I Let Google’s New AI Agent Into My Inbox. It Planned My Birthday Party—and Demoted My Boyfriend.

At its annual developer conference, Google launched Gemini Spark, an always-on AI agent designed to tap into your personal data, handle online tasks, and automate daily interactions. It’s the company’s answer to the OpenClaw agent that shook Silicon Valley early last year, where early users handed over full access to their messaging and calendars—sometimes with awkward, bot-caused blunders.

I decided to test it. I gave Gemini Spark unfettered access to my Gmail, Google Docs, and Calendar. Then I typed a simple request: help me plan a birthday party. Within minutes, the agent dug through my inbox, found a real reservation I’d made at a karaoke bar, and produced a five-page itinerary. It included a guest list, venue rules, nearby restaurants, after-party spots, email invites, and theme ideas. I didn’t have to supervise or leave my laptop open. The whole thing took maybe two minutes.

What made me laugh—and cringe—was the guest list. Gemini Spark scanned my emails and documents and recommended 15 people, the exact maximum for the karaoke room. At the top: my boyfriend of several years, whom I live with. The agent explained, “Your travel history and emails identify [his name] as a close friend and frequent companion, making him a natural first addition.” Close friend? Frequent companion? After I handed over nearly every digital scrap of my life, the AI couldn’t see he’s my partner. It also forgot to include me—the birthday boy—on the list.

Gemini Spark is rolling out this week as a beta for subscribers of Google’s AI Ultra plan, which costs $100 a month. It lives inside the Gemini chatbot as a new tab and works on both desktop and mobile, including iPhones. Instead of prompts, you send “tasks.” The agent can create calendar events, send emails (with your approval), and operate a remote browser to surf the web.

I asked it to book dinner reservations at a sushi spot it suggested. It tried, opened a remote browser, triggered a six-digit verification code to my phone, and then glitched out. I ended up calling the restaurant myself.

When I questioned why it only recommended gay bars for the after-party, Gemini Spark replied that it doesn’t infer identity. It simply scans for exact keywords, past itineraries, and transactions in your Google Workspace history. It pointed to specific emails and documents I’d forgotten existed. Fair enough.

The email draft it wrote for guests was too formal—lecturing about age limits for a crowd of 32-year-olds. I asked for a rewrite, approved it, and let the agent send a test to my boyfriend.

Google warns users about security risks. A help page notes that malicious instructions could trick the agent into posting private data online or sending emails to outside services. That alone should give anyone pause. I wouldn’t recommend giving Spark full inbox access unless you’re ready for potential breaches.

Still, my birthday experiment showed two things. First, the more data you hand over, the more specific the outputs get—but that same access opens security doors. Second, these agents can be technically brilliant and still lack basic common sense. I asked again what it thought of my boyfriend. “Your shared housing, mutual account recovery, and shared travel records show you are close daily companions,” it said. It still wouldn’t call him my partner. After giving it everything, I felt like it barely knew me.