people or objects from photos on an iPhone or Android device. They work in some situations and underwhelm in others. Neither is replacing Photoshop anytime soon, and thus, this battle between generative AI editors ends in a draw.

Beyond generative removal and fill, both Android and iPhone have AI image generation support. On the Android side, it'll depend on your phone manufacturer. For example, Google Pixel phones have Pixel Studio, while Samsung Galaxy phones have Sketch to Image. iPhone users get Image Playground, but all of these image generation clients are effectively the same.

They all leverage text-to-image AI models to generate photos, some of which are better than others. Generally, the model's safeguards will affect their handiness. Image Playground has strict guardrails preventing it from creating lifelike images, whereas Pixel Studio can now generate images of people.

Android vs. iPhone: Security and privacy

Security and privacy are the one area where iPhones have the edge. Most Gemini features require cloud processing and include extensive data sharing policies. By default, your data stored by Google is retained for 18 months before being deleted. Anything that human reviewers looked at is kept for up to three years, even if you manually delete your data.

As such, you need to keep privacy at the forefront when using Gemini on Android. You wouldn't want to share anything you wouldn't be fine handing over to Google or its employees.

Apple treats Apple Intelligence differently — it processes AI tasks on-device or using its Private Cloud Compute servers. They're run using a custom operating system that is made public so independent security researchers can verify there are zero vulnerabilities. With Private Cloud Compute, not even Apple has access to your data.

The notable exception is that using ChatGPT integration will share queries with OpenAI, but you'll need to give permission every time you choose to tap into those resources.

Android vs. iPhone: Verdict

At the moment, Android has the clear edge when it comes to mobile AI. Gemini is a better digital assistant than Siri, and Gemini Live blows Visual Intelligence out of the water. Image generation is a toss-up, and better security and privacy policies aren't enough to pull Apple Intelligence ahead. Apple's weak foundation models mean users will need to use external sources like ChatGPT more often, negating privacy benefits.

Beyond rudimentary use cases like generative photo or video creation and text editing, artificial intelligence is starting to help out in real-world situations. Notably, live translation features that eliminate language barriers in real-time are on the rise.

We've seen features debut in smartphones, earbuds, and even smart glasses. Truthfully, the idea of live translation isn't new, but it's coming into its own in 2025 thanks to improvements in support and performance. Apple recently highlighted Live Translate as a perk of its new AirPods Pro 3, and a similar feature leveraging Google Translate already existed on the Google Pixel Buds Pro 2.

So, is Google's or Apple's version of live translate the best? We tried them both, and although neither is perfect, these are the key ways they compare — and which one we prefer.

All the differences between Apple and Google's live translate features

Live Translate is marketed as a unique earbud feature in 2025, but the legwork is still being done on your smartphone. To use Apple's live translation with a supported pair of AirPods, you'll need an iPhone 15 Pro or newer running iOS 26 or later. While it might seem like an arbitrary cutoff, an iPhone 15 Pro is needed because live translation uses Apple Intelligence, which isn't available on the base iPhone 15 or older models.

This also means that Android users who own a pair of AirPods for wireless audio (yes, they do exist) won't be able to take advantage of Apple's live translation features at all. By comparison, Google's equivalent is much more forgiving. You only need an Android phone running Android 6 or later to use live translation on a pair of Google Pixel Buds.

This includes all Google Pixel phones, but also the vast number of Android phones by other brands that meet the version requirement. You can use a pair of Google Pixel Buds with a Samsung or OnePlus phone and still use Google's live translation feature, for instance. You'll just need to ensure your phone has the latest version of the Google and Google Translate apps installed, as both are required for live translation.

Both the Apple and Google implementations of live translate use off-device processing by default, so you'll need an internet connection. However, you can download languages for offline translation via the Settings app on iOS and the Google Translate app on Android.

Apple also offers an "on-device mode" that will always use offline, on-device processing for translations. However, it's worth noting that this mode only applies to the Apple Translate app, which is used for AirPods translation. If iPhone users want to translate with Siri or Safari, those translations will always be processed in the cloud.

The way you trigger live translation on Pixel Buds and AirPods is also slightly different. For Google Pixel Buds, the preferred method of using Live Translate seems to be using Google Translate on your phone and tapping Conversation mode. Alternatively, you can use Google Assistant to trigger live translation by saying a phrase like "Hey Google, help me speak [language]."

There's no way to use Conversation mode hands-free without using Google Assistant; otherwise, you'll need to pull out the Translate app on your Android phone. Apple's implementation differs slightly. You can use Siri for translation, but the preferred method is to hold down both AirPod stems to trigger live translation.

Do you need the latest AirPods or Pixel Buds to use them?

Both Apple and Google have tremendous support for live translation features, even on older devices. However, Google clearly offers an edge. It has been shipping real-time translation on earbuds since 2017, dating back to the very first pair of Google Pixel Buds. The feature has since been available on every pair released since, from the high-end Pixel Buds Pro 2 to the affordable Pixel Buds A-series and new Pixel Buds 2a.

Apple's feature is currently listed as a beta offering, so it won't be perfect. It was announced alongside the AirPods Pro 3, but it's also coming to AirPods Pro 2 and the ANC model of AirPods 4 via a software update.

So, chances are, if you have a new enough pair of Apple or Google earbuds, you'll be able to utilize live translation.

Are the AirPods Pro 3 or Google Pixel Buds Pro 2 better for live translation?

The aforementioned differences between calling up live translation on AirPods and Google Pixel Buds are the most impactful to the experience. It's nice to be able to walk up to someone who doesn't speak your language, hold both stems, and start hearing a translated version of their speech in your ears. There are no voice commands or smartphone taps required when using a supported AirPods and iPhone, which is ideal.

By comparison, it can be cumbersome to use clunky voice commands or pull out your phone to use Conversation mode with live translation on a pair of Google Pixel Buds. Certainly not a dealbreaker, but it's a minor inconvenience.

In terms of quality, there's still a ways to go before we're conversing at a natural pace with others between languages. The live translation features on both Apple or Google earbuds have a noticeable delay — currently necessary for processing and wireless transmission — that might impede the flow of a conversation.

Additionally, it isn't uncommon to have either feature completely miss a sentence or phrase. All it takes is someone speaking too fast, a noisy environment, or an unfamiliar dialect to throw Apple Translate or Google Translate off its game. However, it's certainly easier to use AirPods or Pixel Buds for live translation than flipping through a language phrasebook or guide.

Like almost everything leveraging AI these days: when it works, it's magical, but you can't depend on it.

Everyone is racing to make the best AI phone, but it's really a two-horse race between Samsung and Google. There are pros and cons to choosing either brand, and I've tested the latest features from both Gemini and Galaxy AI. It's clear to me that, if getting the best AI features on your phone is your top priority, the Samsung Galaxy S25 series is the one to go with.

That's not to say that the Google Pixel 10 is doing anything worse than the Galaxy S25 — it's just a numbers game. Google is very generous, gifting its best AI features to other Android and iOS devices, whereas Samsung's Galaxy AI suite is exclusive to its own devices. If you use a Galaxy S25, you're still getting multimodal Gemini Live, Circle to Search, and AI editing tools in Google Photos. Access to both Google AI and Galaxy AI is a major plus for the Galaxy S25.

That's partly why I'm intrigued by a Verizon deal that claims to give away a Samsung Galaxy S25 or Galaxy S25 Plus for free, which sounds too good to be true. In some ways, it is — you have to add a line on the Unlimited Ultimate or Unlimited Plus plans. That means it's not completely free, but if you're looking to switch plans or already have one of these Verizon plans, this is a sweet way to get my favorite AI phone for less.

✅Recommended if: you already plan to add a new cellular line; you want a flagship Android phone with a robust AI suite; you need on-device AI and want to be able to fine-tune your privacy settings.

❌Skip this deal if: you don't want to be locked into a multi-year carrier contract; you don't see value in having both Galaxy AI and Gemini; Verizon coverage isn't great in your area.

Why I think Samsung is beating Google in mobile AI

There are two key ways the AI experience on Samsung Galaxy phones is better than the comparable experience on Google Pixel phones, and it comes down to feature availability and customization. Starting with feature availability, devices like the Galaxy S25 have access to Galaxy AI and Gemini, which is an impressive combination. You probably know all about what Gemini can do, but you should also be aware of the great things Samsung is doing on the AI front.

The Now Bar and Brief are both designed to use AI to put relevant items in front of you whenever you need them — think something like an AI-powered version of Apple's Live Activities and Dynamic Island, but for the Galaxy S25 series. These features are neat, but they heavily depend on whether you use supported services, like Samsung apps or Google Maps.

I find that the simpler Galaxy AI tools are more handy, like Object Eraser and Audio Eraser, which perform better than Magic Eraser and Audio Magic Eraser on the Google side of things. Additionally, if you're into wearables, Galaxy AI powers a daily Energy Score and Sleep Score with a Galaxy Ring or Watch. This is especially great when you consider these health insights are completely free, whereas Google has some Pixel Watch features behind a Fitbit Premium subscription.

Best of all, Samsung has the hardware and software needed to deliver a private and secure experience. The Qualcomm Snapdragon 8 Elite for Galaxy chipset is great at on-device AI processing, but so is the Google Tensor G5 chip. The difference is that Samsung gives you a helpful toggle that turns off every AI feature requiring cloud processing.

This toggle lets you choose to process data only on your device, which is something you won't find on a Google Pixel or iPhone. As the company explains, "for added privacy, you can prevent collection and processing of your data for Galaxy AI's advanced intelligence features."

It's a close race between Samsung and Google, but I think Galaxy phones have earned an edge due to their support of multiple mobile AI suites and their built-in privacy customization options.

Is buying a Samsung Galaxy S25 on a Verizon plan worth it?

The Samsung Galaxy S25 and Galaxy S25 Plus are excellent AI smartphones, but whether this particular Verizon deal is right for you will depend on a few things. Before you get sucked into the hype of a "free" phone, especially one as good as the Galaxy S25, review the fine print. In this case, be aware that you need to add a line, upgrade, or keep an Unlimited Plus or Unlimited Ultimate plan to get the deal.

The savings that cut the entire retail price of the Galaxy S25 or S25 Plus down to $0 are applied as bill credits over the course of 36 months. If you cancel, switch plans, or upgrade early, you'll be on the hook for the remainder of the phone's cost.

That might sound scary, but it doesn't have to be, if you know what you're getting. With an Unlimited Plus or Unlimited Ultimate plan, you get unlimited talk, text, and data, plus extra perks like mobile hotspot and 5G UW support. Those who already need a great wireless plan can get one at Verizon while also snagging a free Galaxy S25 in the process.

Artificial intelligence is great at some things, and not-so-great at others. Finding the ways AI can help your workflows or improve your daily life can be difficult, especially without compromising quality.

I've found that AI chatbots like Gemini or ChatGPT are excellent at lending a hand when you have a specific problem that requires personalized context. Gemini first helped me tend to my plants earlier this year; I could've used a search engine for ideas, but none would've been able to take the real-world appearance of the plants into account.

For those without a green thumb, I discovered yet another use case for mobile AI while struggling to come up with a design layout for my apartment. After a year of moving things around, buying new furniture, and reading unhelpful articles online, I finally pulled out my Google Pixel 9 Pro XL in a moment of frustration and turned to Gemini for help. I didn't expect much, but was blown away by what Google AI came up with.

Here's how I used AI to solve a complicated interior design issue, and how you can use it to help with your own problems too.

Giving Gemini the right starting point

The problem with my apartment is simple: it's fairly large, but there's only one window in the living room, kitchen, and dining area. The window is actually a sliding glass door, and is flanked by perpendicular walls with double doors, which limits furniture placement. With one natural light source, the wrong layout could make this long and relatively narrow living space feel dark and bland — that's why I needed to turn to AI.

AI is worst when you don't point it in the right direction, and best when you provide it with ample context and thorough prompts. Instead of giving Gemini a vague and generic prompt, I found the exact floor plan for my space for context and clearly stated my goals. The prompt doesn't have to be lengthy, but it does need to be thoughtful and specific.

In this case, after providing the floor plan, I asked Gemini how to lay out the living room to get the most lighting.

Gemini's personalized and specific help

Gemini quickly responded with some general tips, like avoiding blocking the light source, using reflective surfaces, and choosing light colors. These were helpful in their own right, but giving Gemini an understanding of the floor plan returned personalized results. The chatbot understood that "the living room is a central, long space with a large window at the end."

It sounds simple, but Gemini having a basic understanding of the layout already gave it a massive advantage over search engines or interior design websites. As an example of how, one of the sites I read suggested renovating the room to change the positions of windows and doors, which was laughably unhelpful for a renter.

Gemini provided three detailed suggestions, and I went with the "open flow" layout. This involves floating a couch in the middle, facing the glass door, creating an open concept in the room. To avoid moving stuff around like a madman, I asked Gemini to come up with a few examples of what this would look like in my space — another thing I couldn't have done without AI.

The other cool thing about using AI for problem-solving and brainstorming is that you can ask further questions as you go. If you have essential needs, like a TV in a certain spot, Gemini can take that into account and only provide solutions that meet this requirement.

AI also suggested ideas and solutions that never even crossed my mind — like using mirrors around the room to simulate more windows and natural light in the space. After going back-and-forth with AI for about ten minutes, I had an interior plan that resulted in the following:

If you ask me, it's a pretty bright and open space, which is the vibe I've spent a literal year trying to create — and is something I was only able to achieve with Gemini. This was all done using the standard Gemini chatbot, and you could get even more personal help by using Gemini Live's multimodal live video support.

How you can use Gemini to solve life's problems

You can use Gemini to solve problems in your daily life by focusing on three things: context, prompting, and refinement.

Context is the crucial starting point. If you don't provide an AI chatbot with ample context, it only has its training material to go off of, which can lead to unhelpful results. You can go about providing context in a few ways, such as providing AI photos or screenshots, like I did with my apartment's floor plan. Or, you can use features like Gemini Live that share your surroundings via your phone's camera, which are other ways to provide context.

From there, be sure to write a thorough prompt that clearly explains your goals and what you need AI to do. You should use complete sentences for the best results, but avoid writing long paragraphs. Lastly, don't forget that you can refine your prompts as you go and steer the AI chatbot in the right direction. This can help if at first the AI doesn't provide the results you're expecting from your prompt.

With these factors in mind, you can get AI to help you with almost anything, whether it's gardening or interior design — or something else completely.

Apple's infamously slow start in the AI race has been well-documented, and Apple Intelligence hasn't helped it catch up in a meaningful way. Even partnerships with OpenAI and ChatGPT weren't enough to help the iPhone compete with mobile AI suites like Google AI or Galaxy AI.

Apple seemed to learn from its mistakes at the iPhone 16 event; instead of announcing lofty feature goals alongside the iPhone 17 that it couldn't deliver on, it chose not to discuss artificial intelligence at all. The absence of Apple Intelligence talk at the company's recent keynote wasn't really a shock, considering key features announced a year prior still aren't ready yet. It begs the following questions: does Apple even need AI to sell iPhones, and do iPhone users even want it?

The answer to the first question is still up in the air, but we now decisively know that Apple's users do care about AI. Google's Gemini app for iOS recently overtook ChatGPT to become the top free app on the App Store. It's a major win for Google, as it signals that the momentum is shifting from ChatGPT to Gemini. Additionally, it's a sign that even as Apple's internal AI efforts may be stalling, iPhone users are getting their mobile AI fix somewhere else.

Why Apple barely talked about AI in its iPhone 17 event

Apple's fall keynote was light on AI features, with the biggest one discussed being Live Translation on the AirPods Pro 3 and older models. Apple focused on practical implementations of AI, like Live Translation, which bridges communication barriers by translating conversations to your preferred language in real time. The same strategy can be seen with Workout Buddy, an AI feature for Apple Watch that helps coach you with fitness insights during your session.

The positioning of new features as practical solutions to real-world problems is a stark contrast to how Apple has talked about AI in the past. At past WWDC keynotes and launch events, the company touted Apple Intelligence features in traditional generative AI implementations, like text editing, image creation, and voice assistance.

The problem with the latter approach is that Apple's foundation models aren't on-par with those from Google or OpenAI, so Apple Intelligence can rarely compete with Gemini or ChatGPT in traditional gen-AI tasks. Instead, where the company can make its mark in the AI space is with specific features that leverage Apple hardware for a unique task, much like Live Translation or Workout Buddy.

While the company is still working on improving Apple Intelligence's capabilities, it's choosing to talk about those gen-AI features in press releases rather than lengthy keynote segments. After all, it seems pretty clear that competitors provide a better experience in text and image generation as well as voice assistance than Apple Intelligence and Siri. Now, that's reflected by iPhone users in the App Store's top charts.

Why Gemini is a better stopgap for iOS users than ChatGPT

Gemini is now the top free app on the App Store, and this is an impressive feat that shows just how far Google AI has come. Across 483,000 ratings, the Gemini app for iOS has 4.8 out of five stars, which is standout performance. It's also the number one app in the productivity category, ahead of ChatGPT and Google's own Gmail app.

Out of the thousands of reviews for the Gemini app on the App Store, there's one that really exemplifies why Apple users are flocking toward Google for their AI needs. "I wasn't sure if this app would work as advertised but it absolutely does," writes user Jrich828. "Siri does not work as well for me but Gemini has managed to answer all my quirky questions. Love it!"

That's the jist of why iPhone users are starting to use Gemini in droves. It's not just that Google's AI is useful — that wouldn't be enough for people inside the Apple ecosystem to give Gemini a try. Further, it's that Apple Intelligence and Siri are failing to fulfill a need, so iOS users are looking for something better.

While it's just one app review, the above sentiment is extremely common in Apple circles. Siri has been behind the curve dating back to the Google Assistant days, and Gemini can simply run circles around Apple's AI offerings. On top of that, Gemini is really easy to use on iPhone. You can add Gemini widgets to your home and lock screens for easy access, and even open the app with a voice command if you're savvy with Shortcuts.

Even as Apple partners with ChatGPT for Apple Intelligence, users are going out of their way to try Gemini. Aside from the above factors, I suspect iOS users are eager to use Google AI because they're familiar with other Google apps and services, like Gmail, Search, or YouTube. The Google ecosystem may not feel as familiar as Apple's, but it's not foreign either.

Gemini becoming the top free app on the App Store is an indication that iPhone users have spoken. Gemini and ChatGPT are numbers one and two on the platform. Users want AI on their phones, but they'd rather have better third-party options than Apple Intelligence.

Wearable technology is ubiquitous in 2025, with many people donning smartwatches, bands, or rings to gain insight into their fitness and health. We've also seen companies like Apple, Samsung, and Strava add AI coaching options for their products.

However, AI software in the fitness space is only one piece of the puzzle. Hardware companies are also developing new wearable concepts with AI at the core. To get an idea of where we're at and where we're headed, I tested Samsung's Running Coach software and Hypershell's X Ultra exoskeleton hardware at IFA 2025 in Berlin, Germany.

As an amateur distance runner, I wanted to find out if adding AI to my training could really help. Here's how it went.

Personal trainers aren't going anywhere... yet

Samsung debuted Running Coach alongside the Galaxy Watch 8, and our own wearables expert Michael Hicks put it through its paces. His experience was less-than-stellar, as Running Coach underestimated his abilities. That got me wondering whether a more amateur runner like myself would be a better fit for the software.

At Samsung's mobile devices booth at IFA 2025, I filled out a quick survey about my sleep and running habits, and was then given a slip with my running level and a QR code.

Unfortunately, I still ran into Running Coach limitations. The feature is built around fixed goals that are linked to your running level — I got Level 5 after the quick survey at IFA — with little room for personalization.

I think this strategy is a mistake, because it makes Running Coach too centered upon the goal of finishing a race length you may not be interested in. Additionally, when you reach a goal, you simply move to the next level without much focus on being able to complete those distances during training on a week-to-week basis.

For instance, the Level 5 training plan has me trying to finish a 10K under 1:08:20 (11:00/mile) — something I can already do. I can't tell Running Coach that I'm trying to cut my 5K time rather than build up to longer distances, or that I'd rather gain the strength to run five times per week before scaling distances. This seems like something that would be easy enough for Samsung to add with a few extra survey questions.

I'll skip using Running Coach for my own training plan for now, but using Gemini AI on the Galaxy Watch 8 has actually been a big help. You can use it to ask quick questions, like running routes or stretches. It's also helpful for starting workouts and controlling your running watch hands-free.

Hypershell's X Ultra exoskeleton could be a game-changer

I also tried a new AI hardware wearable at IFA, which falls into an emerging product category: exoskeletons. It's a device that straps onto your waist and thighs, featuring a motor on each side for lifting your legs robotically. I admit I was skeptical, and it sounds like something straight out of a sci-fi movie, but it's real, and it's called the Hypershell X Ultra.

The company explained that its goal is to enhance human capabilities. The $2,000 exoskeleton utilizes AI and machine learning to anticipate your next move and provide motorized assistance accordingly. It uses data from more than 12 sensors and its AI MotionEngine Ultra, built on training data, to adjust to your individual stride signature. No one walks the same way, and that's why AI and ML are necessary.

The device connects to an Android app and is designed for physical activities, like walking, hiking, and running.

It's something you really need to feel to believe, but the motorized exoskeleton actually does help lift your leg up in alignment with your stride to reduce exertion. You can feel the Hypershell X Ultra adjusting power and dynamics based on what you're doing — an easy way to test this is to shimmy your legs in place while feeling the motors ramp up in line.

The extra power made it much easier to walk, hike, and climb stairs. I switched between transparency mode, which temporarily disables assistance, and the Eco and Hyper modes to get a feel for how much work the exoskeleton was really doing.

It's doing a lot, but I don't think AI wearable hardware like the Hypershell X Ultra will help you run faster or longer. Not because the device doesn't work, but you still need to build cardiovascular endurance to run well.

Additionally, my short sprints with Hypershell revealed that the waistband slightly limited my mobility. The gain in power you get from wearing the exoskeleton might be countered by less efficient running mechanics.

I don't see people wearing exoskeletons to beat their PRs, without even getting into the "cheating" debate. However, I do see runners potentially using exoskeletons for training. Fitness mode on the Hypershell X Ultra turns the motors against you, making it harder for you to run at your normal pace. It can also reduce impact stress while going downhill, possibly lessening knee and joint pain.

It might sound like torture, but advanced runners already do this with devices like speed parachutes. Personally, I find exoskeletons to be a more elegant solution for resistance training, and I'd definitely use a Hypershell in Fitness mode to intensify my training.

AI hardware vs. software

AI software tools are more accessible to runners than hardware exoskeletons, but I find the latter to be a more exciting training companion. AI running coaches simply don't provide enough personalization options, while the hardware can elevate your training while literally personalizing the experience in real time using AI and ML algorithms. I'd never use it in a race, but I do think it can help you become a better runner.

Note: Hypershell provided travel and lodging for part of my trip to IFA, but had no input in this article.

The Google Pixel 10 series may be the first lineup of true AI smartphones. Other brands have tried to claim that title before, but fell short of delivering a next-generation experience. I'm confident the Google Pixel 10 will change that.

With multimodal Gemini Live, a Tensor G5 processor capable of running on-device AI models, and a Magic Cue feature that offers help without the user needing to ask, the Pixel 10 is the most complete AI smartphone yet. Those aforementioned perks are great, but they're also familiar.

If you're craving something completely new, look no further than Camera Coach. It walks a fine line between artificial intelligence instruction and photo editing. Instead of using Google Photos or Magic Editor to fix a photo after you shoot it, Camera Coach gives you tips on how to improve your output before you tap the shutter button.

As a fan of mobile photography, I'm intrigued by Camera Coach. I don't want AI to consume the art of capturing pictures, but Camera Coach has the potential to be a teaching tool like nothing we've seen before. If used the right way, it could make photography more approachable to the masses.

How Camera Coach works

Camera Coach is a generative AI tool that is debuting in preview alongside the Google Pixel 10 lineup. It's powered by custom Gemini models tuned for photography tips. While some of this may run on-device, a disclaimer notes that "when you use Camera Coach, the contents of the scene will be sent to Google for processing and then deleted."

Offloading some of the AI processing from your phone to Google's cloud servers isn't new or surprising. It comes with benefits, like saving your phone's battery life for things you actually need. However, you might want to consider what you share with Camera Coach for privacy reasons, even if the content is deleted from Google's servers eventually.

Camera Coach is turned off by default. When enabled, it uses AI models to analyze the scene you're trying to capture with contextual awareness. After it understands the scene, Camera Coach makes suggestions to aspects of the shot, like composition, framing, and lighting. All this is happening before you capture the photo, putting the photographer in complete control.

Camera Coach makes skilled photography accessible to all

I'm naturally adverse to letting AI creep into creativity and the arts, but let's be honest, photography isn't exactly an accessible craft. I own a Sony a6400 mirrorless camera for work and a bit of play, although I wouldn't call myself a great photographer. The cost of a great camera like that one is a huge barrier to entry, and learning how to use it is another major undertaking. It's one that I admittedly haven't spent the time to do yet.

This premise is what changed my mind about Camera Coach after initially being wary of the feature. Photography is a hard skill to learn, and not everyone has the time to learn it. We've all heard expressions like "it's the photographer, not the camera that makes the photo" — and they're all true. That's exactly why there's a need for something like Camera Coach to exist.

Smartphones already serve as excellent cameras, and they're so good that now the skills of smartphone users are what need to be improved, rather than specs.

AI might be a better photography instructor on the fly, using real-world scenes to teach Pixel 10 users about principles such as composition or lighting. Camera Coach can suggest changes like moving the subject or using optical zoom to the photographer's advantage. It can even come up with brand-new shot ideas simply from viewing your surroundings.

If you're a photography nerd, you've probably grimaced looking at photos captured with smartphones by your friends and family. You know that angling the camera lens a degree up or down could've made the photo so much better — but they don't. Camera Coach sure sounds like a solution to this problem, as it's effectively a photography instructor you can put in your pocket.

Camera Coach is only great if you learn from it

If Camera Coach makes photography more accessible to the masses, I'm all for it. I hate gatekeeping hobbies and skills, and capturing photos is tough to do right. I'd love to see Pixel 10 users try out Camera Coach for, say, a month. After that, those same users have ideally learned enough to start applying Camera Coach principles on their own.

I'd be lying if I said I wasn't worried about people using Camera Coach as a crutch rather than a learning tool. In that case, we could lose skills thanks to AI, rather than learn them. Camera Coach isn't inherently bad — in fact, it might be great — but it's up to us to use it the right way.

OpenAI has the clear brand advantage in the AI race, and that's something that may prove impossible for competitors to overcome. ChatGPT has more weekly users than Meta or Google has on a monthly basis. Being first comes with benefits, which OpenAI is certainly reaping.

In the time since, though, the competition caught up — at least on a functional level. AI models from Google, DeepSeek, Claude, and xAI have all held top-five spots on LMArena's leaderboard, signaling that the race from basic large language models (LLMs) to artificial general intelligence (AGI) is far from decided. For months, I've been certain that Gemini 2.5 Pro is a better and more versatile thinking model than anything coming out of OpenAI.

Well, OpenAI is hoping to change that narrative with the release of GPT-5, which launched for everyone last week (Aug. 7). It's actually a collection of models that can intelligently pick which one is best for a given prompt on the fly. And at least for now, it's at the top of both the LMArena and WebDev Arena leaderboards — two crucial LLM benchmarks.

GPT-5 doesn't solve all of ChatGPT's hallucination problems, and it's definitely not AGI. There are some areas where Google models still outperform comparable ones from OpenAI. Regardless, GPT-5 looks impressive and free, and that might be all OpenAI needs to maintain its hefty lead over the competition.

Where OpenAI's GPT-5 beats Google's Gemini 2.5 Pro

GPT-5’s smartest feature has nothing to do with compute or its knowledge base. It’s called real-time routing, and it helps pick the right model for your task without any additional user input. Currently, AI chatbots like ChatGPT and Gemini are littered with a variety of new, old, and experimental models best suited for a specific prompt. That’s great, but the onus was on the user to decide which one is best. With names like GPT-o3 and GPT-4o, or Gemini Flash Thinking Experimental, making that decision wasn’t always easy.

Instead, GPT-5 is a simple name for a collection of OpenAI models. It includes a lightweight model for quick, easy prompts and a more thoughtful model called GPT-5 Thinking for complex queries. OpenAI’s real-time router means that all of these models function as one, at least to the user. After inputting a prompt to ChatGPT, the GPT-5 router will decide which one to use, streamlining the UX.

There are still ways to manually control which GPT-5 model responds to your prompt. You can use phrases like “think really hard about this” in your prompt to trigger GPT-5 Thinking. If ChatGPT mistakenly opts for the thinking model, users can tap Get a quick answer to pivot to the lightweight model instead. Both models appear to be very reliable in early testing, clearly citing sources to avoid hallucinations.

These online sources, which include sites like Android Central, are a key reason why hallucinations are down with GPT-5. OpenAI says that with web search enabled on GPT-5 prompts, the model is about 45% less likely to contain a factual error than GPT-4o. Hallucinations are by no means extinct, but they are less prevalent when using GPT-5 in ChatGPT.

Independent tests and benchmarks seem to align with OpenAI’s claims that GPT-5 is better at writing, coding, and health-based tasks. It finally usurps Gemini 2.5 Pro on both the LMArena and WebDev Arena leaderboards, claiming the top spot overall. Specifically, GPT-5 seems to have the edge in text-based and coding prompts. I used a sample prompt from OpenAI to test GPT-5’s coding capabilities in ChatGPT, and came away seriously impressed.

Where Gemini still beats ChatGPT

GPT-5 isn’t better than Gemini 2.5 Pro in every area, and OpenAI still doesn’t seem to be able to match Google’s image and video generation. That’s reflected in LMArena’s text-to-image, text-to-video, and image-to-video benchmarks — Google’s Imagen 4 and Veo 3 tools sweep the graphical suite of generative AI tests.

I wanted to test that outcome in the real world, so I fed ChatGPT and Gemini the same prompt: "Generate an image of Johnny Thunderbird holding up a Big East Tournament trophy at Madison Square Garden." Here’s how the images turned out:

Anecdotally, the generated images align with the benchmark results — Google’s video- and image-generation chops are much stronger than OpenAI’s. For starters, Google gave me a generated image in about 10 seconds, and I had to wait nearly two minutes for an image to come back from ChatGPT.

From an accuracy standpoint, Gemini won handedly. It knew I was talking about my alma mater’s basketball program from the mascot reference, and correctly recreated Madison Square Garden with a ton of detail and flair. Meanwhile, ChatGPT produced an image with the wrong team and I’m not convinced the generic background is really MSG; it looks like it could be any basketball court.

So, if you’re the kind of person who likes to switch between LLMs based on your task, perhaps choose GPT-5 for writing or coding and Gemini for image or video generation.

GPT-5 is great, but GPT-4o just won't go away

A wrinkle in OpenAI’s GPT-5 rollout is user attachment to older models. GPT-5 was more than two years in the making, as OpenAI reportedly struggled to make an AI model worthy of the title. As it turns out, users didn’t want one. OpenAI is bringing back GPT-4o — a model that was supposed to be sunset in favor of GPT-5 — because of user backlash (via Tom’s Guide).

It might seem silly for those of us who touch grass, but ChatGPT users have seemingly created parasocial relationships with OpenAI models. They’d rather use the AI model they “know” over GPT-5, even if the newer one is better in every way.

I think you can extrapolate this premise to the broader AI race. In some ways, it really doesn’t matter whether Google, OpenAI, Claude, or DeepSeek has the best-performing model. People are going to stick to the models they like and are familiar with, and if that’s the case, OpenAI’s lead in this space may be insurmountable for the competition.

Google is finally challenging OpenAI in the broader market for AI, but most Android users care about the features that are available on their specific phone. That's where it gets tricky — each major Android manufacturer has its own set of unique AI features. On top of that, many phones tap into external sources, like ChatGPT or Perplexity, for a more versatile experience.

Currently, Samsung offers the best mobile AI experience. Galaxy AI is arguably the best mobile AI suite, and Samsung works with Google to ensure some of the most useful Gemini features are optimized for Galaxy phones. The "best of both worlds" situation is why I love using AI on Samsung phones, but Motorola is dangerously close to beating Galaxy AI at its own game.

Moto AI uses a mix of AI models and services, includes Motorola-exclusive features, and relies on custom hardware buttons to make accessing it all easier. On paper, that's all the makings of a standout mobile AI suite. In reality, the execution is somewhat questionable. That being said, if you're looking for a sleeper contender that can win the mobile AI race with the right moves, it's Motorola.

Motorola doesn't have a horse in the AI race — and that's a good thing

When it comes to mobile AI, the most important thing is giving users choice — how much AI they want to use, which models and services they prefer, and whether compute is handled on or off device. For some Android brands, that philosophy might conflict with their broader AI strategy. It's unlikely that Google is going to integrate third-party AI models natively in Pixel phones, as that would hurt the company's efforts with DeepMind, Gemini, and Google AI.

Motorola isn't trying to compete with Gemini or ChatGPT, and that gives it an advantage. Moto AI can use whatever AI model or service is best at handling a given task without needing to worry about exclusivity.

Motorola currently uses Google, Microsoft, Meta, and Perplexity natively as part of Moto AI. I don't know of any other Android brand that uses a greater variety of AI models and services to power its mobile AI suite than Motorola.

With this kind of utility, the user always wins. Moto AI picks the best model or agent for the job.

Motorola's Razr 2025 lineup was the first to integrate Perplexity in a phone, and it powers a Next Move feature that intelligently surfaces the AI features you'll need most in a given situation. Meta's Llama models are used to power Catch Me Up, the notification-summarization feature. And, you can "Ask Copilot" when using Moto AI to tap into Microsoft and OpenAI smarts.

If that wasn't enough, Moto AI's backbone is built on Google AI. You also get all the Google features that are available cross-platform on Android, like Gemini with live video and screen sharing support.

Moto AI's feature set is rock solid

Motorola also seems to be picking the right AI features to develop, because many of them are actually useful. Pay Attention is an audio-recording feature that can transcribe and summarize meetings, conversations, and more on your smartphone.

I recently tested an expensive dedicated AI recorder, so there's clearly demand for this kind of tool. Eventually, I came to the realization that it could've just been an app. With Moto AI, it effectively is one, and it can help you pay more attention in meetings or calls.

Remember This is another handy note-taking tool. Instead of audio recordings, it's intended to help you capture photos or screenshots and write text-based notes on the fly. You can snap a picture of a piece of paper you don't want to lose with a written note to help you remember what it is. I wish you didn't have to ask Moto AI's universal search for help recalling saved memories, but the concept is still helpful.

Finally, generative AI-based photo editing is a bit controversial, but it is solid with Moto AI. Magic Canvas will create AI-generated images from scratch, and Group Shot is Motorola's version of Best Take — it captures multiple photos to ensure that everyone in the photo has their eyes open in the resulting image.

The Moto AI suite is thorough but not overwhelming, partially because features like Next Move help show only the tools you'll need in a given situation.

Moto AI only needs a few tweaks to be a game-changer

At the same time, Moto AI is a bit rough around the edges. The inclusion of the AI Key on the Motorola Razr Ultra 2025 was an excellent move, and it's already available on other Motorola phones. Hardware buttons are an important part of mobile AI, because having the power button responsible for too many actions is a recipe for disaster.

The problem is, the AI Key in its current state is too limited to be all that useful. Moto AI doesn't start listening when you trigger the AI Key — you need to also tap the mic button on the screen, which kind of defeats the point. You also can't remap it to Gemini or specific features aside from Remember This and Pay Attention. It'd be nice if it worked more like triggering a classic assistant, where Moto AI would start listening immediately.

The AI Key is a microcosm for the Moto AI suite as a whole. With a few small changes, the AI Key could be incredibly useful and a differentiator for Motorola phones.

The same is true for Moto AI at large. If Motorola made it as easy to recall Remember This memories as Nothing's Essential Space, I'd love it. If Motorola borrowed Samsung's toggle for only running AI features on-device, it would address the needs of privacy-conscious users.

My wishlist of things I'd like to see Motorola improve is quite small when you think about it. The company could tweak all of these aspects of Moto AI in a software update. Considering Moto AI only exited the beta testing stage relatively recently, I'm optimistic Motorola is heading in the right direction. If it plays its cards right, it could offer the best mobile AI suite on Android.

Almost every phone on the market is bundled with AI features in 2025, but Samsung might offer the most well-rounded AI suite of any brand. When you buy a flagship Samsung Galaxy device, you get exclusive Galaxy AI features, fine-tuned multimodal Gemini support, and even more Google AI features.

Considering that Galaxy AI does a few things better than any other phone maker, Samsung users have an edge that those rocking the other best Android phones don't. One UI 7 brought some great AI tools to the Samsung Galaxy S25 series earlier this year, and now, One UI 8 is doing the same for brand-new Samsung foldables.

I've spent the last few days testing the Samsung Galaxy Z Flip 7 in and out, and came up with five AI features I actually found myself wanting to use. Some are new and some are old, but you should be aware of every single one of them if you're buying a new Galaxy Z Flip 7 or Galaxy Z Fold 7. In fact, many of them are available on Samsung Galaxy phones at large.

Gemini Live

Flip phones didn't have the best support for multimodal Gemini Live before Samsung released the Galaxy Z Flip 7. You couldn't use screen sharing or live video streaming features on the Galaxy Z Flip 6, and while you could do it on Motorola Razr phones, the experience was quite terrible. With little optimizations and buttons hidden behind the camera lenses, it's tricky to even start a live video session on my Motorola Razr 2025 cover screen.

Meanwhile, the new Galaxy Z Flip 7 supports a version of Gemini Live that's fully optimized for the cover screen. You can't use screen sharing for obvious reasons — the 4.1-inch cover screen is too small for it to be useful — but live video sharing is perfect. After testing multimodal Gemini Live on phones from every major Android brand, this implementation of it might be my favorite yet.

Being able to clutch the small flip phone in your hand while talking with Gemini Live is nice, and the form factor helps ensure you always have a comfortable grip. It's easier to just hold your arm outward to show Gemini Live what you're looking at, and flip the phone around to show yourself or listen to a response when needed. You could also prop the Galaxy Z Flip 7 up in tent mode while working with your hands.

The versatility of using Gemini Live on a flip phone makes it better than using a slab phone in my book, which are heavy and cumbersome to wield while doing other things. If we're talking about flip phones specifically, no other model has optimized Gemini quite like the Samsung Galaxy Z Flip 7.

AI Select

AI Select might be the most underrated Galaxy AI feature. You'll find it in the Edge Panel — swipe from the top-third of the screen to open it. After that, AI Select will be the first option from the top. You can use it to select basically anything on your screen with artificial intelligence, and get quick access to context-based options.

For example, you tap an image on your screen to download just that photo to your device without having to manually crop out the rest of your screen. Or, if you're looking at a flyer, you can use AI Select to automatically add the event to your calendar. Other options include selecting text, creating GIFs, translating languages, or generating something new with Generative edit and Drawing assist.

Actions that would be tedious on an average smartphone, such as creating a GIF from any video on your screen, are made simple. It's like Circle to Search, but instead of searching the web, you can tap into all the Galaxy AI features on your device.

Object Eraser

Almost every phone offers a generative edit and object removal tool, but in my testing, Samsung's Object Eraser is the most impressive. Apple's Clean Up and Google's Magic Editor try to do the same thing, but every time I try using Object Eraser, I'm shocked about how well it uses generative AI to remove something in a photo and fill in the gaps.

To test Object Eraser on the Galaxy Z Flip 7, I took a photo of myself holding a Galaxy Watch 8 and tried to remove it. This was a particularly tricky situation, as it not only had to fill in the gaps in my hand, but also the striped and textured carpet with various colors in the background.

There's a dropoff in sharpness and texture in the parts of the image where Object Eraser used generative fill, but it's hard to tell that at first glance. This extreme example is proof that if you need to remove something in the background of an image, your Galaxy phone is up to the task.

To try it, open a photo in the Samsung Gallery app, tap the Galaxy AI button in the center of the toolbar, and select or circle anything you want to remove. It's that simple.

Interpreter

The form factor of a foldable is uniquely positioned to help solve certain problems, and language barriers are one of them. With Interpreter on the Galaxy Z Flip 7, you can use AI to quickly converse with someone who speaks a different language than your own. This feature leverages both screens, and when the Galaxy Z Flip 7 is on a table, both people can see and hear a translation of what the other person is saying.

This feature isn't new to the Galaxy Z Flip 7, but the bigger Flex Window cover screen definitely makes it more enjoyable to use when sitting across from someone. It's certainly a Galaxy AI feature you should keep in your portfolio. Beyond the cover screen, you can also find Interpreter directly in the call UI for use during phone calls, as well as in the Edge Panel.

Call Assist

There's one AI feature that I truly can't live without, and that's Call Assist. I use it or a comparable version on every device I test to screen calls and avoid wasting time with robocalls or scammers. Straight from the Call UI, you can trigger Call Assist. From there, you can utilize the Live Translate feature we just covered, or use my favorite option: text call.

With text call, Galaxy AI will convert the person's words on the other end of the call to text, and give you the chance to reply — also with text. This is handy when you don't know whether you should pick up, but you also don't want to miss anything important.

I always use features like Call Assist when I'm expecting an important call from a business or unknown number, and simultaneously don't want to pick up a spam call by mistake. It's already come in handy a few times while using the Galaxy Z Flip 7 as my daily driver.

Plus many, many more

Those are just a few of the Galaxy AI features I find myself using on the Samsung Galaxy Z Flip 7, and the good news is, they're on a lot more Galaxy devices too. To close this list out with an honorable mention, Samsung still gives users the option to only process their data on device when using Galaxy AI features. When I reviewed the Galaxy S25, I wrote that every Android flagship should copy this toggle, and I'm glad to see it on the Galaxy Z Flip 7 as well.

Gemini is great, but Samsung phones like the Galaxy Z Flip 7 take things a step further with Galaxy AI features you can only get on Galaxy. Meanwhile, most of Google's best AI features are crossplatform. Not everyone cares about mobile AI, but if you do, there's a case to be made for choosing a Samsung Galaxy phone over the competition.

Whether we like it or not, artificial intelligence has changed our lives in just a few short years since ChatGPT brought it mainstream. The AI revolution has mostly shaken up the software world, with specialized AI hardware devices flopping — we haven't forgotten the disasters that were the Humane AI Pin or Rabbit R1 just yet.

When a new AI device pops up, contrarians are quick to point out that it "could've been an app." Sometimes that's true, sometimes it's not. AI voice recorders are particularly gaining traction among those who want new bleeding-edge hardware, and Plaud is leading the charge. The company's Note and NotePin devices aim to go with you daily to record conversations, meetings, and notes, transcribing them later in the Plaud app.

I wore the Plaud NotePin on my wrist for a week to figure out whether it's really a helpful companion, or an overpriced AI product that, well, could've been an app. Here's how it went.

Plaud nailed the NotePin's design and form factor

I wasn't impressed by the respective form factors of the AI Pin or the Rabbit R1. They looked too bulky and inconvenient to wear or carry daily. With these kinds of devices, they're only useful if you actually want to carry them. An AI assistant or voice recorder is of no use if you forget it at home or it's buried in a bag.

Luckily, Plaud struck the right balance of utility and compactness with the NotePin. It's a small, pill-shaped device that can conform to be worn in a few different form factors. There's a smartwatch-like wristband, a lanyard necklace, a clip, and a magnetic "pin" option included in the accessory kit. It's worth noting that only the pin and clip are included with the NotePin — the other items will cost extra.

I wore the wristband most often, as it's incredibly comfortable and unobtrusive for anyone used to wearing watches daily. It might look a bit strange wearing a NotePin on one wrist and a smartwatch on the other, but I did it anyway.

With up to 40 days of standby time and up to 20 hours of continuous recording, you should be able to wear your NotePin for plenty of time in between charges. I found those battery life estimates to be very generous — you will probably need to charge more often than that. But with only two hours needed to recharge, it won't be much of a hassle.

Plaud's pricing model is extremely hard to justify

AI hardware products are extremely difficult to price out. Devices that rely on cloud servers incur ongoing costs for the companies supporting those off-device features. While subscriptions aren't fun, I actually think they're necessary for AI hardware. A product being sold with a one-time, upfront cost is almost certain to be bricked in the long run — the economics almost necessitate that.

Plaud went with the worst of both worlds for its NotePin pricing model. It has an extremely high upfront cost for what it is, retailing for $159. On top of that, you'll need to pay for a subscription to unlock all AI features beyond the initial free trial period. Plaud's AI Pro plan costs $100 per year and still comes with rate limits. The AI Unlimited tier costs $240 per year for the full, unabridged experience.

To put it lightly, this pricing model is insane. With the unlimited plan, your first year of using the NotePin would cost $399, and that's before any extra accessories. Using it beyond that would incur greater subscription fees over time.

The subscription model isn't the issue. The problem is that you can't ask customers to pay extreme hardware costs and subscription fees. Plaud should've followed in Whoop's footsteps, including the hardware for free while charging monthly for its subscription plan.

The recording and transcription features are fine, but not outstanding

It doesn't help that I wasn't blown away by Plaud's ability to record my conversations and transcribe them. The Plaud NotePin didn't do a bad job by any means, but it wasn't outstanding, and it needed to be to justify its cost.

The way it works is simple: you long-press the NotePin to start recording, press again to stop recording, and let the recording upload to the Plaud app on your Android phone. After that, it'll need to be uploaded to the cloud for processing and returned to your device.

Like anything dependent on the cloud, this will take time. While uploading and transcribing your recording, the app says it'll take "a few minutes to complete" and that it will "continue if you navigate away from this page." So, I closed the app for about 10 minutes, let Plaud do its thing, and checked back in.

Instead of seeing a summary, I received a new message stating, "The network is unstable, it may take a little longer than expected." That was unusual, so I ran an internet speed test on my phone. Sure enough, my download speed was 891Mbps at the time that message appeared, which is more than fast enough for what the Plaud app was trying to do.

It felt like Plaud was trying to pass the buck on the user for its slow processing times, since my Wi-Fi 6E connection certainly wasn't the issue. The 10-minute waiting period for a roughly 30-minute conversation to process was frustrating enough, but playing the blame game was even worse.

To be clear, it wasn't all bad. Most transcriptions took about three to five minutes to process, which still isn't fantastic but is passable.

Additionally, the Plaud app provides more detailed transcriptions than any voice recorder app I've used. It creates an organized summary of the conversation and identifies the name of each speaker.

Beyond that, you get a mind map showing the flow of the entire conversation, plus an outline with isolated topics you can jump to at any time. Plaud also has an AI chatbot that can answer specific questions about your voice recordings, which some users might find really helpful.

It's not a matter of if — Android apps CAN do it better

In this case, it does feel like the Plaud NotePin could have been an app, as clichéd as that might sound. Something that uses on-device processing, like a Gemini Nano model, would be infinitely better than Plaud's reliance on the cloud. It'd be more private, more secure, and generally more helpful.

In fact, that's exactly what the Google Recorder app does on Google Pixel phones. Google switched to using Gemini Nano on-device to process recordings in 2024, supporting real-time transcription and processing of long recordings. While the Recorder app doesn't natively support a chatbot function, you can upload recorded files to Gemini to get transcriptions, summaries, and contextual conversations.

While testing the Plaud NotePin with the Google Recorder app, I vastly preferred using my Pixel 9 Pro XL for recordings and transcriptions. For one, I might have to wait as long as 10 minutes for the NotePin to transfer, upload, and process recordings. Meanwhile, the Pixel 9 Pro XL was transcribing my conversation instantly while I was speaking.

It's true that the Google Pixel 9 Pro XL is expensive, but the phone you already have in your pocket probably supports this feature. Apple, Google, and Samsung all offer apps that do the same thing, and they're included with your phone for free. You could buy a refurbished Pixel 7a for under $200 to get the Google Recorder app for much less than a full year of paying for the Plaud NotePin and its unlimited subscription plan.

I've been a loyal Google Recorder user since 2022, when I started using it alongside the Pixel 7 series. After switching to the Plaud NotePin for a week, I don't feel a strong incentive to replace it with a Pixel for my recording needs. With a Pixel phone, Google Recorder, and Gemini, you can get a similar experience with no subscription.

OpenAI made artificial intelligence mainstream with the release of ChatGPT in 2022. AI and machine learning were around well before then, but ChatGPT's emergence marked a turning point. Soon after, AI would become everywhere — on our phones, in our favorite apps, and part of our most-used services.

Whenever something new pops up, like ChatGPT, it takes a while for researchers to learn its true effects on our minds and daily lives. That process is only beginning, but an early study from MIT's Media Lab highlights some concerning trends regarding AI usage. It hasn't been peer-reviewed yet and has a small scope, including 54 participants aged 18 to 39 years old from the Boston region.

The results found that people regularly using ChatGPT to write essays had the lowest brain engagement of the groups studied and "consistently underperformed" at all levels compared to the other groups, which used either Google Search or only their own brains to write essays. The study demonstrates "the pressing matter of a likely decrease in learning skills" resulting from LLM use, per the report.

If you use AI tools, that probably sounds scary. You might be wondering whether using generative AI and large language models (LLMs) regularly puts your critical thinking skills at risk. While there's reason for concern, using AI for the right things probably won't be as dangerous to your mind as you might think.

What the MIT study found

First, it's important to understand the scope of the study and what its results actually mean. The team of researchers at MIT wanted to figure out how using AI to write an essay affected participants' learning and brain function. The study contained three groups: a brain-only group, an LLM group, and a search engine group.

The participants were asked to write three essays, one per session while remaining in their group. Those who were in the brain-only group could not use any external tools to write their essays in any of the sessions, for example. Then, in the fourth and final session, LLM users were asked to use only their brains to write an essay, and brain-only users were permitted to use LLMs — 18 participants took part in this final test.

All the while, researchers measured brain activity across 32 regions using an electroencephalogram (EEG), a device that records electrical signals firing in the brain. The tests aimed to record not only how cognitive function behaved during essay writing while using brain-only, LLMs, or search engines but also how it changed following prolonged LLM usage.

Aside from measuring brain activity with EEGs, the researchers conducted interviews with participants after each session. They also used human teachers to score essays while having an "AI judge" evaluate the works.

The results painted a clear picture: brain activity and engagement decreased with the amount of external tools needed. The study found that the brain-only group had the strongest level of brain connectivity, with the search engine group displaying intermediate levels and the LLM group performing the weakest.

When the LLM users were asked to use only their brains to write the final essay, the participants "showed weaker neural connectivity and under-engagement of alpha and beta networks." In contrast, brain-only groups who were then permitted to use an LLM for the final essay "demonstrated higher memory recall, and re‑engagement of widespread occipito-parietal and prefrontal nodes."

In layman's terms, those who used an LLM to write the essays produced lower-quality work, couldn't recall what they had submitted, and had lower brain engagement. Additionally, those who had relied on LLMs before being asked to only use their brains performed worse than the brain-only group.

"The use of LLM had a measurable impact on participants, and while the benefits were initially apparent, as we demonstrated over the course of 4 months, the LLM group's participants performed worse than their counterparts in the Brain-only group at all levels: neural, linguistic, scoring," the study found.

What does this mean for you, an AI user?

The study's results are yet to be peer-reviewed, but let's face it, they aren't exactly surprising. Like any other muscle, your brain needs to be used regularly to stay in shape. If you stop using your brain to write essays, reports, emails, or text messages often, those skills will start to deteriorate.

People who often vouch for AI in learning and work environments often compare chatbots and LLMs to a calculator. When the calculator first became prolific, teachers wanted it banned. Now, calculators are allowed to be used in educational settings and are a vital part of our everyday lives — they're on our phones, computers, and even smartwatches.

I do think generative AI is like a calculator, but maybe not in the way its advocates think. As a writer, I haven't used complex math regularly since my early college days, when I was getting my required math courses out of the way. Since then, a calculator has been my best friend, and guess what? I'm not as good at mental math as I was before.

Similarly, I haven't regularly handwritten words in years — I'm an all-digital person now and have been for some time. When I get handed a form at the doctor's office, my penmanship isn't nearly as good as it once was, and I get hand cramps pretty quickly. All that is to say, the MIT study's findings are really just reinforcing common sense. If you create a dependency on artificial intelligence and stop using a particular skill, you will absolutely lose it.

To me, that's not as scary as it might sound. It all depends on how you incorporate AI into your daily life. Want to use generative AI to make a wallpaper or a cool image to share with your friends? Want Gemini to replace Google Assistant as your voice-based helper on an Android phone? Need it to analyze long documents? I don't think any of these use cases present a significant risk of harming your cognitive abilities.

When you use AI to do something you can do yourself, you're at risk of losing those skills. If you use ChatGPT to write all your emails, documents, and reports, there is a good chance your skills will start to fade away — that's essentially what the MIT study found.

But that's not how I use artificial intelligence, and I don't think that's how most people use it, either. I use AI to do things I couldn't possibly do, and that's why it's helpful. That includes analyzing and summarizing documents that are thousands of pages long or generating an image that I couldn't create because I'm no artist. In those instances, AI is expanding your capabilities, not replacing them.

If you look at AI as a way to elevate your possibilities rather than replace the things you'd rather not do, you probably don't need to change your AI usage habits. However, if you want to use AI to eliminate certain parts of your workflow, think about whether you're willing to lose those skills in the process.

The Gemini app serves as the central home for Google's AI efforts on iOS and Android, but the company changed things up last month when it introduced a standalone NotebookLM app on mobile. Considering that NotebookLM started out as a limited research project in 2023, earning its own mobile app is a big milestone for the AI-powered study tool.

Although the Gemini mobile app has almost completely replaced the web-based equivalent for me, I was skeptical that this would be the same for the NotebookLM app.

Despite the name, NotebookLM isn't a note-taker like Google Keep or Notion. It's more of a research and study tool, capable of using AI to parse and converse about complex topics with a lot of specific context.

As it turns out, I was both right and wrong to be skeptical of NotebookLM's Android app. It's going to be a helpful companion for existing NotebookLM users, but it likely won't be a great first impression of the service for new users. The Gemini-powered research tool is simply better suited for the desktop — let me explain why.

What's good about the NotebookLM mobile app

Let's start with a high-level overview of what NotebookLM does. It takes a collection of sources about a topic to form a "notebook." Each of these sources can contain up to 500,000 words, and you can put 50 of them into a single notebook. This gives your NotebookLM project its own context based on the sources provided, whether they be websites, YouTube videos, PDF files, or your own notes.

After you've selected your sources, a Gemini model within NotebookLM gets to work. It's not always clear which Gemini model is at work — Google seems to update it often, but we know it's at least a tweaked version of Gemini 2.0 Flash Thinking. Considering the Gemini 2.5 series is already stable and public, it's possible the model working behind the scenes has been updated.

From there, NotebookLM can do a lot — including chat with you based on the context of your sources, provide summaries, generate spoken Audio Overviews, and more.

The mobile version of NotebookLM works best with notebooks you've already created. After opening the app for the first time, you're greeted with a list of your existing notebooks alongside the option to create a new one. With sources already uploaded, one of your current notebooks is ready to go for text-based chats or spoken Audio Overviews.

Audio Overviews are the best part of the NotebookLM app, as they can be downloaded to your device for offline listening. Of course, this takes a while — the generation process for an Audio Overview takes around 5-10 minutes, and the download takes additional time on top of that. It's such a lengthy process that the app reminds you that you can close it and do something else while an Audio Overview is created. And sometimes, it'll fail entirely.

But when it does work, it's pretty glorious. You can listen to the podcast-style Audio Overview based on your study topic while driving, taking the train, or doing just about everything else. Downloaded content helps eliminate some of the stutters you might experience while trying to play an Audio Overview on a mobile network. You can also "join" an Audio Overview to get your podcast hosts on the exact topic you need help with, which is neat.

Audio Overviews are perhaps the most popular feature included with NotebookLM — so popular that they've since been added to the Gemini app and an experiment in Search Labs. Really, it feels like the NotebookLM mobile app for iOS and Android is just a player for Audio Overviews, because aside from the chat function, there's not much else this app can do at the moment.

That doesn't have to be a bad thing, but it's important to keep your expectations tame if you try this app as a potential replacement for your favorite note-taker.

What's bad about the NotebookLM mobile app

The limitations of the NotebookLM mobile app compared to the web client are why I kept wanting to use the fully-featured desktop experience instead. One of the best NotebookLM features added recently is called Discover sources, and it effectively takes the work out of your hands when creating a new notebook. You'll type in a few keywords, like "Algebra I" or "The American Revolution," and NotebookLM will find good sources on the web and automatically add 10 of them to a new notebook.

There's also an I'm feeling curious button, which is a nod to Google Search's I'm feeling lucky button. As you might have guessed, this tool picks a random topic, finds sources about it, and adds it to a new notebook. It's perfect for the curious-minded person who just loves to learn new things.

I was excited to use this feature in the NotebookLM mobile app, but I quickly found out that it wasn't there. It's quite disappointing because the process of adding sources to a notebook using the Android app is tedious. It involves a ton of window-switching and copying that isn't suited for a small screen, even if the NotebookLM shortcut in the Android share card is handy.

Put simply, it feels like a missed opportunity, and it isn't the only one. You also can't create a note on the NotebookLM mobile app. On the web, you can make a brand-new, text-based note in a notebook on the fly and later convert it to a source. This feature is notably absent on the mobile app, and it severely limits its capabilities as an Android note-taker.

For now, the NotebookLM web client isn't going anywhere

To be fair, Google is aware of these limitations. In its introductory blog post, the company said that "this first version of the app includes many of the core features of NotebookLM — with more refinements and features to come in the months ahead." I'm hopeful that in the future, the NotebookLM app will become on par with its desktop counterpart.

For now, though, the NotebookLM app on iOS and Android isn't that useful. It's the only way to download Audio Overviews to your mobile device, so if that feature sounds appealing to you, it's worth using. Otherwise, you should probably stick to the web client — that's what I'm going to do.