PeopleofSuperintelligence

Researcher

Patrick Schramowski

Patrick Schramowski is a researcher at DFKI whose work on the ethics and safety of large AI models runs from language models to image generators. After an M.Sc. in computer science at the University of Dortmund in 2017 and early work at start-ups, he did his Ph.D. at TU Darmstadt's Machine Learning Lab from 2017 to 2023 in Kristian Kersting's group. A 2022 Nature Machine Intelligence paper he led found a moral direction inside language models, and he coauthored LAION-5B, which won a NeurIPS 2022 Outstanding Paper award. He then built safeguards for image data and generators, including Safe Latent Diffusion at CVPR 2023 and LlavaGuard at ICML 2025. He is a research group leader at DFKI.

Now at
DFKI
Badges

Timeline

  1. 2011–2017

    Dortmund studies and start-ups

    He studies computer science at the University of Dortmund from 2011 and receives his M.Sc. there in 2017. Alongside the degree he is CTO of MateApps GmbH from 2014 to 2016 and co-founder of Pflegix GmbH in 2016.

  2. 2017–2023

    A PhD at TU Darmstadt

    He joins the Machine Learning group at TU Darmstadt in 2017 and works there as a Ph.D. student at the Machine Learning Lab until 2023. He is part of Kristian Kersting's Artificial Intelligence and Machine Learning group, and his research explores the ethical implications of large-scale, self-supervised, transformer-based models.

  3. 2022

    A moral direction in language models

    With Cigdem Turan, Nico Andersen, Constantin Rothkopf and Kristian Kersting, he reports in Nature Machine Intelligence in 2022 that language models such as BERT hold a dimension running from good actions to bad actions. The approach works with any language model and needs no access to its parameters, and it can swap verbs to make a sentence less offensive.

  4. 2022

    LAION-5B wins Outstanding Paper

    He is one of 16 authors of LAION-5B, an open dataset of 5.85 billion CLIP-filtered image-text pairs, which receives a NeurIPS 2022 Outstanding Paper award. He adds a pretrained model that filters potentially offensive, threatening or disturbing content from the data.

  5. 2022–2023

    Safe Latent Diffusion

    With Manuel Brack, Björn Deiseroth and Kristian Kersting, he introduces safe latent diffusion in 2022, together with the I2P test bed of real-world prompts covering concepts such as nudity and violence. The method removes inappropriate image parts during the diffusion process without additional training, and the paper appears at CVPR 2023.

  6. 2023

    Supervising the LeoLM models

    LAION releases LeoLM in 2023, a suite of German-language foundation models built on Llama-2 and trained on hessian.AI's supercomputer 42. Björn Plüster leads the project, and he is one of three researchers who give it support, guidance and supervision.

  7. 2024–2025

    LlavaGuard image safeguards

    With Lukas Helff, Felix Friedrich, Manuel Brack and Kristian Kersting, he coauthors LlavaGuard, posted in 2024 and published at ICML 2025. It is a suite of vision-language-model safeguards from 0.5B to 7B parameters that rate the safety of images against flexible policies.

  8. As of October 2026

    Research group leader at DFKI

    He is a research group leader at DFKI, where he has been a researcher since 2022. In 2026 he coauthors No Safe Dose with Felix Friedrich, Lukas Helff, Niharika Hegde and Kristian Kersting, which finds that a text-to-image model's output unsafety rises from 16.6% with no unsafe training images to 25.5% with 5% unsafe images.

Career

  1. Since 2022DFKIResearch Group Leader
  2. From 2016Pflegix GmbHCo-founder
  3. 2014–2016MateApps GmbHCTO

Education

  1. 2023TU DarmstadtPh.D.
  2. 2017TU Dortmund UniversityM.Sc.

From the biography.

  1. 11% of AI doctorate holders from 2010 to 2016 founded or led a company or lab by 10 years after graduationIncludes Patrick Schramowski. We estimate how often AI doctorate holders founded or led a company or lab by 10 years after graduation, how waiting times compare between founders and top jobs, and which universities trained the most.
  2. Nonprofit roles are 2.3 times as common in AI safetyIncludes DFKI. We estimate where people focused on safety, alignment and evaluation work. Nonprofit roles appear for 25.1% of safety people with known affiliations, compared with 11.0% in other areas.
  3. The top 10 organizations employed 22% of people shaping AI in 2025: 51% of company researchers and 21% of university researchersIncludes DFKI. We estimate the share of people shaping AI who worked at the top 10 organizations each year from 1955 to 2025, comparing concentration across companies, universities, and institutes and labs.
  4. The United States employs an estimated 48% of people shaping AIIncludes DFKI. We estimate where the people shaping AI work today, how universities, companies and research institutes split across countries, and how country shares have changed since the 1960s.

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 Patrick Schramowski yet. Send us a free-licensed photo.

  1. Patrick Schramowski aiml.informatik.tu-darmstadt.de. Checked 22 October 2026.
  2. Mit Künstlicher Intelligenz gegen anstößige Inhalte – TU Darmstadt tu-darmstadt.de. Checked 22 October 2026.
  3. Tactful machines – TU Darmstadt tu-darmstadt.de. Checked 22 October 2026.
  4. LAION-5B: An open large-scale dataset for training next generation image-text models arxiv.org. Checked 22 October 2026.
  5. NeurIPS 2022 Awards neurips.cc. Checked 22 October 2026.
  6. Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models arxiv.org. Checked 22 October 2026.
  7. LeoLM: Igniting German-Language LLM Research | LAION laion.ai. Checked 22 October 2026.
  8. LlavaGuard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models arxiv.org. Checked 22 October 2026.
  9. No Safe Dose: How Training Data Drives Unsafe Image Generation arxiv.org. Checked 22 October 2026.