I let Gemini watch my family for the weekend — it got weird

This weekend, I turned my home into a test lab for Google’s new Gemini for Home AI and subjected my family to 72 hours of surveillance as it watched, interpreted, and narrated our every move. My purpose? To find out if an AI that sees everything is actually helpful or just plain creepy.

“R unpacking items from a box,” read one notification from the Nest camera on a shelf in the kitchen. “Jenni cuts a pie / B walks into the kitchen, washes dishes in the sink / Jenni gets a drink from the refrigerator,” it continued. Sometimes, the alerts sounded like the start of a joke, “A dog, a person, and two cats walk into the room / Two chickens walk across the patio.”

But these weren’t jokes. They were mostly accurate descriptions of the goings-on in and around my home, where I’d installed several Google Nest cameras powered by Gemini for Home. This is a new AI layer in the Google Home app that interprets footage from the cameras and — combined with Nest’s facial recognition feature — delivers a written description of the events, including who or what is present, what they are doing, and sometimes even what they’re wearing.

For example, now, instead of alerts saying “animal detected on the porch,” I get more descriptive versions telling me that it’s two chickens or one dog. One of those requires immediate action on my part (my husband is not a fan of chickens pooping on our outdoor couch). The other I can ignore. Alerts like the one from the Nest Doorbell at 1AM, which clarifies it’s my son at the front door trying to get in, are less anxiety-inducing than one that just says “person detected.”

Using AI to improve the constant barrage of notifications from security cameras is a major upgrade — one that every manufacturer from Ring to Arlo to Wyze is chasing. But how much can we really trust AI to watch over our home?

While Gemini didn’t hallucinate strangers or wildlife in my home as others have reported, its daily summaries — called Home Briefs — in the Google Home app leaned more toward fiction than fact. This is where Gemini AI in the home becomes problematic.

I don’t record my family inside our home. I leave the security to outdoor cameras and set any indoor cameras to switch off when we’re home. But I wanted to give Gemini as much information as possible, so I packed my home with cameras all set to record.

I installed the new Nest Doorbell 2K, Nest Cam Indoor 2K, and Nest Cam Outdoor 2K (all wired), along with some earlier models compatible with Gemini. This provided surveillance for the major traffic routes inside and outside my house, as well as a large swath of my backyard.

Gemini’s descriptions and the Home Brief, which are currently in early access, require a $20 per month or $200 per year Google Home Premium Advanced subscription, which also includes 24/7 video recording. The AI only analyzes video, not audio, presumably because it’s using a vision language model for processing. Gemini AI also powers a new feature in the Google Home app called Ask Home, which lets you search recorded video, something also offered by Ring.

When I tested Ring’s version of video search, I found it helpful for keeping tabs on my outdoor cat. Google’s version is better, because it understood the context of my requests. When I asked both to show me the last time chickens were on my porch, Gemini surfaced the most recent sightings, whereas Ring showed me the best match, which was from three days ago.

But beyond finding errant animals, I have struggled to find value in this blow-by-blow account of the actions in my home. And yes, it feels creepy.

The real-time alerts were mostly accurate and straightforward. I only spotted two major errors: it mistook my dog for a fox in the backyard, and while I had a package sitting on my doorstep waiting to be picked up, anyone approaching the door became someone delivering a package.

However, Gemini appears to have some selective interpretations. I tried to get it to say I was holding a knife — even brandishing it threateningly at the camera. But it would just use words like chopping or carving in the descriptions, no mention of a knife. And it was a big knife!

When my husband left the house carrying a shotgun, the alert said he was carrying a garden tool. It’s unclear whether this is intentional (I’ve reached out to Google), but not identifying weapons seems like a significant oversight for a security-focused system. I would like to know if there is a person with a gun on my porch or wielding a knife in my home.

The daily Home Briefs is where it got weird. Around 8:30 each night, the Google Home app presents its interpretation of the day’s events in a new Activity Tab. You can customize this to focus on things you’re most interested in; for me, that’s animals and teenagers arriving home past curfew.

The summaries were about 80 percent correct, with a few less concerning confusions — it said a pizza oven was delivered when I was actually taking it away. But my main issue was with the heavy-handed editorializing in the summaries. Gemini took accurate real-time descriptions of events and turned them into a narrative full of declarative sentences that were simply incorrect.

For example, on Halloween, the summary said “Jenni and R were seen interacting with the trick-or-treaters and enjoying the festive atmosphere.” While I did pass out candy, my daughter R wasn’t home. Another summary detailed how my husband and I had spent an enjoyable evening relaxing on the couch along with others. We were home alone.

Home Brief is an interesting concept. Receiving an end-of-day summary rather than being constantly distracted by alerts throughout the day could help reduce notification fatigue. But it doesn’t mean I can turn off the real-time notifications. I don’t want to wait for a cheerful evening Home Brief to find out about the person who broke into my car at 6 in the morning, or that a fox got into the hen house.

Google Home needs a way to prioritize urgent alerts over less important ones. It also needs to integrate these with my smart home so that I can use specific events — such as chickens on the porch — to trigger automated responses.

But making stuff up as the Home Brief did is inexcusable for a system used for home security. Google Home notes that Gemini may make mistakes and offers short clips under the Home Brief so you can check its work, for example, to see that my husband and I were sitting alone in the living room after all. But while Gemini may be able to describe what’s happening, it’s these attempts at interpreting what matters that fall far short.

While I found the AI descriptions for outdoor cameras useful, Gemini’s “intelligence” has not convinced me to start running cameras inside the house. If I’m going to consider letting AI narrate my family’s life, it will need to be a lot more useful and a lot more accurate.