Hands-On With Gemini Spark: I Gave It Access to My Life and It Friend-Zoned My Boyfriend
Google’s Gemini Spark can scan Gmail, Docs, and Calendar to plan tasks, but this hands-on found it impressive, glitchy, and unsettlingly personal.
Intelligence analysis by GPT-5.4 Mini
WIRED tests Google’s new Gemini Spark beta, an always-on agent that can mine personal data to plan events, draft messages, and browse the web. It works fast, but its guest list, failed bookings, and privacy implications show both the power and creepiness of the tool.
Google made a new helper called Gemini Spark that can look through a person’s emails, calendar, and documents to help get things done. In the test, it quickly built a birthday plan from very little input, like a super-fast assistant sorting a messy desk.
It was smart enough to find the real party reservation and suggest people to invite. But it also acted strangely personal, like a robot trying to guess who belongs at the party without really understanding the people.
It also failed at some jobs, like finishing a dinner booking. The story shows that this kind of helper can save time, but giving it so many keys can also make things awkward and risky, like lending a bicycle to a machine that sometimes forgets where the brakes are.
Analysis
What Spark did
Google is rolling out Gemini Spark as a beta feature for AI Ultra subscribers. The tool lives inside Gemini, works on mobile and desktop, and can browse the web, create calendar events, and draft emails after the user approves them. In the hands-on test, the reporter connected Gmail, Docs, and Calendar, then asked Spark to help plan a birthday party.
Spark pulled the real reservation details from email and turned a vague request into a detailed planning document in minutes. It assembled an event overview, nearby dining ideas, invite text, theme suggestions, and a guest list based on information it found in the reporter’s account history. The article notes that it even surfaced the last four digits of the card used for the reservation deposit, showing how much personal context the system can combine.
Where it breaks down
The demo also showed the limits of the agent approach. When Spark tried to book dinner reservations, it opened a remote browser, triggered a verification text, and still failed to finish the task. That makes the agent look useful for research and drafting, but not yet dependable for end-to-end execution.
The most memorable problem was social, not technical: Spark described the reporter’s boyfriend as a "close friend and frequent companion" and left the reporter off the guest list for his own party. Google says the system is not making identity inferences; instead, it is matching exact names, keywords, trips, and transactions from stored files and emails. The piece also notes Google’s warning about prompt-injection attacks, where malicious content can trick an agent that has access to private data. The takeaway is clear: Spark is powerful, but the more it knows, the more it can surprise, misread, or expose the user.
Key points
- Google is beta-testing Gemini Spark for AI Ultra subscribers at $100 per month.
- Spark can use a user’s Gmail, Docs, and Calendar data to plan events, draft emails, and browse the web with approval.
- In the test, it produced a detailed birthday plan, but it also made awkward guest-list choices and left the birthday person out.
- The agent failed to complete a restaurant booking even after opening a remote browser and triggering a verification text.
- Google warns that connected agents can be vulnerable to prompt-injection attacks.
