What happens when an AI product stops feeling like software?
Google's Gemini Spark is producing that reaction in early users. The consistent pattern: technically the most impressive AI experience they've had, and also genuinely unsettling in ways previous models weren't. Those two feelings aren't contradictory. At the frontier of AI capability, they tend to arrive at the same time.
The "terrifying" part of these reactions rarely involves the model doing something dangerous. It's usually a capability gap closing faster than expected. When an AI handles a complex, ambiguous request with something that feels less like text generation and more like actual comprehension - tracking what you meant rather than what you literally typed - the experience reads as both impressive and disorienting. GPT-4 triggered similar reactions in early 2023. Claude 3 Opus did it again in 2024.
Gemini Spark appears to be the next product to push into that territory. Whether the capability is genuine or a surface-level effect (models have gotten very good at sounding insightful without being so) is something only extended use in real workflows will confirm. Demo impressions consistently overestimate frontier models - the real test is whether a tool holds up during your third hour of actual work on a Tuesday afternoon, not during a first-session showcase.
For people building content, running campaigns, or managing client work: Google has historically produced strong Gemini benchmark numbers that didn't fully translate into the actual user experience. If Spark is generating reactions this strong across early testers, it suggests that gap may have narrowed. Availability and pricing details are still emerging.