Stanford-Led Study: 35% of Websites Created After ChatGPT Contain AI-Generated Content

A joint research team from Imperial College London, the Internet Archive, and Stanford University has published a paper, "The Impact of AI-Generated Text on the Internet," that empirically documents the pervasiveness of AI-generated text across the internet.

Their estimate suggests that roughly 35% of websites created after ChatGPT's release in November 2022 contain AI-generated text. The study also flags the spread of "unnaturally upbeat prose," underscoring once again how deeply AI-generated content is affecting the internet's information ecosystem.

**Why it matters:** Large-scale empirical data now backs the claim that AI is substantially transforming the quality of content on the internet. By quantifying the impact of what has long been debated as "AI pollution," the study could shape future content policies and search engine design.

Figma Rolls Out AI Agents That Create and Edit Designs Directly

Figma has begun offering a feature that lets AI agents create and edit designs directly on the Figma canvas. Notably, it also incorporates safeguards to prevent unintended designs from being generated.

With Figma — the flagship of design tools — getting serious about agent capabilities, UI/UX design workflows could be in for a major shift. That said, how well the guardrails against "unintended UI generation" hold up in real-world use remains to be seen.

The Reasoning Model Paradox: "The More They Think, the More Biased They Get"

A paper published on arXiv, "More Thinking, More Bias: Length-Driven Position Bias in Reasoning Models," reports a counterintuitive finding. Chain-of-Thought (CoT) reasoning and reasoning-tuned models like DeepSeek-R1 have been assumed to reduce shallow heuristic biases by thinking things through. In reality, however, longer reasoning traces were found to amplify position bias.

Tested across 13 reasoning mode settings, 12 showed a positive correlation between reasoning trace length and the Position Bias Score (PBS). Even the 671B-parameter DeepSeek-R1, whose overall PBS is low, still exhibited bias in the quartile with the longest traces.

**Why it matters:** The results cast serious doubt on the assumption that reasoning models become more accurate the longer they deliberate. In particular, they show that reasoning models must not be treated as "order-robust" in multiple-choice evaluation pipelines.

AI Cognition Evolves Unevenly: Superhuman on Language, Catastrophic on Vision

The paper "Uneven Evolution of Cognition Across Generations of Generative AI Models" evaluated the cognitive profiles of generative AI models using a psychometric framework based on the Wechsler Adult Intelligence Scale (WAIS). The results were striking.

The models scored above the 98th human percentile in verbal comprehension and working memory, yet below the 1st percentile in perceptual reasoning — revealing an "extremely uneven cognitive architecture." A longitudinal study spanning six generations and two model families found a major dissociation between "verbal presentation" and "visual presentation" even in abstract quantitative reasoning, confirming a strong architectural bias toward language-based symbol manipulation.

**Why it matters:** This suggests that scaling alone may never reach "balanced, human-like general intelligence." Accurately understanding the strengths and weaknesses of today's LLMs is essential for deciding where to deploy them.

LLM "Planning" Turns Out to Be Myopic — Revealed by Reasoning Trace Analysis

"Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning" proposes a new method for extracting and quantifying search trees from LLM reasoning traces. Analyzing the game of Connect Four, the researchers found that LLMs search more shallowly than humans, and that performance is predicted by the "breadth" of the search rather than its "depth."

Most intriguingly, even though LLMs expand deep nodes within their reasoning traces, their actual move choices are best explained by a "myopic model" that ignores those deep nodes. This stands in contrast to humans, where deep search is what drives performance.

**Why it matters:** Even when an LLM appears to be "thinking long and hard," it may in fact be making decisions based on shallow information. This raises fundamental questions about how we evaluate AI reasoning capabilities.

---

% of post-ChatGPT websites are AI-generated: ITmedia AI+ / research by Stanford University and others

Figma AI agent features: ITmedia AI+ / Figma announcement

More Thinking, More Bias: arXiv:2605.06672

Uneven Evolution of Cognition: arXiv:2605.06815

Extracting Search Trees from LLM Reasoning: arXiv:2605.06840