Key Takeaways

  • Grace Lee's research revealed OpenAI's GPT-5.6 Soul improved its design output by actively sidestepping common 'bad AI smell' anti-patterns, not by developing genuine creativity.
  • Specific patterns now avoided include the ubiquitous Bento box layout in dashboards, oversized typefaces in hero images, and certain color palettes or high border radiuses that became generic AI signatures.
  • This design improvement came from identifying 'holes' in a projected design manifold where previous GPT-5.5 outputs clustered due to their generic aesthetics.
  • For founders, this means AI can get better by learning what to avoid, but achieving truly unique design still hinges on user-injected reference points to guide the model beyond mere pattern avoidance.

The AI Design Paradox: Avoiding Bad Taste Isn't Creativity

For ambitious builders, the promise of AI generating stunning, unique designs is tantalizing. Yet, often, AI outputs feel… generic. You know the look: high border radius, specific muted color palettes, predictable layouts. It's what Tyler Cosgrove called “the the the the, you know, high border radius on the edges. There's a little color on the on the side.” This isn't just a hunch; Grace Lee's research, discussed by John Coogan on TBPN, exposed the mechanics behind OpenAI's GPT-5.6 Soul finally shedding this 'bad AI design smell.'

Lee's method was illuminating. “How did OpenAI's Soul finally learn design taste? She projected 1,000 websites by GPT-5.6 Soul into a design manifold and discovered big holes. These holes were where GPT-5.5 previously generated outputs with quote bad AI smell,” Coogan explained. Instead of learning what makes good design, GPT-5.6 Soul learned what makes bad design and simply avoided it. This is a subtle but critical distinction. The AI didn't suddenly become a design savant; it just stopped making the obvious mistakes.

Coogan highlighted three key anti-patterns Lee identified: “One, the Bento box layout in dashboards. Two, large typefaces in hero images.” These were not inherently 'bad' design choices but had become so overused by AI that they signaled a machine-generated output. By consciously steering clear of these predictable elements, GPT-5.6 Soul's outputs became more palatable, less obviously AI-generated, and, frankly, less boring. It's a sophisticated form of pattern avoidance, not a burst of genuine aesthetic insight.

Your AI Won't Be a Genius, But It Can Stop Being Basic

This insight cuts deep for any founder building with AI. If you're relying on a model to generate creative outputs, understand its limitations. “But you're not necessarily like making the model more creative by you by removing these like patterns that always comes to you,” Tyler Cosgrove noted. The model isn't truly more creative; it's just better at disguising its machine origins by not falling into common traps. This means the default output of even advanced AI will trend towards 'inoffensive but not innovative' if left unchecked.

The real leverage, Coogan and Cosgrove suggested, comes from injecting external influence. “If you at least inject like one reference point, you'll usually land somewhere,” Cosgrove said. This is the human touch, the specific aesthetic guidance that pushes the AI beyond its learned anti-patterns into something truly unique. Without it, your AI will simply produce the most statistically improbable, yet still generic, output. The goal isn't to make the AI a Picasso, but to give it enough guardrails and nudges to avoid being the design equivalent of stock photography.

What to Do With This

If you're using AI for product design, marketing materials, or content generation, don't just prompt for "innovative design." Instead, actively identify and prohibit 'generic AI smells' for your specific domain. Pull your last 5 AI-generated designs and pinpoint repetitive elements (e.g., specific layout types, font sizes, color gradients). Then, inject at least one specific visual or stylistic reference point into your prompts—a mood board, a URL to a favorite website, or a precise color palette. This week, try a few iterations: one with your standard prompt, and one with a prompt explicitly banning a known AI anti-pattern and including a direct visual reference. Compare the outputs. You'll quickly see the difference between avoiding bad and aiming for distinctiveness.