Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Correspondence
  • Published:

Large language models should be used as scientific reasoning engines, not knowledge databases

Nature Medicine volume 29, pages 2983–2984 (2023)Cite this article

Subjects

Large language models (LLMs) such as GPT-4 have emerged into the public view in recent months. These artificial intelligence (AI) models boast impressive capabilities1,2 and can analyze and write texts with human-like performance, including for medical and scientific applications3,4. LLMs can be used to respond to student assignments and to scientific and medical examination questions, and to create text for research proposals or scientific publications5. Yet LLMs often fall short when it comes to factual accuracy6. These models have been known to generate false statements that do not reflect scientific facts7, known as ‘hallucinations’. For example, LLMs can hallucinate about the chemical structure of molecules or add nonexistent citations to scientific text7, potentially leading to the spread of misinformation. LLMs may also misinterpret the relevance of varying levels of evidence, inadvertently promoting concepts that may have been ruled out or deemed inconclusive in the original source. Their inconsistency and omission of facts is likely to hinder the use of LLMs in clinical practice1 and in scientific work.

We have observed, however, that this negative outlook for LLMs arises from users treating the models as knowledge databases. We contend that this is not the best use of LLMs; they should instead be best used as reasoning engines or inference engines, which can provide extra-human cognition. LLMs can perform logic, analogy, causal reasoning, extrapolations and evidence evaluation. This is despite the models being trained solely on a large body of textual data, and not specifically in solving complex logical problems4.

This is a preview of subscription content, access via your institution

Relevant articles

Open Access articles citing this article.

Access options

Buy this article

  • Purchase on Springer Link
  • Instant access to full article PDF
Buy now

Prices may be subject to local taxes which are calculated during checkout

References

  1. Anonymous. Nat. Med. 29, 505–506 (2023).

    Article  Google Scholar 

  2. Anonymous. Nat. Cancer 4, 151–152 (2023).

    Article  Google Scholar 

  3. Singhal, K. et al. Nature 620, 172–180 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Bubeck, S. et al. Preprint at https://doi.org/10.48550/arXiv.2303.12712 (2023).

  5. Sanderson, K. Nature 616, 639–640 (2023).

    Article  CAS  PubMed  Google Scholar 

  6. Zhavoronkov, A. Nat. Med. 29, 532 (2023).

    Article  CAS  PubMed  Google Scholar 

  7. Sanderson, K. Nature 615, 773 (2023).

    Article  CAS  PubMed  Google Scholar 

Download references

Acknowledgements

The authors disclose the use of GPT-4 (OpenAI) to check spelling and grammar during the writing of this article.

Author information

Authors and Affiliations

  1. Department of Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen, Germany

    Daniel Truhn

  2. Experimental Pathology, Department of Pathology, Memorial Sloan Kettering Cancer Center, New York, NY, USA

    Jorge S. Reis-Filho

  3. Else Kroener Fresenius Center for Digital Health, Technical University Dresden, Dresden, Germany

    Jakob Nikolas Kather

Authors
  1. Daniel Truhn
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Jorge S. Reis-Filho
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Jakob Nikolas Kather
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Jakob Nikolas Kather.

Ethics declarations

Competing interests

J.S.R.-F. reports a leadership (board of directors) role at Grupo Oncoclinicas, stock or other ownership interests at Repare Therapeutics and Paige.AI, and a consulting or advisory role at Genentech/Roche, Invicro, Ventana Medical Systems, Volition RX, Paige.AI, Goldman Sachs, Bain Capital, Novartis, Repare Therapeutics, Lilly, Saga Diagnostics, MultiplexDX and Personalis. J.N.K. declares consulting services for Owkin (France), DoMore Diagnostics (Norway) and Panakeia (UK) and receipt of honoraria for lectures from Eisai, Roche, Bayer, MSD and Fresenius. D.T. and J.N.K. hold shares in StratifAI GmbH, Germany.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Truhn, D., Reis-Filho, J.S. & Kather, J.N. Large language models should be used as scientific reasoning engines, not knowledge databases. Nat Med 29, 2983–2984 (2023). https://doi.org/10.1038/s41591-023-02594-z

Download citation

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s41591-023-02594-z

This article is cited by

Search

Advanced search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing