The pitch for on-device AI on smartphones has been steady for about two years now: your phone runs powerful language and vision models locally, your data stays on your device, and the experience is faster than cloud-based alternatives because there’s no round trip. The reality, as of May 2026, is more uneven than the marketing.

I’ve been using three current flagship phones — the Pixel 10 Pro, the iPhone 17 Pro, and the Samsung Galaxy S26 Ultra — for the past several months specifically to evaluate which on-device AI features get used in normal daily life and which ones turn out to be one-time demos. Here’s what I’ve learned.

The features that have actually become daily-use

Live translation during voice calls is the standout success across all three platforms. Google’s implementation on the Pixel is the most polished, but Samsung and Apple have both reached usable quality. The lag is around one to two seconds per phrase, the translation quality is good for the major language pairs, and the use case is unambiguously valuable for anyone who deals regularly with non-native speakers.

Photo organization and search has quietly become indispensable on every platform. Searching your photo library by description — “my dog at the beach last summer” — works well enough that I’ve stopped thinking about it as an AI feature and started thinking about it as just how phones work now. The on-device indexing means search is instant, even on libraries with tens of thousands of photos.

Visual search of objects in the camera viewfinder — pointing your phone at something and asking what it is — has gotten genuinely useful. The plant identification, product lookup, and text translation use cases all work well. The accuracy on plant ID specifically has improved significantly in the past year.

Audio transcription of meetings, voice memos, and ambient conversation is the fourth daily-use category. The on-device speech recognition on all three platforms is fast and accurate enough that I no longer use cloud services for routine transcription work.

The features that remain marketing

On-device generative image creation — the “magic edits” that all three platforms have promoted heavily — gets used occasionally but isn’t part of most people’s daily phone use, based on the usage data that’s been published. The quality is good enough for casual edits and the privacy story is clean, but the tasks people actually want done with image generation tend to involve more complex prompts and longer generation times than on-device models handle well.

On-device summarization of long documents and articles is technically functional but, in my experience, not actually used much by people. The cognitive overhead of opening a document, triggering summarization, reading the summary, and deciding whether to read the full document tends to exceed the cost of just skimming the document yourself.

The promised “AI assistant that knows your context” remains the perpetual just-around-the-corner feature. Apple’s restructured Siri, Google’s Gemini integration, and Samsung’s Bixby revival all promise this, and all three deliver something less than promised. The fundamental challenge is that “knowing your context” requires access to your messages, email, calendar, and app data in ways that the operating systems are still cautiously navigating.

The hardware story underneath

The reason on-device AI works at all on current phones is that the neural processing units in flagship chips have gotten genuinely capable. The current generation of mobile NPUs delivers compute performance that would have required a desktop GPU five years ago, and the memory architecture has been redesigned to allow language models to run efficiently in the available RAM.

The current sweet spot for on-device language models is around 3-7 billion parameters, with quantization to 4-bit precision. Models in this range can be loaded into RAM, can run inference at conversational speeds, and can handle a meaningful subset of useful tasks. They can’t compete with cloud-based 100B+ parameter models on complex reasoning, but they’re more than adequate for the narrow tasks they’re typically asked to do.

Battery impact has been the surprise constraint. Heavy use of on-device AI features measurably affects battery life. A day of normal phone use that includes thirty minutes of cumulative AI feature use will consume noticeably more battery than the same day without those features. The chipmakers have been working on power efficiency, but the gains haven’t kept up with the increase in model size and usage.

The cloud-versus-device tradeoff in practice

The reality of how AI features work on current phones is that almost everything is hybrid. Some operations run entirely on-device. Some run in the cloud and stream results to the device. Some start on-device and fall back to cloud for queries that exceed local model capability.

The fallback behavior is mostly invisible to users, which is both convenient and concerning. You don’t always know whether a particular query was processed locally or sent to a cloud server. The privacy claims of “on-device AI” can become technically true while functionally misleading if the query routing isn’t transparent.

For people who care about this distinction — and there are good reasons to, particularly in professional contexts where data sensitivity matters — the platforms have started providing better controls. iOS lets you disable cloud fallback for specific feature categories. Android offers similar controls but they’re harder to find. Samsung’s settings remain the most opaque of the three.

What I’d watch for the next year

Three things, mainly. First, the next generation of mobile NPUs (expected in late 2026 flagships) is likely to roughly double the practical model size that can run on-device, which will probably bring some currently-cloud-only features into the on-device tier.

Second, the operating system AI integration is going to keep getting more aggressive about looking at your personal data — messages, files, calendar — to provide context-aware assistance. How well this works will depend heavily on whether the platforms can deliver useful results without creating privacy problems.

Third, the regulatory environment is shifting. The EU’s AI Act enforcement provisions and Australia’s evolving framework will start to affect what features can be shipped without disclosure, particularly around user data processing.

The honest summary is that on-device AI on phones in 2026 is a real technology that delivers real value in narrow ways, and is also significantly more limited than the marketing suggests. For most users, the features that actually matter — translation, photo search, transcription — work well. The features that show up in keynote demos but don’t fit into daily routines remain demo material. That’s roughly what we should have expected, and it’s a healthier place to be than the inflated expectations of 2024.