AI Taste Simulation

AI taste simulation is the source’s distinction between AI producing more tasteful-looking outputs and AI having taste of its own. In Can Silicon Valley give AI good taste?, Sophie Hagney argues that curated examples, tastemaker feedback, and preference data can make generated text or images less generic, but this still leaves the system imitating, averaging, or applying human standards.

The concept extends Human Taste as AI Training Signal / 人的品味作为AI训练信号 while pushing against an easy conclusion from it. If taste can be expressed in examples, ratings, datasets, or instructions, it can shape model behavior; but Hagney’s argument is that Embodied Taste also depends on lived perception, timing, scarcity, and independent encounter with the world. That makes AI taste simulation adjacent to Human Judgment Under AI and AI Content Devaluation rather than a solved product feature.

Key Claims

• AI output can become more elegant without AI possessing independent taste.
• Human-curated datasets can encode a group’s preferences into model behavior.
• Averaging preference data may produce consensus aesthetics rather than distinctive judgment.
• Taste includes discovery, timing, scarcity, and embodied attention, not only pattern matching.
• A chatbot can use preference-like language while still lacking human-style preference experience.
• The more AI-generated output floods a medium, the more valuable human judgment, authorship, and off-path discovery may become.

Connections

• Sophie Hagney - source guest grounding the distinction.
• Taste Labs - startup case for curated human taste data.
• Embodied Taste - human side of the distinction.
• Human Taste as AI Training Signal / 人的品味作为AI训练信号 and Research Taste - adjacent taste-as-judgment branches.
• AI Slop, AI Content Devaluation, and AI-Generated Content Quality Gap - quality and attention problems the concept responds to.
• Claude, Anthropic, and Model Value Embedding / 模型价值观嵌入 - chatbot preference-language and model-persona branch.