William Brannon

We conduct the first large-scale audit of consent and usage policies for 14,000+ web domains that underpin major AI training corpora like C4. We find a rapid rise in AI-specific restrictions and inconsistencies between Terms of Service and robots.txt, with large portions of widely used corpora now formally off-limits for AI training. The work highlights an emerging “data commons” crisis that will shape the future of both commercial and academic AI.

Covered by outlets including Nature… Show more

We conduct the first large-scale audit of consent and usage policies for 14,000+ web domains that underpin major AI training corpora like C4. We find a rapid rise in AI-specific restrictions and inconsistencies between Terms of Service and robots.txt, with large portions of widely used corpora now formally off-limits for AI training. The work highlights an emerging “data commons” crisis that will shape the future of both commercial and academic AI.

Covered by outlets including Nature, The New York Times, Vox, and MIT Technology Review. Show less

We introduce AudienceView, an LLM-powered tool that helps journalists make sense of large volumes of audience comments. The system clusters themes, links them to specific comments, visualizes sentiment and distribution, and supports idea generation for follow-up stories. Through user interviews, we explore how such tools can fit into newsroom workflows while keeping human judgment central.

We demonstrate the Bridging Dictionary, an interactive tool that shows how politically charged terms are understood differently by Democrats and Republicans. A printable version drills down on nearly 800 top terms. Both versions surface partisan usage patterns, sentiment, and examples, using LLMs applied to large-scale political text from Twitter. The goal is to help journalists and communicators choose language that reduces miscommunication across political divides.

We trace licenses, sources, and attribution for 1,800+ text datasets widely used to train and fine-tune language models. The audit reveals widespread omission and mislabeling of licenses on popular hosting sites, plus sharp divides in what data is legally usable for commercial AI (e.g., low-resource languages and creative tasks are often restrictively licensed). We release both the results of our audit and an interactive UI, the Data Provenance Explorer, to help practitioners inspect and filter… Show more

We trace licenses, sources, and attribution for 1,800+ text datasets widely used to train and fine-tune language models. The audit reveals widespread omission and mislabeling of licenses on popular hosting sites, plus sharp divides in what data is legally usable for commercial AI (e.g., low-resource languages and creative tasks are often restrictively licensed). We release both the results of our audit and an interactive UI, the Data Provenance Explorer, to help practitioners inspect and filter training data by provenance.

Covered by outlets including the Washington Post, IEEE Spectrum and VentureBeat, as well as a Nature Machine Intelligence editorial. Show less

We propose ConGraT, a self-supervised method that jointly trains a language model and graph neural network to align text and graph representations on text-attributed graphs. Using a CLIP-style contrastive objective extended with graph structure, ConGraT outperforms prior methods on node/text classification, link prediction, and language modeling. We also show how it can uncover more textually grounded communities in social networks, and release code and datasets.

This paper synthesizes the landscape of foundation model training data and argues that current practices need improvement in many dimensions: authenticity, consent, privacy, documentation, and others. We map the limitations of existing tools and standards, and outline what technical infrastructure and policy changes are needed to make training data more transparent and trustworthy. The goal is to give researchers, developers, and policymakers a shared roadmap for responsible data-centric… Show more

This paper synthesizes the landscape of foundation model training data and argues that current practices need improvement in many dimensions: authenticity, consent, privacy, documentation, and others. We map the limitations of existing tools and standards, and outline what technical infrastructure and policy changes are needed to make training data more transparent and trustworthy. The goal is to give researchers, developers, and policymakers a shared roadmap for responsible data-centric AI.

ICML 2024 Spotlight paper (top ~3% of submissions). Show less

We compare how 1,600+ news events spread and fade across Twitter (X) and U.S. talk radio. News on Twitter moves faster, decays more quickly, and is substantially more negative and outraged than on radio, across different user groups. The findings show how social media may shape the news agenda: by discussing stories first, social media may set the pace and emotional tone of the news cycle.

Presented at the Computation + Journalism Symposium and the International Conference on… Show more

We compare how 1,600+ news events spread and fade across Twitter (X) and U.S. talk radio. News on Twitter moves faster, decays more quickly, and is substantially more negative and outraged than on radio, across different user groups. The findings show how social media may shape the news agenda: by discussing stories first, social media may set the pace and emotional tone of the news cycle.

Presented at the Computation + Journalism Symposium and the International Conference on Computational Social Science (IC2S2). Show less

We compile detailed metadata and tools for examining dozens of popular instruction-tuning datasets used to train and align language models. The project lets practitioners inspect licenses, languages, domains, and other fine-grained characteristics of the data behind these models. Our analysis highlights just how fractured and opaque current data transparency practices are, and offers practical tooling to make data-centric model development more informed and responsible.

Spotlight paper… Show more

We compile detailed metadata and tools for examining dozens of popular instruction-tuning datasets used to train and align language models. The project lets practitioners inspect licenses, languages, domains, and other fine-grained characteristics of the data behind these models. Our analysis highlights just how fractured and opaque current data transparency practices are, and offers practical tooling to make data-centric model development more informed and responsible.

Spotlight paper at the GenLaw @ ICML 2023 workshop. Show less

We analyze a large corpus of professionally dubbed media to see how humans actually localize video across languages. Contrary to common assumptions, human dubbers often prioritize vocal naturalness and translation quality over strict lip-sync and character-length constraints, and we document how timing and source audio shape the final dub. These insights challenge standard objectives in automatic dubbing and suggest new directions for model design.

Presented at ACL 2023 and covered by… Show more

We analyze a large corpus of professionally dubbed media to see how humans actually localize video across languages. Contrary to common assumptions, human dubbers often prioritize vocal naturalness and translation quality over strict lip-sync and character-length constraints, and we document how timing and source audio shape the final dub. These insights challenge standard objectives in automatic dubbing and suggest new directions for model design.

Presented at ACL 2023 and covered by industry outlets such as Slator and Papercup. Show less

We introduce RadioTalk, a large corpus of U.S. talk radio transcripts (2.8 billion words from 284,000 broadcast hours) with rich metadata on stations, shows, speakers, and locations. The dataset is designed to support both NLP and social science research on conversation, media, and politics. We demonstrate its value with initial analyses of content, style, and speaker patterns across the talk radio ecosystem.