A new Perspective paper argues that the important question about generative artificial intelligence is no longer only whether these systems can stand in for humans in lab studies. Instead, the authors ask whether—and when—people are beginning to delegate parts of their own judgement to large language models (LLMs). The paper is written by Henrique Ferraz de Arruda and Yamir Moreno and was posted on arXiv on June 10, 2026.

The paper does not report a new experiment. It lays out a way to study a social change the authors see happening. They collect existing findings and examples across many areas—health, law, finance, education and personal support—and argue for treating LLMs as “social actors.” By that they mean systems whose outputs shape what people decide, not tools that only provide neutral facts.

The authors give a clear definition of “AI delegation.” It is more than asking for information. Delegation, on their definition, is when a person accepts an AI’s recommendation or framing as the basis for a real decision, with little independent checking or comparison to other options. For example, when someone asks an LLM whether a symptom needs medical attention and then decides what to do based mainly on the answer, that is delegation rather than mere information‑seeking.

Why this matters, the paper says, is that individual decisions influenced by AI can add up. The authors describe feedback loops: human behaviour produces data, that data helps shape LLMs, the LLMs influence more human decisions, and the cycle repeats. This coupling could change social norms and collective outcomes. The paper highlights possible risks such as behavioural homogenisation (people’s choices becoming more similar), erosion of personal autonomy, and the transmission of values embedded in model outputs.