About the project

A Sloan Foundation project on what errors in AI-driven modeling can teach science: where it came from, what we ask, and how to take part.

This project grew out of several lines of work that were each moving toward the same question. Here’s how they came together, and where we’re going next.

  1. error-in-science

    A philosopher sorts the errors that can creep into an experiment by the stage at which they arise. Hon 1989

  2. ai-climate

    Climate scientists start running their models through the same experiments so they can be compared: CMIP, the Coupled Model Intercomparison Project, still going three decades later. Meehl et al. 2000

  3. error-in-science

    A hypothesis is well supported only if it has passed tests likely to catch errors: the idea of severe testing. Mayo 1996

  4. opacity

    Computer simulations are called “epistemically opaque”: no person can trace every step from input to output. Humphreys 2004

  5. error-in-science

    Nineteenth-century microscopists turned a mistake, nerve fibers damaged by handling, into a test for identifying nerve tissue. An error became a tool. Schickore 2005

  6. credibility

    Parts of climate models become meeting points between field scientists and modelers, binding the two groups together. Sundberg 2007

  7. error-in-science

    Errors in computer models come from many sources: code bugs, approximations, data, and the choices built into the model itself. Parker 2008

  8. opacity

    Skeptics argue climate models are so tangled that no success or failure can be pinned on any one part. Lenhard and Winsberg 2010

  9. error-in-science

    Some systems don’t just survive shocks and mistakes; they gain from them. Taleb calls this antifragility. Taleb 2012

  10. error-in-science

    Failure is central to how science succeeds. Firestein 2015

  11. credibility

    Climate modeling is a social enterprise: its knowledge is built by teams and communities, not lone modelers. Winsberg 2018

  12. ryan

    Ryan begins interviews and site visits with climate scientists at NSF NCAR and NASA GISS.

    Talks: NSF NCAR, 2022 NASA GISS, 2023

  13. matt

    Matt shows that metadata and documentation are how scientists record what they know about their data and models, including known limits and errors. Mayernik 2019

  14. opacity

    Explanations bolted onto AI models after training can mislead; build understandable models instead. Rudin 2019

  15. credibility

    CMIP shapes climate research, but what each team learns about its own model tends to stay with that team. Touzé-Peiffer et al. 2020

  16. matt

    Model comparisons earn trust by coupling institutions (rules, governance) with data infrastructure. Credibility belongs to the community, not any single model. Mayernik 2021

  17. opacity

    Explainable AI arrives in climate science, but different explanation methods can disagree about the same model. Others ask what would make such explanations genuine. Mamalakis et al. 2022; Buijsman 2022

  18. ai-climate

    FourCastNet, an AI weather model, makes fast global forecasts. Pathak et al. 2022

  19. ryan

    Ryan shows that climate scientists can and do trace model errors to their sources, using physical reasoning, simplification and carefully varied reruns, answering the skeptics. O’Loughlin 2023

  20. ai-climate

    FourCastNet’s instability, which kept it to forecasts under ten days, is traced to math built for a flat map. Rebuilt for a sphere, it runs a year. Bonev et al. 2023

  21. credibility

    In applied data science work, errors become “a site of collaboration” that reveals and reshapes how people work together. Lin and Jackson 2023

  22. opacity

    Climate scientists publish guides for choosing among explainable-AI methods, a sign of how widely they’re now used. Bommer et al. 2024

  23. ryan

    Ryan and coauthors urge AI climate research to look beyond after-the-fact explanations, toward knowing which part of a model does what. O’Loughlin et al. 2025

  24. ai-climate

    An AI forecast model goes into operational use at ECMWF, Europe’s weather center, alongside its physics-based one. ECMWF 2025

  25. credibility

    Climate model developers describe debugging that resists standard procedures, aiming for models that are “good enough”. Proske and Melsen 2025

  26. ryan

    Pilot interviews at NSF NCAR. A first finding: scientists often avoid the word “error”, preferring “bias”, “limitation” or “statistical distinguishability”. Parker 2008 as the starting typology

  27. ryan

    Ryan identifies four kinds of antifragility in modeling: flexibility, new hypotheses, learning from error signals, and credibility won by overcoming failure. O’Loughlin, manuscript

  28. error-in-science

    Allchin argues for “negative knowledge”: knowing about errors, and how they were corrected, is knowledge too. Allchin 2026

  29. merge into main

    These lines of work were all moving toward one question none had asked directly: how can errors, and the ways scientists handle them, make AI-driven science better?

  30. main

    The team forms: Ryan O’Loughlin (philosophy of climate modeling, Queens College, CUNY) and Matthew Mayernik (data and credibility, NSF NCAR), with an advisory board of philosophers, STS scholars and scientists.

  31. main

    The Alfred P. Sloan Foundation funds the project: two years, through 30 September 2028.

  32. next

    Interview plan finished and ethics approval (IRB) obtained.

  33. next

    The advisory board meets twice a year.

  34. next

    Interviews begin at NSF NCAR, the LEAP Center and NASA GISS, with document analysis alongside.

    Work with AI models? Share an error story

  35. next

    Call for extended abstracts and invited speakers.

    Submit an abstract

  36. next

    First round of interviews complete.

  37. next

    Two-day workshop for philosophers, STS scholars and scientists, with funding for PhD students and early-career researchers.

    Come to the workshop

  38. next

    First findings presented at a conference.

  39. next

    Follow-up conversations with interviewees; the framework takes shape.

    Give feedback on the framework

  40. next

    Papers submitted: two in philosophy of science, one in STS, one for climate scientists and one for the public, plus a workshop report.

    Get updates

  41. next

    Project ends.

  42. next

    A longer-term research program on error, AI and scientific modeling, reaching beyond climate science.

    Using AI in another science? Collaborate with us

The questions

Understanding

When scientists diagnose why an AI model failed, what do they come to understand about the model, and what does that mean for explainable AI?

Antifragility

When do errors leave models and research communities stronger, not just repaired, and better off than immediate success would have?

Credibility

How do the ways teams diagnose, document and share errors build or erode trust in the science and data they produce?

  1. How do errors in AI-driven models contribute to or constrain scientific understanding, and what do practices of error diagnosis reveal about model explainability?
  2. In what ways do modeling communities develop antifragile responses to error, using failures to strengthen rather than merely repair their systems?
  3. How do errors and their management shape the credibility of AI-driven models and the social organization of modeling teams?

How we work

Interviews

20 or more people at each of NSF NCAR, the LEAP Center and NASA GISS: scientists, software engineers and data specialists.

Documents

Reports, papers, code repositories, issue trackers and dataset records, to see how errors get written down, or don’t.

Analysis

Bringing what we hear into conversation with philosophy of science and STS, to say when errors help science.

Workshop

Two days for philosophers, STS scholars and scientists to test the framework.

Take part

Share an error story

Work with AI-augmented models? Tell us about a time one went wrong, and what happened next.

Share a story

Come to the workshop

For philosophers, STS scholars and scientists, with funding for PhD students and early-career researchers. The call opens in February 2027.

Workshop details

Give feedback

Read our framework as it develops and tell us where it’s wrong.

Get updates

Bring another science

Similar questions are arising in genomics, astronomy and beyond. If you work there, we’d like to compare notes.

Get in touch

Interviews are voluntary and confidential, and no one is named without their consent.

Acknowledgments

Funding
This project is supported by the Alfred P. Sloan Foundation (2026–2028).
Institutional home
Research Foundation of CUNY – Queens College, City University of New York
Partner institution
NSF National Center for Atmospheric Research, home of co-PI Matthew Mayernik
Research sites
The scientists and institutions who make our fieldwork possible.
Advisory board
Thanks to our advisory board for their guidance.
Interviewees
Thanks to everyone who shares their experience with us. Interviewees are not named.

References

  • Allchin, D. (2026). Toward a philosophy of error in science. Oxford University Press.
  • Bommer, P. L., Kretschmer, M., Hedström, A., Bareeva, D., & Höhne, M. M.-C. (2024). Finding the right XAI method: A guide for the evaluation and ranking of explainable AI methods in climate science. Artificial Intelligence for the Earth Systems, 3(3). https://doi.org/10.1175/AIES-D-23-0074.1
  • Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., & Anandkumar, A. (2023). Spherical Fourier neural operators: Learning stable dynamics on the sphere. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 2806–2823. https://proceedings.mlr.press/v202/bonev23a.html
  • Buijsman, S. (2022). Defining explanation and explanatory depth in XAI. Minds and Machines, 32(3), 563–584. https://doi.org/10.1007/s11023-022-09607-9
  • ECMWF. (2025, February 25). ECMWF’s AI forecasts become operational. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
  • Firestein, S. (2015). Failure: Why science is so successful. Oxford University Press.
  • Hon, G. (1989). Towards a typology of experimental errors: An epistemological view. Studies in History and Philosophy of Science Part A, 20(4), 469–504. https://doi.org/10.1016/0039-3681(89)90020-4
  • Humphreys, P. (2004). Extending ourselves: Computational science, empiricism, and scientific method. Oxford University Press.
  • Lenhard, J., & Winsberg, E. (2010). Holism, entrenchment, and the future of climate model pluralism. Studies in History and Philosophy of Modern Physics, 41(3), 253–262. https://doi.org/10.1016/j.shpsb.2010.07.001
  • Lin, C. K., & Jackson, S. J. (2023). From bias to repair: Error as a site of collaboration and negotiation in applied data science work. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1), 1–32. https://doi.org/10.1145/3579607
  • Mamalakis, A., Barnes, E. A., & Ebert-Uphoff, I. (2022). Investigating the fidelity of explainable artificial intelligence methods for applications of convolutional neural networks in geoscience. Artificial Intelligence for the Earth Systems, 1(4). https://doi.org/10.1175/AIES-D-22-0012.1
  • Mayernik, M. S. (2019). Metadata accounts: Achieving data and evidence in scientific research. Social Studies of Science, 49(5), 732–757. https://doi.org/10.1177/0306312719863494
  • Mayernik, M. S. (2021). Credibility via coupling: Institutions and infrastructures in climate model intercomparisons. Engaging Science, Technology, and Society, 7(2), 10–32. https://doi.org/10.17351/ests2021.769
  • Mayo, D. G. (1996). Error and the growth of experimental knowledge. University of Chicago Press.
  • Meehl, G. A., Boer, G. J., Covey, C., Latif, M., & Stouffer, R. J. (2000). The Coupled Model Intercomparison Project (CMIP). Bulletin of the American Meteorological Society, 81(2), 313–318. https://doi.org/10.1175/1520-0477(2000)081%3C0313:TCMIPC%3E2.3.CO;2 · Free copy (PDF, NOAA GFDL)
  • O’Loughlin, R. (2023). Diagnosing errors in climate model intercomparisons. European Journal for Philosophy of Science, 13(2). https://doi.org/10.1007/s13194-023-00522-z
  • O’Loughlin, R. (in preparation). Manuscript on antifragility in scientific modeling.
  • O’Loughlin, R. J., Li, D., Neale, R., & O’Brien, T. A. (2025). Moving beyond post hoc explainable artificial intelligence: A perspective paper on lessons learned from dynamical climate modeling. Geoscientific Model Development, 18(3), 787–802. https://doi.org/10.5194/gmd-18-787-2025
  • Parker, W. S. (2008). Computer simulation through an error-statistical lens. Synthese, 163(3), 371–384. https://doi.org/10.1007/s11229-007-9296-0
  • Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., Hassanzadeh, P., Kashinath, K., & Anandkumar, A. (2022). FourCastNet: A global data-driven high-resolution weather model using adaptive Fourier neural operators. arXiv:2202.11214. https://arxiv.org/abs/2202.11214
  • Proske, U., & Melsen, L. A. (2025). How climate model developers deal with bugs. Earth’s Future, 13(8), e2025EF006318. https://doi.org/10.1029/2025EF006318
  • Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. https://doi.org/10.1038/s42256-019-0048-x
  • Schickore, J. (2005). ‘Through thousands of errors we reach the truth’—but how? On the epistemic roles of error in scientific practice. Studies in History and Philosophy of Science Part A, 36(3), 539–556. https://doi.org/10.1016/j.shpsa.2005.06.011
  • Sundberg, M. (2007). Parameterizations as boundary objects on the climate arena. Social Studies of Science, 37(3), 473–488. https://doi.org/10.1177/0306312706075330
  • Taleb, N. N. (2012). Antifragile: Things that gain from disorder. Random House.
  • Touzé-Peiffer, L., Barberousse, A., & Le Treut, H. (2020). The Coupled Model Intercomparison Project: History, uses, and structural effects on climate research. WIREs Climate Change, 11(4), e648. https://doi.org/10.1002/wcc.648
  • Winsberg, E. (2018). Philosophy and climate science. Cambridge University Press. https://doi.org/10.1017/9781108164290