The
Morning
Brief
Thursday, September 3, 2026
Today's Signal
Today's content surfaces a persistent tension between human perception and machine cognition — AI tools are reshaping how interfaces get made, but research increasingly shows AI vision diverges from how real users actually see. Underneath this runs a deeper craft question: as AI generates more of the visual world, the timeless principles governing human perception (Gestalt, Fitts's Law, color semantics) become more critical, not less, as the baseline designers must defend.
Deep Read
AI Can't See What Your Users See — and That Gap Is a Design Risk
Jakob Nielsen's Substack
Neural network vision models process low-level visual cues like proximity differently than humans do, meaning AI-generated mockups may pass an internal logic check while failing actual users at the perceptual grouping stage. As AI agents and generative tools take on more interface design work, this machine-human perception gap becomes a structural risk rather than an edge case. The practical implication: use AI for candidate generation, but treat human first-read testing as non-negotiable verification — not a nice-to-have.
In the Feed
Dark Gestalt: How Perceptual Principles Get Weaponized Against Users
Jakob Nielsen's Substack
Deceptive interfaces don't break visual design rules — they exploit them, using proximity and similarity to bind unrelated elements and camouflage harmful actions inside benign visual treatments. The Amazon Prime FTC settlement puts a regulatory price tag on what happens when a false perceptual model gets shipped at scale.
Runway's Solaris Generates UI Frame by Frame — No HTML, No CSS, No Code
AlphaSignal
Runway's Solaris produces interactive interfaces entirely through real-time visual generation, eliminating the traditional front-end stack and enabling interfaces that adapt per user and per interaction. It's the clearest signal yet that the Figma-to-code pipeline may not be the endpoint — generative visual rendering could bypass static component systems entirely.
Disclosing an AI's Intent Cuts Persuasion in Half — Its Label Does Nothing
Jakob Nielsen's UX Roundup
A controlled experiment found that AI content labels had zero effect on how persuaded people were, but revealing the AI's explicit goal to persuade reduced that effect by 50%. For designers building AI-powered products, the implication is pointed: transparency about purpose is the real trust lever, not provenance watermarks.
Quick Takes
Gestalt principles aren't soft craft knowledge — they're the specification layer that AI-generated interfaces will most reliably violate, making them more essential to design system governance, not less.
Jakob Nielsen's Substack →Runway's Solaris is the first tool that makes you genuinely ask whether a design system built on static components and tokens is solving the right problem for an AI-generated interface world.
AlphaSignalThe PM-builds-AI-workflow story is really a design systems story in disguise: persistent context, reusable patterns, and architecture beat raw tool capability every time.
Lenny's Newsletter →EU AI watermark mandates may be solving the wrong problem — research suggests hidden intent, not unknown provenance, is what actually makes AI-generated content dangerous.
Jakob Nielsen's UX Roundup →