I’ve spent the last six weeks daily-driving three different flagship phones from three different manufacturers, all of which heavily advertise their on-device AI capabilities. I switched SIMs around like a maniac, used each phone for two full weeks as my actual main device, and tried every advertised AI feature at least a dozen times under real-world conditions.

The conclusion isn’t going to make the marketing departments happy. Most of the on-device AI features being pushed in 2026 are mid at best. A few are genuinely useful. A handful are actively annoying. Here’s the honest breakdown of what’s worth caring about and what’s not.

What “on-device AI” actually means now

Quick technical context because this matters. When manufacturers say “on-device AI,” they generally mean that some specific AI features run on the phone’s neural processing unit rather than sending data to a server. This has two real benefits — privacy (your data stays on the phone) and latency (no round trip to a data centre).

The catch is that on-device models are smaller and dumber than the cloud models the same companies offer. A photo edit that runs on-device uses a model in the low billions of parameters. The same task in the cloud might use a model fifty times larger and produce noticeably better results. So when you see an AI feature advertised, the first question to ask is which version of the model you’re getting.

In Australia we have an additional wrinkle: some features are gated by region and a few that work brilliantly in the US are either turned off or significantly limited here. Always check before you assume.

The features that genuinely work

A short list of on-device AI features that I’d genuinely call useful after extended testing:

Real-time call transcription. Surprisingly good across all three phones I tested. The accent recognition has gotten meaningfully better in the last year — I’m getting around 95% accuracy on transcription of Australian English with a few regional accents thrown in, which is dramatically better than the 70-something percent I was getting on the same task in 2024.

Live translation. This one’s genuinely transformative if you’re travelling or working with non-English speakers regularly. The on-device translation is now fast enough to feel real-time and accurate enough to be trustworthy for casual conversation. Don’t trust it for medical or legal contexts.

Photo magic eraser / object removal. Yes, every manufacturer has a version. Yes, they all work. The differentiation between phones is smaller than the marketing suggests but the feature itself is genuinely useful — I use it weekly to clean up photos that would otherwise have been deleted.

Voice recorder summarisation. If you record meetings, lectures, or interviews, the on-device summarisation has reached a level where it’s saving me real time. Caveat: summaries get specific names wrong with some regularity, so verify before quoting.

Adaptive battery management. Less flashy but probably the AI feature that’s quietly making the biggest difference. The on-device model learns your usage patterns and shifts background activity to match. End-of-day battery life on all three phones I tested is meaningfully better than equivalent hardware would have been two years ago.

The features that don’t quite work yet

A few that the marketing oversells:

Generative photo composition. All three phones I tested offer some form of “AI completes the photo” or “generative expand” feature. The results range from unconvincing to actively misleading. Anything more complex than expanding a blue sky tends to produce artefacts that anyone paying attention will notice. Use sparingly if at all.

On-device AI image generation. Slow, mediocre quality, and the prompt understanding is years behind cloud tools. Why this exists as a phone feature is beyond me. If you want AI-generated images, do it from the cloud tools on your laptop.

Real-time content awareness suggestions — the “you’re looking at a restaurant, here are reservations” style proactive suggestions. Three problems: they fire too aggressively, they’re often wrong about context, and they’re irritating. I turned this off on all three phones within 48 hours.

Voice assistant general intelligence. The on-device voice assistants are still pretty basic. They can set timers and send messages but anything more conceptual gets handed off to the cloud, at which point you’re back to standard cloud assistant behaviour with a slight delay. The marketing implying you have a brilliant on-device AI you can chat with all day is misleading.

What’s happening behind the marketing

Worth pulling on this thread because it explains the inconsistent quality you’re seeing across features. Phone manufacturers in 2026 are basically integrating two different things and calling both “on-device AI.”

The first is genuinely on-device — small models running on the NPU, handling specific narrow tasks like noise reduction, photo editing, transcription. These features work because the models are purpose-built for the task and they don’t need to be smart in any general sense.

The second is what one of the Team400 team once called “AI theatre” in a conversation — features that are marketed as AI but are mostly conventional software with a slight ML tweak applied. The “AI camera modes,” the “AI fitness coaching,” the “AI smart suggestions.” Most of these would have shipped exactly the same way in 2019 with different branding. The actual machine learning content is minimal.

Knowing the difference helps you ignore the marketing and focus on what’s actually new.

The privacy question

One thing the on-device pitch genuinely delivers on: privacy. When transcription is happening locally, that audio doesn’t leave your phone. When photo editing is happening locally, your images aren’t being processed on someone’s servers. That matters more in 2026 than it did five years ago, with the various regulatory and breach stories that have made the news.

Worth checking carefully though. Some features advertised as “on-device” are actually hybrid — partial processing locally, additional processing in the cloud. The fine print in the settings menu is where you find the truth. I’d recommend a 20-minute audit of your phone’s AI settings menu to actually understand what’s running where. You’ll likely find a couple of features that are sending more data off-device than you realised.

The Australian Privacy Foundation has been pushing for clearer labelling around this and the manufacturers have started to comply, but the disclosures are still buried in places most users don’t read.

Which phone wins on AI features

I’m not going to name specific phones because the field shifts quickly and any specific recommendation will be obsolete in six months. What I will say:

The differences between flagship-tier on-device AI features are smaller than the marketing makes out. If you’re picking a phone based on AI features specifically, you’re probably overweighting that input. Pick based on camera, battery, durability, ecosystem fit, and price-to-quality ratio. The AI features will be roughly comparable across all the major flagships.

Where the AI difference does matter is between flagships and mid-range. Mid-range phones in 2026 are still running noticeably weaker on-device models. The transcription, translation, and photo editing features that work brilliantly on a $1,500 phone are mediocre on a $700 phone. That gap will probably close in the next two years as cheaper NPUs improve, but for now it’s real.

What I’d actually use

If I had to pick the three AI features I genuinely use day-to-day: call transcription (constantly useful), live translation (genuinely transformative when you need it), and adaptive battery (invisible but real). Everything else is occasional novelty or pure marketing.

The next 18 months in this space will be interesting. The on-device models are getting bigger and the hardware is getting faster. Some of the features that currently disappoint will work meaningfully better by 2027. Worth checking back in then.