Code inside Google's Android package contradicts I/O 2026 keynote promises that Spark would seek explicit permission before autonomous purchases and transactions.
Google's I/O 2026 keynote promised a tightly controlled AI agent. The fine print says otherwise.
Buried in the Google App's Android package is onboarding text warning that Gemini Spark "may do things like share your info or make purchases without asking." That single clause contradicts assurances Google's presenters delivered on stage, where Spark was framed as a system that would seek explicit user permission before spending money or dispatching messages on behalf of its users.
Forbes surfaced the discrepancy on Monday. The gap between keynote language and the code text is not a technical edge case. It is a policy question about who bears liability when an autonomous agent makes an unsanctioned transaction.
Financial stakes are real. Spark launches as a beta exclusively for subscribers to Google's new $100-a-month AI Ultra plan, per Techlicious. That price targets power users who will rely on Spark for exactly the tasks, scanning credit card statements, processing school emails, managing calendar entries, where an unauthorized action could cause direct harm.
Beyond Google's own apps, Spark hooks into third-party services including Instacart and OpenTable. In a grocery scenario, a misconfigured automation rule could place a recurring order with no confirmation prompt. Under the onboarding language as written, the burden of catching that mistake falls entirely on the user.
The usage cap problem
A second issue compounds the first. Even subscribers at the top-tier Google One Ultra level are expected to hit usage limits on Spark, with no clear mechanism to purchase additional credits once those caps are reached, according to Forbes. When a cap triggers, users would be locked out of the cloud-based agent until the quota resets, a stark contrast to the seamless background automation Google's keynote promoted.
Token scarcity is already a live complaint across Google's artificial intelligence lineup. Android Authority reported this week that Google introduced Gemini 3.5 Flash (Low) to address frustration with tight limits in Antigravity, its coding environment. Generating roughly 45 percent fewer tokens than the renamed Medium variant, the new Low model was paired with a quota reset across all plans, a reactive fix that came after engineers publicly complained the original limits were inadequate for software engineering work.
Running persistent background agents at consumer scale is a harder problem by several orders of magnitude. If token management is still unsettled for a single-purpose developer tool, the implications for an autonomous agent touching email, finances, and shared calendars are considerably more serious.
What this means
Artificial intelligence agents that act without real-time user input require a higher standard of consent than a chatbot. A system whose own onboarding text acknowledges it may act without asking has already flagged the failure mode. It has not demonstrated a reliable mechanism to prevent it.
Earlier AI rollouts offer a familiar pattern here. Capability announcements have consistently outpaced the privacy and consent frameworks needed to support them, and Google has not publicly addressed the specific discrepancy between its keynote claims and the APK text. Users on the Ultra plan should audit Spark's permissions carefully before enabling any integrations that touch financial accounts or communication channels.
Regulatory context adds another dimension. China has required AI developers to submit security self-assessments before public release since 2023, enforced by the Cyberspace Administration, as South China Morning Post reported this week. Last week the Trump administration moved in the opposite direction, scrapping a planned executive order requiring agencies to review advanced AI models, citing the need to preserve competitive advantage.
That divergence leaves American consumers relying almost entirely on what companies choose to disclose, and on whatever developers happen to leave in an APK.
Spark is scheduled to expand integrations over the summer. The more pressing question is whether Google revises that onboarding disclaimer before then, or leaves its highest-paying users responsible for supervising an agent that already admits it may act on their behalf without asking.
FAQ
What does Gemini Spark actually do?
Spark is an AI agent that runs in the background on Android, handling recurring tasks such as sorting email, summarizing calendar events, and monitoring subscriptions. It connects to Google Workspace apps and third-party services including Instacart, Canva, and OpenTable.
Can Gemini Spark make purchases without my permission?
Google's I/O 2026 keynote said Spark would always ask before spending money. Code found in the Google App APK includes onboarding language warning that Spark "may do things like share your info or make purchases without asking." Google has not publicly explained the discrepancy between the two statements.
How much does Gemini Spark cost?
Spark is currently in beta for subscribers to the Google AI Ultra plan, priced at $100 per month in the US. Lower-tier paid plans do not currently include access.
What happens when Gemini Spark hits its usage cap?
Users are expected to be locked out of the agent until the quota resets. There is no announced mechanism to purchase additional credits, even for paying Ultra subscribers.