What Happens to Design After AI?

What It Covers

A16z general partner Anish Acharya speaks with Microsoft VP of Design John Maeda and Impeccable founder Paul Backus about how AI reshapes design work, why design vocabulary produces better AI outputs than engineering language, and how tools like Impeccable counter model-generated aesthetic homogeneity in software interfaces.

Key Questions Answered

• •Design vocabulary as AI leverage: Designers consistently get better outputs from Claude than engineers using identical models because they deploy specific terminology — vertical rhythm, negative space, visual weight — that engineers lack. Building this vocabulary into agent harnesses, as Impeccable does, closes the gap and improves output quality for non-designers immediately.
• •Slop is a moving target requiring active anti-attractors: AI default aesthetics shift over time — purple gradients gave way to beige backgrounds, instrument serif fonts, and eyebrow text. Impeccable counters this by generating randomized seeds from brief user interviews and scripts that steer outputs into underrepresented regions of the model's latent space rather than defaulting to overused patterns.
• •LLMs approximate taste from outputs, not inputs: Models train on the finished artifacts of human creative decisions, not the reasoning behind them. This means AI can replicate the surface appearance of taste for a specific audience at a specific moment, but cannot reconstruct the judgment process — making human viewpoint amplification a more productive goal than taste replication.
• •Agentic experience design (AX) is the next design frontier: As visual UI becomes increasingly automated, designers should redirect attention toward designing for agent-readable interfaces — robots.txt files, CLI help commands, error messages, API shapes, and information architecture. These non-visual affordances require the same structured thinking as UX but remain largely unaddressed by current AI tooling.
• •Cognitive delegation versus cognitive surrender: Using AI plan mode without critically reviewing outputs represents cognitive surrender — accepting model judgment by default. Effective AI collaboration requires maintaining an active point of view throughout the process. Tools should function as back-and-forth collaborators that preserve user agency rather than pipelines that absorb decision-making from the human operator.