Skip to main content
AI

Google’s Gemini for navigation is full of trivia, but can struggle with directions

Morning Brew took a few walking and biking trips with Google’s new AI assistant.

5 min read

Google’s Gemini navigation assistant promises “a knowledgeable friend in the passenger seat.” More recently, that “friend” can now bike or walk along with you, too.

I set out as both a pedestrian and a cyclist on both coasts to test Gemini’s value, and found it to be a chatty companion that was perhaps better at dispensing trivia than holding a map right side up.

Google first brought Gemini to driving directions in November, then expanded it to walking and cycling directions in recent weeks. The idea is to give travelers a hands-free, conversational way to gauge ETAs, ask about things like attractions or restaurants along a route, or just wonder aloud where exactly they are.

If you’ve given Google more digital permissions than I have, you can also have Gemini text a friend that you’re running late or check your calendar for upcoming events.

The rollout comes as automakers are increasingly looking for ways to integrate generative AI into in-car experiences—in some cases, with Google’s help. GenAI could provide a hands-free way to interface with directions as well as navigate all of the digital amenities that modern vehicles now offer. Why not offer the same functionality to walkers and cyclists?

For one, I felt a little foolish talking aloud to an AI while walking down a crowded city street. Maybe it’s just me, but I found the experience of asking my phone, “What neighborhood am I in?”—a query suggested by Google—to be a bit embarrassing when passersby were in earshot.

It certainly did know which neighborhoods I was in, however, whether it was San Francisco’s Mission District or the border of the Lower East Side and the East Village in Manhattan. Or even a much smaller neighborhood in a suburban Bay Area city. And its grasp of place-based trivia is extensive.

Walking down Delancey Street in Manhattan, I asked it, intentionally vaguely, about a nearby diner featured in a documentary. It identified the modern iteration of Shopsin’s General Store in Essex Market as the subject of a 2004 film, I Like Killing Flies. It could identify and tell me the history of various theaters in the Mission District. It successfully suggested local stores where I could buy a Valentine’s Day present, and identified convenience stores to buy particular items.

Map troubles

I’ve tried to use AI for biking navigation before, and my very cursory foray was fraught. Last summer, I tested ChatGPT’s ability to give me info on a route along the Empire State Trail through a remote part of the Adirondack Mountains. (Editor’s note: Patrick is extremely qualified to conduct these tests. He’s an avid runner and cyclist, routinely taking a spot on the podium for Morning Brew’s Strava Activity Challenges.)

Tech news that makes sense of your fast-moving world.

Tech Brew breaks down the biggest tech news, emerging innovations, workplace tools, and cultural trends so you can understand what's new and why it matters.

The guidance it gave me at the time sounded well and good at first, until I realized much of it was unmoored from any geographic reality. It told me a stretch of trail was 25 miles in one response, then 15 miles in the next. (It was actually 41 miles.) ChatGPT imagined roads with spacious shoulders and rolling terrain instead of much more real steep, narrow hills. This seemed to be one use of AI that was more prone to hallucinations (this was a whole generation of GPT ago, though, to be fair).

One would assume Gemini to be more tethered to Google Maps. I didn’t get it on any remote mountain trails, thanks to snow and frigid temperatures, but Gemini struggled to turn conversations into routes in Livermore, California, and in Manhattan.

In one case, I asked Gemini to tweak its directions to incorporate a nearby trail instead of street sidewalks. Gemini described the route I wanted to a T. But when Maps tried to alter the directions accordingly, it inexplicably added in 13 miles and a tour of the neighboring city of Pleasanton. It seemed that despite the trail stretching between the two cities, there was only one trailhead labeled as such on the map. Trying to get it to route along the correct trail by voice ended up being more trouble than it was worth.

At another point, Gemini and I talked through a plan for Valentine’s Day shopping. It dutifully identified nearby florists, but this was on February 8—too far from the actual holiday. It then crafted a route that would hit a jewelry shop and a chocolate store; however, I told it I didn’t want to go to that particular jewelry store, and it seemed to understand. We discussed a new route that would encompass a gift shop and a chocolate store. But when I set off, having been assured of updated directions, I found it was incorrectly routing me once more to the jewelry store I’d rejected. OK, I can take the hint.

These types of negotiations sometimes resulted in a desired route, but I ultimately found it easier to just pull over my bike or stop walking and type things in by hand. Maybe if you’re behind the wheel on the road, it makes sense. And if you don’t mind awkwardly conversing with your phone on the street, it can definitely offer you some fun facts on your immediate surroundings. But for now, I’ll be sticking to the classic Google Maps when I simply need to get from Point A to Point B.

Tech news that makes sense of your fast-moving world.

Tech Brew breaks down the biggest tech news, emerging innovations, workplace tools, and cultural trends so you can understand what's new and why it matters.