AI agents on my phone? No thanks, I still can’t trust a simple Google search

It’s funny how LLMs were supposed to make researching stuff easier. Yet, I find myself double-googling the LLM’s replies, just to make sure it didn’t hallucinate something.

Sure, don’t get me wrong — if my search is low-stakes enough, or simple enough, I can trust the quick AI blurb that I get as a response. But, when researching something, I am typically navigating multiple links and multiple search prompts, just to be triple-sure. And sometimes that’s justified, as I keep getting different answers, numbers, or logic in reply to my requests.

Is that better or worse than the old days, when we could find the answer to our questions on a random subreddit or Facebook group? I don’t know.

What I do know is that

Now is certainly not the time for AI agents to enter our smartphones

You’ve probably heard of that numerous times already. The industry’s next half-baked dream is to have on-device computing with “agentic AI” on our smartphones.

Technically, agentic AI was added to some devices in 2025. Samsung’s and Google’s phones are able to take cross-app action, right? Like “Find me the next team x football game and add it to my calendar”.

The Now Bar on Galaxy phones is also, technically, agentic — if useless.

Currently, AI agents are very basic, but there’s a push to increase their authority in 2026, to a point where they may be able to buy stuff for you. Like asking it “Find when Taylor Swift is performing nearest to my home location and book me two tickets. Make sure we have transportation and a place to stay the night”. And the AI agent will… presumably, hopefully, in theory, maybe go out and do it all for you.

We can all see the problem with this, right?

I find it increasingly difficult to believe that any reasonable person would blurt out such a request into an AI entity that… used to suggest that we eat rocks and drink glue just a few months ago. An AI that often mixes up years, events, or makes up entire stories and presents them as fact.

Sure, sure. You don’t get progress without taking a risk and pushing for that next big thing. We’ve all seen tech companies rush undercooked products before. Foldable phones, S Pens, smartwatches, they all have one or two red spots on their reputations, due to the pressure to always innovate and always be first.

And while the push for early adoption has always had a certain tint of absurdity to it, you could always see that tech enthusiasts could find something in it, something that maybe doesn’t work right but gives them a taste of what the future might be.

Most importantly, the tech we were rushing to before didn’t have the potential to mess up your entire schedule or cause financial harm.

Like, the first S Pen was pretty rudimentary and limited by the phone’s hardware. The current S Pen is the best stylus you can get on a phone. The first Apple Watch was a laggy, uncompleted mess. The current Apple Watch is the best-selling watch (not smartwatch) in the world.

These things didn’t function right, but early adopters enjoyed them because they could see the foundation they were building, giving us a glimpse into the future.

Can we see the foundation for agentic AI right now?

Legitimate concerns over AI agents

If a search engine gives you a wrong fact, you just close it and do the research yourself. If an AI agent has permission to book a non-refundable flight or send a sensitive email based on a "hallucination", the consequences can affect financials, professional or personal relationships, or at the very least — your schedule.

An AI agent will need… a lot of access — to your data, your device, your payment information. Yeah, we have all the Secure Enclaves and On-Device Processing promises in the world. Yet, we do live in a world that has already seen a few cloud data breaches, correct?

Not to mention that researchers are already seeing a new form of malicious attacks. Emails and websites can be “injected” with prompts that are invisible to users, but will be picked up by AI. And they may just instruct your assistant to do things you absolutely do not want it to.

Can the companies ever guarantee 100% accuracy?

Presently, I just can’t see it. Maybe call me a Luddite. But we’ve had, what… three years of AI meme farming behind us — and multiple subreddits with funny screenshots to document it all.

No matter which LLM you prefer, you are probably already well trained to catch it hallucinating and double-check your results. It happens more often than we’d like, right? And manufacturers and developers believe we will trust an AI agent that is supposed to interpret intent and take action autonomously? Sure.

User distrust in AI seems to be at an all time high. And I am not even going to dive into the negative sentiment built up by current “AI slop” trends, hardware price increases, artist struggles, and job replacement threats.

AI has an image problem. Now is absolutely not the time to give it more authority. That push seems absurd to me.

But hey, maybe just me. What’s the first prompt you would give to your autonomous agentic AI assistant?

Try Noble Mobile for only $10

Get unlimited talk, text, & data on the T-Mobile 5G Network plus earn cash back for data you don’t use.

Buy at Noble Moblie