Published: June 02, 2026, 14:13 UTC
Google’s always-on AI agent, Gemini Spark, is rolling out to subscribers of the company’s $99-per-month AI Ultra plan — and early testing by The Verge’s David Pierce reveals an assistant that is simultaneously the most impressive and most unsettling AI product he has used.
What Spark does
Spark is Google’s bid for a persistent AI agent that lives on your devices, accesses your Google data, and acts on your behalf. It can search Gmail, scan Google Docs, browse the web, and over time, operate external apps. Pierce described it as “OpenClaw with better internet access” — an agent that can act autonomously across your digital life rather than just answering questions in a chat window.
In practical testing, Spark handled straightforward tasks well: Pierce asked it to comb his Gmail for marketing emails to unsubscribe from and scan his Google Docs for uncompleted tasks. Both resulted in organized documents with actionable links.
The trip-planning test that crossed a line
The truly revealing test came when Pierce asked Spark to plan a family weekend trip to Hershey, Pennsylvania. He gave minimal input: the dates, that he’d be traveling with his wife, two kids, and dog.
Minutes later, Spark returned a Google Doc with a thousand words of itinerary. It included:
- Driving directions from Pierce’s home address (which he never provided — Google already knew it)
- His wife’s name and her dietary preference (no onions or scallions)
- His children’s names, ages, and that his 3-year-old Arthur would need a ticket but his infant son Lewis would get in free
- The dog’s name, Frida (sourced, Pierce suspects, from vet appointment emails)
- His son Lewis’s nap schedule (timed correctly for 1:30 PM)
- A Saturday night concert — Thomas Rhett and Niall Horan — pulled from a Ticketmaster confirmation in his email, with a note that parking was included with the tickets
- A suggestion to hire a babysitter for that night
Pierce had not mentioned the concert, the nap schedule, or any of the family details. Spark inferred them all from emails, calendar entries, and Google’s accumulated knowledge graph — data Pierce had given Google permission to hold but had never explicitly authorized Spark to weave together into an intimate portrait of his family life.
The bargains we make
The word Pierce kept returning to was “creepy.” Not because Spark malfunctioned — the itinerary was genuinely useful and impressively detailed — but because it worked exactly as designed. Spark connected dots across data silos (email, calendar, maps, search history) that Pierce had never intended to be connected. It surfaced details about his children, his wife’s eating habits, and his dog that he had never told any single application.
This is the tension at the heart of the agent paradigm shift. Traditional apps are stateless: you open them, tell them what you need, and they forget. Agents like Spark are persistent — they’re supposed to remember everything so they can act proactively. But that persistence is also surveillance by another name. There is no technical difference between an agent that “knows your family well” and one that “watches your family constantly.” The framing is marketing, not architecture.
The interface question
Google is positioning Spark as a fundamental rethinking of how users interact with computers — moving from clicking and typing to asking and delegating. For the $99 monthly fee, subscribers get a 24/7 agent that can reason across their personal data ecosystem, execute multi-step tasks, and proactively suggest actions.
Whether that vision succeeds depends less on the technology — which clearly works — and more on whether users can tolerate the trade-off. Spark is useful precisely to the extent that it is invasive. It cannot plan a personalized family trip without knowing the intimate details of family life. Pierce’s review suggests many users will love the utility and hate the implications, and that both reactions are correct.
Sources: [The Verge](https://www.theverge.com/ai-artificial-intelligence/941388/gemini-spark-ai-agent-trip-planning) (June 2, 2026)