Research

The emergence of generative AI has intensified long-standing tensions between content producers and platforms over the value of online content. We examine how news organizations responded to this shift by analyzing historical robots.txt snapshots from the Internet Archive's Wayback Machine, constructing a monthly panel of news publishers across the United States, Europe, and the Rest of the World from May 2022 through December 2025. Using interrupted time series and Callaway-Sant'Anna staggered difference-in-differences designs, we study publisher responses to the release of GPTBot in August 2023.

We document three main findings. First, blocking accelerated after GPTBot's release across all three regions, with U.S. and European publishers increasing blocking at similar rates and more strongly than publishers in the Rest of the World. Second, paywalled outlets block more quickly and more strongly than free outlets. Third, blocking extended beyond AI-related crawlers: GPTBot blocking was followed by increased blocking of non-AI bots, with the strongest evidence of this spillover in the U.S. and Europe. These findings suggest that the rise of generative AI has contributed to a broader tightening of automated web access among news publishers, with implications extending beyond AI training to the governance of crawling on the web.

Music composition is a domain where structured technical knowledge and open-ended creative judgment jointly determine the quality of the output, which makes it a demanding test of whether generative AI augments or substitutes for human creativity. In partnership with an online composition platform, we randomly assign composers access to an AI assistant that offers real-time harmonic and melodic suggestions inside their existing workflow.

We estimate the causal effect of AI access on compositional productivity, on quality as judged by independent human raters, and on the harmonic complexity and diversity of the music produced. Because the platform records what composers do with every suggestion they are shown, we can distinguish AI that scaffolds human judgment from AI that replaces it, and we stratify on measured skill to ask whether the gains accrue to beginners or to experts. A post-trial willingness-to-pay measure tests whether access to AI changes the perceived value of creative tools.

Platforms increasingly require creators to disclose when content is AI-generated, but little is known about what disclosure does to the creators who comply. We study YouTube's "Altered or synthetic content" label. For a large panel of videos we track each video's disclosure state alongside its views, likes and comments, re-reading the full panel on a rolling cycle so that the moment a label appears is observed directly rather than inferred.

We exploit the staggered timing of these labels in an event-study design, using not-yet-labelled videos as controls, and separate the effect of a label appearing from the baseline behaviour of AI-generated content and from platform-wide shocks. The results speak to how transparency mandates interact with creator incentives and to the design of AI disclosure policy.