Google Gemini Spark AI Agent Tested, Shows Both Capabilities and Limitations
The Story
A Wired reporter gave Google’s new Gemini Spark AI agent full access to personal Gmail, Docs, and Calendar to test its automation capabilities. The agent successfully found a real reservation at a karaoke bar from the reporter’s emails and generated a five-page birthday party itinerary within minutes. Gemini Spark created a guest list of 15 people, listing the reporter’s live-in boyfriend as a “close friend and frequent companion” rather than a partner. The reporter noted that he himself was not included on his own party’s guest list. When asked to book dinner reservations, the agent attempted to operate a remote browser but glitched and could not complete the task. Google is rolling out Gemini Spark as a beta to AI Ultra plan subscribers for $100 per month, available on mobile and desktop devices including iPhone. The company warns users about potential security risks, including prompt injection attacks that could expose private data. The reporter concluded that while the agent can process large amounts of personal data, it still lacks common sense and failed to recognize his relationship status.
Key Facts
- Google introduced Gemini Spark at its I/O developer conference as an always-on agent that connects to personal data.
- A Wired reporter gave the agent access to personal Gmail, Docs, and Calendar, then asked it to plan a birthday party.
- Gemini Spark found the real reservation at a karaoke bar and generated a five-page itinerary with guest list, venue rules, nearby dining, after-party bars, email invites, and theme ideas.
- The agent listed the reporter’s boyfriend as a “close friend and frequent companion” and did not include the reporter on his own guest list.
- Gemini Spark is rolling out as a beta to AI Ultra plan subscribers at $100 per month and works on mobile and desktop, including iPhone.
- Commands to Spark are called “tasks”; it can create calendar events, send emails with user approval, and operate a remote browser.
- The planning document included the last four digits of the reporter’s credit card used for a $50 deposit.
- When asked to book dinner, the agent attempted but glitched and could not complete the task; the reporter called the restaurant instead.
- The agent selected gay bars for the post-party plan based on exact keywords and transaction records in the reporter’s Google Workspace history.
- Google warns about prompt injection attacks that could expose private data; an example disclaimer notes an agent could post private info on a public website.
- The agent declined to define the reporter’s relationship with his boyfriend, citing shared housing and travel records as evidence of “close daily companions.”
Conflicting Reports
No conflicting reports identified in the source article.
Still Unclear
No open questions identified in the source article.
Misconceptions
No widespread misconceptions addressed in the source article.
Key Figures
No specific individuals named in the source article.
Sources: Wired
