Google warns of large-scale prompting campaign to clone Gemini

Google has reported coordinated attempts to reproduce the capabilities of its Gemini conversational model by repeatedly querying it and collecting answers to train alternate systems. In at least one detected instance adversaries issued on the order of a hundred thousand prompts across multiple non-English languages, using the returned text as training material for a lower-cost knockoff. The firm treats this pattern as an intellectual-property breach that undermines the costly investments made to develop large foundational models. The technique at the center of the alert—training a new system on another model’s outputs—has a name in engineering circles and is attractive because it short-circuits the enormous compute and data barriers to building a competitive LLM. Google says the activity appears commercially motivated and traces to actors worldwide, though it declined to identify specific organizations behind the campaigns. This disclosure sits against a backdrop of earlier disputes over dataset sourcing and reuse; those prior episodes complicate straightforward claims of ownership and are likely to shape regulatory and courtroom debates. Technically, defenses will need to span telemetry-based detection of mass-query patterns, rate-limiting, API access controls, and output watermarking or poison injection strategies to make harvested outputs less useful for downstream model training. Economically, successful cloning reduces barriers for new market entrants, pressures incumbents’ margins, and may accelerate commodity LLM offerings that undercut premium cloud and AI services. Legally, firms will push for clearer protections over model behavior and outputs, while challengers will press limits around fair use, interoperability, and research prerogatives. For practitioners, the immediate priority is operational: improve monitoring, distinguish benign research from systematic extraction, and harden contractual enforcement. For policymakers, the episode raises questions about the appropriate balance between competition, innovation, and protection of expensive AI assets. If left unaddressed, mass replication of state-of-the-art models could homogenize offerings and erode incentives for foundational research funding. Google’s public account functions as both a warning to customers and a call to the industry to invest in defensive engineering and legal frameworks.