PeopleofSuperintelligence

Researcher

John Kirchenbauer

John Kirchenbauer is a language-model researcher and first author of "A Watermark for Large Language Models", an Outstanding Paper at ICML 2023. He studied violin at the Oberlin Conservatory of Music and computer science at Washington University in St. Louis, then completed a PhD at the University of Maryland with Tom Goldstein in 2026. Later work includes a study of watermark reliability and the goldfish loss for limiting memorization, both in 2024, and the Common Pile dataset of openly licensed text in 2025. Since August 2026, he is a Vector Distinguished Postdoctoral Fellow at the Vector Institute in Toronto, working with Colin Raffel on fully open language model development and mitigating the risks posed by modern AI.

Badges

Timeline

  1. 2017–2020

    From violin to computer science

    He earns a Performance Diploma in Violin from the Oberlin Conservatory of Music in 2017, and BS and MS degrees in computer science from Washington University in St. Louis in 2020.

  2. By 2023

    A PhD in Goldstein's lab

    By 2023, he is a PhD student in Tom Goldstein's lab at the University of Maryland, and he interns at Google X during the degree. His earlier post is research engineer at Carnegie Mellon University's Software Engineering Institute.

  3. 2023

    A watermark for language models

    In 2023, he is first author of "A Watermark for Large Language Models", written with Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. The method selects a randomized set of "green" tokens before each word is generated and softly promotes them during sampling, so a statistical test can detect the watermark in a short span of tokens without access to the model's API or parameters.

  4. 2023

    An ICML Outstanding Paper

    The watermark paper receives an Outstanding Paper Award at ICML 2023, one of six chosen from 1,865 accepted papers.

  5. 2024

    Watermark reliability and goldfish loss

    In 2024, he publishes "On the Reliability of Watermarks for Large Language Models" at ICLR 2024, which finds that watermarks remain detectable after human and machine paraphrasing. He also coauthors "Be like a Goldfish, Don't Memorize!" at NeurIPS 2024, which introduces the goldfish loss, a training objective that excludes a random subset of tokens from the loss computation to prevent verbatim reproduction of training text.

  6. 2025

    The Common Pile dataset

    In 2025, he is one of 27 authors of The Common Pile v0.1, an eight-terabyte collection of openly licensed text from 30 sources designed for LLM pretraining. Two 7-billion-parameter models trained on it, Comma v0.1-1T and Comma v0.1-2T, reach performance competitive with LLMs trained on unlicensed text, such as Llama 1 and 2 7B.

  7. May 2026

    Defending a Maryland dissertation

    In May 2026, he defends his dissertation, "Enhancing Trust and Transparency in Language Model Development and Deployment", and completes his University of Maryland PhD with Tom Goldstein in 2026. The thesis opens with a method for watermarking language-model outputs and continues with training data attribution, memorization, and proactive watermarking by content creators to make training data membership testing more tractable.

  8. August 2026

    Postdoctoral fellow at Vector

    In August 2026, he starts a postdoc at the Vector Institute in Toronto, working with Colin Raffel. As a Vector Distinguished Postdoctoral Fellow, he works on fully open language model development and mitigating the risks posed by modern AI.

Career

  1. Since 2026Vector InstituteVector Distinguished Postdoctoral Fellow
  2. Until 2023GoogleIntern
  3. Carnegie Mellon UniversityResearch Engineer

Education

  1. 2026University of MarylandPhD
  2. 2020Washington University in St. LouisBS and MS in computer science
  3. 2017Oberlin Conservatory of MusicPerformance Diploma in Violin

From the biography.

  1. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes John Kirchenbauer. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  2. 16% of AI career moves in the 2020s went from a university to a companyIncludes John Kirchenbauer. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  3. 23% of people shaping AI who left a university for a company later returned to a universityIncludes John Kirchenbauer. We estimate that 468 of the 2,011 people shaping AI who left a university for a company (23%) later returned to a university after a median of 5 years, and 163 of 829 (20%) made the reverse round trip.
  4. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes John Kirchenbauer. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of John Kirchenbauer yet. Send us a free-licensed photo.

  1. John Kirchenbauer - Vector Institute for Artificial Intelligence vectorinstitute.ai. Checked 21 October 2026.
  2. PhD Defense: Enhancing Trust and Transparency in Language Model Development and Deployment (UMD Computer Science talks) talks.cs.umd.edu. Checked 21 October 2026.
  3. UMD Researchers Strong Showing with 24 Papers and an Outstanding Paper Award at ICML 2023 cs.umd.edu. Checked 21 October 2026.
  4. A Watermark for Large Language Models proceedings.mlr.press. Checked 21 October 2026.
  5. ICML 2023 Awards Detail icml.cc. Checked 21 October 2026.
  6. On the Reliability of Watermarks for Large Language Models (arXiv:2306.04634) arxiv.org. Checked 21 October 2026.
  7. Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs proceedings.neurips.cc. Checked 21 October 2026.
  8. The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text (arXiv:2506.05209) arxiv.org. Checked 21 October 2026.
  9. John Kirchenbauer jwkirchenbauer.github.io. Checked 21 October 2026.