OpenAI Codex head says AI struggles with creative design due to taste

OpenAI Codex head says AI struggles with creative design due to taste

Artificial intelligence remains unable to master creative design because it lacks human taste and the subjective judgment required to determine when a visual element actually works. According to Andrew Ambrosino, OpenAI's head of Codex, design is significantly harder to measure and grade than software code.

Ambrosino explained on an episode of Lenny's Podcast that the process of training a model to distinguish between good and bad design is a little bit more tedious and onerous than verifying if code compiles. Because of this gap in aesthetic judgment, he said, Let's give it up for the human brain for now.

The "Generic" AI Aesthetic

Without specific constraints, AI-generated designs often default to a predictable look. Dylan Field, CEO of Figma, stated on Hard Fork that because models are trained on a distribution of data, they typically produce results that people recognize as average.

Community reports describe this default AI aesthetic as featuring floating panels, soft gradients, oversized rounded corners, and dramatic shadows — a specific Inter-and-purple vibe that signals the work was AI-made. Other frequent issues include instructions leaking into UI copy and broken mobile layouts.

Grammy-nominated musician Bas echoed the need for human intervention during a 2024 Harvard conference, noting that human taste is still required to prompt AI to create something people will actually enjoy.

The Evolution of Codex

The current iteration of Codex is distinct from the original 2021 code-completion model, which was deprecated in 2023. Today, Codex is an agentic coding tool that plans, runs, and verifies code from natural-language tasks. It operates across a terminal CLI written in Rust, a desktop app with an in-app browser, a cloud experience, and IDE extensions for Windsurf, Cursor, and VS Code.

By early 2026, Codex became a credible tool for interface design, provided it is given the right verification loops, skills, and references. To move beyond generic outputs, OpenAI guidance suggests a five-step loop for turning reference images into responsive UI:

• Feeding the agent screenshots via the terminal or image flag.
• Providing concrete constraints in the prompt.
• Running a dev server in a separate terminal.
• Keeping prompts small and focused.
• Committing successful iterations and reverting bad ones.

Creative Collaboration and Industry Use

Despite the struggle with innate taste, some OpenAI staff use Codex as a creative collaborator to accelerate ideation. Chad Nelson, a Creative Specialist at OpenAI, described the tool as acting like a junior art director that understands style guides, fonts, composition, and brand books.

Nelson detailed a project for an electric SUV brand called ELARIS, where he used Codex to generate a full campaign concept in a single day. The AI built a web tool and workflow system that automatically constructed advert concepts based on customer profiles — including income level, age range, language, and target city, to generate slogans and select visuals for themes such as weekend ready and everyday luxury.

Integration and Ecosystems

In February 2026, OpenAI and Figma announced an official partnership to turn the Figma MCP beta into a bidirectional integration. This allows users to implement Figma frames in code via get_design_context and push running UI back to editable Figma frames using generate_figma_design.

To bridge the taste gap, third-party projects like Open Design, an independent, local-first open-source layer, wrap Codex in curated design-system libraries and skill sets. This approach aims to remove the toil of maintaining design context by hand for every project. While Claude Code is noted for codebase-aware design decisions, Codex, particularly on GPT-5.4, is recognized for strong visual polish and its ability to handle delegated, sandboxed builds.