Taste Labs found the internet was homogenizing before AI arrived

Taste Labs analyzed more than two million websites over a decade and found the web's palettes and layouts were already converging before generative AI existed. Founder Thais Castello Branco argues AI didn't start that collapse, it accelerated it and stripped out context, so a pet shop site and a finance firm site end up looking the same.

• Mission — Taste Labs, out of stealth a few weeks before this talk, says its mission is to end AI slop by decoding subjective domains like design and writing the way coding and math have already been decoded.
• Two workstreams — The company works with frontier labs to evaluate models and build post-training data or RL environments, and separately helps app-layer companies and agents produce better output without touching the model layer.
• Deterministic vs. not — Contrast, alignment, and color palette selection become near-deterministic once context is defined tightly enough that experts agree; aesthetics stays a matter of expert disagreement, so it leans on data instead.
• Three signatures of slop — Castello Branco names repetition, lack of fit (a pet shop site and a finance firm site converging on the same design), and low intent as the three characteristics that repeat across slop.
• 2 million sites — Taste Labs analyzed over 2 million websites going back roughly a decade using a Wayback-Machine-style approach, plus a synthetic set of AI-generated sites for comparison, and found the internet was already homogenizing in color palettes and layouts before AI, with AI accelerating repetition across unrelated categories.
• Probes — The team mined features (colors, typography, layout, audience) from the data and trained small classifiers called "probes," each detecting one characteristic; stacking probes and measuring their combined frequency predicted slop with high accuracy, beating typical LLM-as-a-judge methods.
• Judgment, not taste — Castello Branco avoids the word taste, preferring "judgment," and argues that as production cost goes to zero, judgment becomes more valuable, not something everyone should be expected to acquire.
• Inference time — She argues fixing slop only at the model layer is insufficient because intent and context are exchanged at inference time, when the user actually interacts with the system, so both layers need work.
• Creativity API — One in-progress system, nicknamed the "creativity API," acts as an inspiration machine for agents to intentionally produce out-of-distribution output rather than raising temperature and hoping for randomness.
• Brand API — Taste Labs' first public product, already in beta with design partners, extracts a brand from a URL into structured components an agent can follow and a person can judge against.
• Demo example — Using the General Intelligence Company of New York's brand, Claude's default slide deck output is shown against a version produced after brand extraction, which the comparison shows as far higher fidelity to the original brand.
• Fallback repository — For users without an existing brand, Taste Labs is building an index of pre-built, cohesive brand systems to retrieve from, rather than generating a new one from scratch each time.

In their words

so we analyzed over 2 million websites from the past like 10 years kind of like way back machine style to try to understand all the trends across like design6:27

so one was that you already kind of saw a a bit of like a collapse uh on the internet before even AI6:54

this performed better, by the way, than like most LLM as a judge methods of like asking an LLM to like judge if that uh is like great human quality versus like AI-generated slop8:18

if you have repetition and let's say one person asks for a website for their pet shop and the other one asks for a website for their finance firm and somehow those designs converge and look the same5:16

Disclosure · Castello Branco is the founder of Taste Labs and is describing and demoing her own company's research and products (probes, the creativity API, the Brand API).

One thing to add — One thing to add — the claim that probes "beat" LLM-as-a-judge methods is stated without benchmark numbers or methodology in the talk, so it's worth treating as a company claim rather than a verified result. The finding that homogenization predates AI is the most falsifiable and interesting part of the talk, since it reframes AI as an accelerant rather than the root cause.</note> </invoke>

One thing to try tonight
Pick one AI design tool you use, feed it your own site's brand URL alongside a generic prompt, and compare the two outputs side by side for repetition, fit to your brand, and whether it guessed your intent correctly.