I Resist Generative "AI"

Esther Weidauer presents a broad case against the current generative-AI boom, focusing on text and image generators and systems built around them rather than machine learning as a whole. Drawing on experience in software development, search, neural networks, psychology, and statistics, she explains that large language models generate text by repeatedly predicting likely next words from the prompt, which includes the conversation history up to a limited context window. So-called AI agents, she says, are not a fundamentally new intelligence: they are programs that repeatedly prompt an LLM and use its outputs to trigger tools such as web searches, payments, or emails. Other machine-learning applications, including weather forecasting and image classification, are distinct from the generative systems she opposes.

Weidauer challenges common defenses of generative AI. Its continued existence does not prove its value, and today’s adoption is partly sustained by investor subsidies that conceal operating costs. Calling it “just a tool,” she argues, obscures the ongoing services, labor, materials, and infrastructure required to run it, from content moderation to replacing failed data-center GPUs and curating training data. Tools are also shaped by social rules and political and economic arrangements, so they are not automatically neutral. Likewise, usefulness alone does not establish that a technology is beneficial: its value must be weighed against wider consequences. The piece promises to examine environmental, psychological, economic, and epistemological harms, as well as political dynamics and claims about AI’s future and consciousness.