Refreshments are available starting at 10:30 a.m. The seminar will begin at 10:45 a.m.
Abstract
When a machine can write like a person, a set of ordinary questions gets hard to answer: “Did you write this?” “Who writes like this?” “Where does training data come from now?” “Can I believe what it says?” Chris Callison-Burch’s lab works on all four questions. For the first, his group built Real or Fake Text, a game that asks people to find the sentence where the machine takes over, and Robust AI Detection, a benchmark that measures how detectors do when you vary the generator, change the domain, or let someone edit the text. The second question is what you ask when the first has no reliable answer. Style and authorship embeddings, trained on supervision an LLM produced, let us ask who writes like this, and the same representations let us rewrite one author’s words in another author’s style. The third question turns the first around. If generated text is good enough to fool a person, is it good enough to train on? DataDreamer, FineInstructions, and CoSyn are how Callison-Burch’s group finds out: synthetic instruction data at pre-training scale, and image training data written as code by a model that never sees an image. The vision models they trained on it beat every open and closed model they compared them with. The fourth question is answered by automated verification of scientific claims. Callison-Burch leads the University of Pennsylvania’s effort on DARPA’s Scientific Feasibility program, ongoing work that runs in parallel with Hopkins’ own. His group mines the NSF awards database for claims researchers make and builds agents that weigh them against the literature using Kani and ReDel, two open-source agent frameworks that the group maintains.
Speaker Biography
Chris Callison-Burch is the Raj and Neera Singh Professor of Artificial Intelligence at the University of Pennsylvania, where he directs the online master’s in AI program and teaches the graduate AI course, which enrolls more than 500 students each fall. He was a research faculty member at Johns Hopkins from 2007 to 2013 and chairs the advisory board of its Human Language Technology Center of Excellence. With his students, Callison-Burch works on machine-generated text detection, style and authorship embeddings, synthetic data for text and images, open-source agent frameworks, and automated verification of scientific claims. He has published more than 250 papers, been cited over 40,000 times, and is a Sloan Research Fellow. He has served as general chair of the Association for Computational Linguistics and program co-chair of the Conference on Empirical Methods in Natural Language Processing, and in 2023 testified before Congress on generative AI and copyright. In 2026 Callison-Burch received the Lindback Award for Distinguished Teaching, Penn’s highest teaching honor, and took part in the Isaac Asimov Memorial Debate at the American Museum of Natural History.