I Relied on my Phone’s AI Features for a Week. Here’s What Worked

Over the past year, I’ve had AI features on my phone, including Gemini from the start. Every few weeks, I’d experiment with something new, hoping it would be helpful, only to return to the basics as the novelty faded.

While AI features are now integrated into most modern smartphones, many still seem like novelties that are tried once and quickly forgotten. I decided to immerse myself in AI features for a week, utilizing photo edits, voice typing, assistant requests, and writing suggestions.

I tested AI features on both Samsung and Pixel devices to understand the experience differences and identify which features are practically useful and which can be ignored.

I initially thought the hype around Pixel’s AI erasers was mainly due to social media, but I soon realized otherwise.

Within a day, I used it to remove an unwanted object from a vacation photo of my mother. The outcome was satisfactory, not flawless but significantly cleaner. My mother was pleased with the photo and didn’t notice the subtle AI alterations, which means the tool met expectations.

Pixel’s Best Take exceeded my expectations. After taking a burst shot of a group of five individuals, the final image was an ideal group photo. Although it’s not always clear how it chooses faces, the selections were apt, and the result didn’t appear digitally stitched.

Conversely, Samsung’s Generative Edit was overly aggressive, which was evident in the results. Attempting to remove a car from a street scene resulted in a visibly altered pavement. For anyone familiar with the original scene, the edits would be evident.

Nikhil Azza / Foundry

Conclusion: Samsung’s Generative Edit works for quick social media posts but lacks finesse on closer inspection. Pixel’s AI photo enhancements are more refined and dependable.

Writing and voice typing that actually held up

Keyboard assistant suggestions were inconsistent, but voice typing was notably effective. Over the week, I dictated about 40 messages on both devices. Gboard’s voice input managed background noise in a moving car surprisingly well. Its accuracy was high enough that I needed fewer edits than I would have through fast typing on a small screen.

Nikhil Azza / Foundry

Writing suggestions for work responses were more beneficial than expected. Quick drafts were improved without needing a complete rewrite. On two occasions, I used the rewrite option for urgent emails, and the outcomes surpassed my initial drafts.

Samsung Writing Assist took it a step further by rewriting an entire paragraph. I composed a blunt work message when tired, and Samsung transformed it into a more professional output, which I now use frequently.

Conclusion: While these features may not feel groundbreaking, they can make a challenging day a bit more manageable.

Assistant tasks that were inconsistent

This is where the week became uneven. Simple tasks like reminders, alarms, and quick voice notes functioned well on all phones, provided the requests were short and clear.

However, complex requests faltered. I instructed Gemini on Pixel to locate a Chinese restaurant open after 9pm on Friday, add it to my calendar, and text my colleague its name. Despite being a showcased multi-stage task, it only completed the first step, requiring me to handle the rest manually, which took longer than if I had done it all myself.

Samsung’s Bixby faced similar limitations: it handled simple, single tasks efficiently but slowed down with more than two commands. By the weekend, I had stopped attempting complex requests altogether.

Conclusion: Keep interactions simple or avoid them completely.

The background features you don’t notice until they’re gone

Call screening emerged as the most effective Pixel AI feature throughout the week, despite my initial reluctance to test it. It flagged seven calls as likely spam, accurately identifying six. The one exception was a legitimate call from my banker, which is an acceptable error rate.

Pixel’s battery optimisation also made a significant impact on battery life. By Saturday, I noticed ending the day with more charge than earlier in the week, despite similar usage patterns. The feature seemed to identify which apps I actively used and limited background activity for the rest.

Nikhil Azza / Foundry

Conclusion: These features stood out as the most valuable, functioning seamlessly in the background without requiring user intervention.

I had anticipated using Circle to Search as a substitute for many Google searches, but it underperformed. It provided incorrect results twice in a matter of days, leading me to lose trust in it by Wednesday. While it is useful for identifying products — its intended purpose — for more crucial searches, I reverted to typing.

Nikhil Azza / Foundry

Otter didn’t live up to expectations during the only recording that mattered that week. A 20-minute conversation in a coffee shop was marred by background noise, resulting in a transcript riddled with errors. A traditional note-taking method would have been more effective.

The Pixel Recorder performed well with a quiet 15-minute voice memo I recorded during a walk, highlighting the importance of location and noise conditions for successful transcription, which Otter struggled with under challenging conditions.

Conclusion: Proactive suggestions, such as app recommendations or shortcuts, primarily impeded me, leading to most of them being disabled.

What I’m actually keeping

Magic Eraser and Best Take will remain on my phone, along with voice typing for specific scenarios. Among all the features, Call Screening is the one I would miss the most if it were removed, and I also appreciate the unobtrusive battery optimization.

Samsung Writing Assist is helpful when I’m in a hurry, which is often the case, and Circle to Search is only useful for product identification, nothing beyond that.

By the week’s end, all other AI features had been turned off or forgotten. This isn’t necessarily because they are ineffective but because they didn’t offer any tangible benefits for me. The tools that endured were the ones I didn’t have to think about, which is likely the best indicator of their effectiveness.

While companies aim to make these AI features essential, after a week of reliance, my conclusion is that most are not yet sufficiently useful or precise.

Read next: Better than AI? The four improvements we want instead.