Why AI browsers haven't taken off (skyfall.dev)

OpenAI’s new macOS AI browser, ChatGPT Atlas (written in SwiftUI), and similar products like Perplexity Comet promised big time-savings, but user adoption has been lukewarm. Despite ChatGPT’s massive user base, people report mixed experiences driven by privacy worries (reluctance to funnel full browsing histories to single companies), slow and error-prone behavior, and limited real-world usefulness. Simple tasks—like finding the cheapest Raspberry Pi across countries and shipping options—often fail because the browser opens wrong storefronts or ignores shipping/pricing nuance, requiring lengthy prompting and human correction. Technically, current AI browsers suffer from grounding, latency, and safety gaps: frontier models take many seconds to follow links, retrieval and reasoning errors produce suboptimal recommendations, and prompting overhead often negates any time savings. There are also security risks from content-based “jailbreaks” (Brave demonstrated a Perplexity exploit where a malicious page induced an account takeover via the browser’s summarizer). For the AI/ML community this highlights priorities: better on-device or privacy-preserving architectures, faster and more reliable retrieval-grounding, robust instruction parsing and sandboxing to prevent DOM-injection style exploits, and UX designs that keep humans effectively in the loop. Until those gaps close, AI browsers will remain niche rather than mainstream.