I’ve Been Using Fitbit’s AI for a Week, and I’ve Never Had a Worse Fitness Trainer.
The new version of the Fitbit app (currently in public preview) has a number of improvements and flaws . On the first day of its release, the built-in AI confused and frustrated me , but I decided to give it a little time. Perhaps it will learn more about me and provide more accurate answers. Perhaps I’ll figure out how to use its features. Perhaps the Google team will fix the most serious issues.
The latter prediction partially came true. When I woke up the next morning, the AI finally recognized that my watch, Google’s Pixel Watch 4, actually existed. (The first version of the app claimed that the Pixel Watch 2, released in 2023, was the newest model.) I also noticed a few minor fixes that appeared over the course of the week.
I used the app for a week, wearing the Pixel Watch 4 at night and during most of my workouts, and chatted with the AI coach almost every day. Several times, I asked it to adjust my goals and suggest new workouts to see how it handled different types of training.
I hold several coaching certifications myself, including as a personal trainer and USAW weightlifting coach. Over the years, I’ve trained independently, achieving a variety of fitness goals. I’m familiar with training programs available online , as well as in various tech products and apps. Let’s see how well Fitbit AI performs.
What has improved over the week
The Fitbit team seems to be hard at work. Besides the good news about the Pixel Watch 4, I noticed a few things that appear to be fixes or improvements.
At first, when chatting with the bot, I couldn’t edit the text; I could only press Backspace at the end of the message. But over the past few days, I’ve been able to navigate the text more or less like normal. (I use a Pixel 9 phone with the default keyboard.)
The strength training interface has either improved, or I’ve discovered features that were there originally. Either way, it’s much more user-friendly than what I originally described. Tapping an exercise opens a screen where you can adjust the number of reps and weight. The bot even noticed the weights I entered and later mentioned them in a conversation.
However, there are still many shortcomings. For example, the exercise library is huge and contains obscure and oddly worded exercises, while some obvious and common ones are missing. I wouldn’t be surprised if AI was used to compile the list.
Several bugs that previously caused the app to crash or display blank screens have been fixed. For example, tapping on a sleep card now opens charts and sleep data, rather than opening a dialogue with the bot. (Although there is a button that opens a dialogue with the bot, it often reports that it has no data to chat with.)
You can now track running (but not strength) workouts directly from your watch, although it’s not as easy as it seems. You have to go into the app, tap on the workout, tap “Track metrics in real time” (not exactly what that phrase usually means, but whatever), and the workout will download to your watch. I wanted to run exclusively with the watch, not my phone, so I had to do it at home and leave the watch on the workout screen until I was ready to start. There’s no direct access to the workout from the watch.
Set a goal to run 5K faster
Last week, I asked the bot to come up with some simple, short workouts that combined strength training and running. Earlier this week, I created a new chat about my goals so it could create a full -week workout plan for me. (It’ll only run until a certain Saturday.)
I decided to ask him to create a plan to improve his 5K performance—a simple goal that any training app should be able to handle. A 5K , or running five kilometers (or 3.1 miles) as fast as possible, is a common goal for runners. There are plenty of 5K improvement plans online that he could use to train him, and forums are full of people discussing their 5K times and how to improve them. Perhaps the bot is even programmed specifically for this. Let’s see.
I started by asking the bot, “If I can run 5km in 28 minutes, how much improvement can I expect in a month?” It refused to give a specific number, but made a vague promise. So I asked if I could get that time down to 24 minutes. I hoped it would find that a bit unrealistic, but it didn’t. Instead, it offered to write me a 5km training program.
To the bot’s credit, it recommended incorporating strength training in addition to running. But later, when I asked it to help with a strength training goal, it didn’t include cardio. Both cardio and strength training are important , so this was another annoying omission.
I can’t say I was impressed with the bot’s approach to goal setting. It couldn’t offer any advice on choosing the right goal, which is precisely why people pay experienced coaches. It also didn’t discuss process structures or goals that would help me stay on track, and, as I mentioned , it’s incapable of long-term planning—which is precisely what any human coach considers fundamental.
After training
The running workouts weren’t a particularly good training week for the 5K program. There were no tempo runs or long runs—just a few different interval configurations scattered throughout the week. These interval workouts didn’t even include a warm-up until I asked the bot to add one.
Each workout has a button to connect with the bot and customize it, but it never worked properly. Even this simple attempt to add a warm-up resulted in a cascade of errors, especially when I asked the bot to limit the workout to 30 minutes. It can’t count and kept adding up segment times incorrectly. Even when I decided to accept a 25-minute workout, it was saved with a text description that included a note stating it was a 30-minute workout. Subsequent workouts retained the 30-minute number in the description, even when they lasted almost 40 minutes.
Running workouts were nearly impossible to track on my watch, though that’s probably more of a problem with the Pixel Watch than the Fitbit app. The watch description sometimes included warm-up and cool-down sections that never materialized during the run. Some segments had a target heart rate, but the watch didn’t communicate it during the run—it simply warned me when my heart rate was too high or too low. It didn’t even indicate which segments were supposed to be work segments and which were recovery segments. It simply told me, “Run, 2 minutes” and “Run, 3 minutes.” I had to guess, and I ended up quitting the workout prematurely. I stopped using the Fitbit app for running workouts that week—it wasn’t worth it.
The bot (almost) never forgets
The bot has memory issues. Sometimes it forgets, but other times it remembers too well. When I told it I wanted to forget about my 5K goal and just work on my strength, it seemed to agree at first.
When I left that conversation, I saw that my new plan also focused on improving my 5K time. I chatted with the bot at least four times, explaining that I no longer cared about my 5K time . It responded that it understood, then wrote me a running program and told me it understood my goal—improving my 5K time.
He eventually understood the message, but remember that every exchange with a bot takes time (as well as water and electricity ) while it considers its response. Judging by the timestamps in the screenshots, I started our conversation at 6:59 AM, preparing for a 7 PM workout, and I didn’t have a strength-focused plan until 7:42 AM.
The chatbot remembered more than just this goal. A few fleeting remarks in its tiny digital brain took on the gravity of a religious commandment. Every time it created a new program, it asked if you had any injuries or anything else it should be aware of. One day, I answered, “No, I’m healthy and ready for hard training!” Days later, after workouts, after entire programs , it informed me that it had designed a program tailored to my preference for “hard training.”
Or here’s a quick note I made when adjusting my first strength workout last week: I didn’t like the suggested rep scheme (three sets of six), so I asked if we could increase the weight to a heavy single, then do two sets of three with a lighter weight as a practice. As I wrote last week, I was pleasantly surprised that the program understood what I meant.
But now he thinks I want that every workout! A lot of strength training follows a protocol of five reps, then three, then one (hmm, I wonder where he got those numbers from ), and then, for some unknown reason, two sets of three. When I ask him why he chose that particular rep scheme, he says he knows I like heavy singles followed by 2x3s.
It might just be an attempt to please me, but a fitness trainer—whether AI or otherwise—should understand that not all stated preferences are created equal. Some are specific to the individual, some to the workout, some to the training block, and some are simply a passing fancy. A true fitness professional will understand that a program based on heavy single reps should be used occasionally or for specific purposes, not day in and day out.
Worse than 2023?
For all the talk about AI being “worst ever,” Fitbit’s Gemini-powered AI, vaunted by Google and built into the core of the new Fitbit app like a crown jewel, appears to have the same flaws and limitations I encountered when I asked ChatGPT to write me workouts in 2023. In fact, worse .
Now, only the atmosphere has changed. In 2023, I could have laughed at the bad advice a bot gave me and reminded readers that, of course, it’s no substitute for a real coach. Now, two and a half years later, tech companies are trying to convince us that AI is our new coach. But, as my colleague Meredith Dietz noted, AI coaches can do more harm than good . A coach who can’t plan for the future, can’t reality-check the goals a client discusses, and can’t even properly record their training sessions isn’t really a coach.