The complexity inherent in software has not been eliminated but rather shifted, a concept articulated by Larry Tesler's law of conservation of complexity. AI assistance, while appearing to simplify interfaces by removing modes and menus, has redirected this complexity. It has moved to the user, who now crafts specifications and verifies AI output, and to developers, who manage evaluations and prompt engineering. Furthermore, platform developers face constant, unannounced model changes, and a new role, the reviewer, has emerged to validate AI-generated content.Simplicity in interfaces is achieved by transferring the workload elsewhere, now at an accelerated pace due to AI. While AI can offer initial assistance and improve efficiency for some, it has also led to a slowdown in experienced developers due to the burden of verification. The core issue is not the failure of AI tools but the difficulty in tracking the time spent on these new verification tasks.Deterministic systems like payroll remain a hard floor of complexity due to their absolute requirement for accuracy. However, even retrieval-grounded AI tools have shown significant hallucination rates in legal domains. The inherent difficulty of complex professional domains like underwriting or clinical practice cannot be automated away, as accountability for decisions remains crucial.The "jagged edge" of AI involves complexity absorption that is uneven and hidden, with organizations adopting AI in functions but not necessarily deploying agents effectively. This perceived advancement masks a shift where machines absorb intelligence as soon as they perform a task. The lack of transparency in AI output, unlike older interfaces that failed audibly, makes it difficult for users to discern correct answers from confident-sounding misinformation.Allocation of complexity is a critical design decision, with human-AI interaction guidelines often skipped. The default, cheapest option for developers is a blank text box, which places the burden of specification and verification onto the user. The stakes have risen significantly from the minutes saved by eliminating menus to the ability of individuals to distinguish between accurate and confidently presented incorrect information.