Editorial

AI, libraries, and literacy: revisiting the Journal of the Medical Library Association's policy on the use of generative artificial intelligence in submissions

Alexander J. Carroll, AHIP¹, Jill T. Boruff, AHIP², Michelle Kraft, AHIP, FMLA³

doi: http://dx.doi.org/10.5195/jmla.2026.2516

Volume 114, Number 4: 330-335
Received 05 2026; Accepted 06 2026

ABSTRACT

In the fall of 2023, the Journal of the Medical Library Association (JMLA) introduced its policy on generative artificial intelligence (AI) submissions, predicting that these tools would become embedded within the workflows of clinicians, educators, researchers, and students. Those predictions dramatically underestimated AI's impact. On the three-year anniversary of that editorial, we revisit the issue in light of AI's unprecedentedly rapid adoption across knowledge work. We examine the educational challenges AI poses, arguing that excessive cognitive offloading threatens the deep thinking on which learning depends. We then consider the challenges AI creates for researchers, from hallucinated citations to fabricated diseases. We suggest that health sciences librarians, who offer expertise in approaching information systems with discernment, are well suited to be leaders in this space. To help prospective JMLA authors weigh whether and when to use AI, we propose the Compliance, Verifiability, and Disclosure (CVD) Framework. Throughout, we reaffirm our conviction that education and research must remain human-centered endeavors.

Keywords: Artificial Intelligence; information literacy; AI in Libraries; AI literacy; literacy skills; lifelong learning; scholarly communication; scholarly publishing; scholarly writing.

INTRODUCTION

In the fall of 2023, the Journal of the Medical Library Association introduced its policy on generative artificial intelligence (AI) submissions. In our editorial introducing this policy, we predicted that AI applications would become embedded into the workflows of clinicians, educators, researchers, and students, which would in turn affect health sciences librarianship [1]. Although our policy advocated against unrestricted use of AI tools during manuscript preparation, we recognized the transformative potential of these tools and encouraged the JMLA community to explore using them.

Our predictions dramatically underestimated AI’s impact [2, 3]. The speed of these tools adoption is unprecedented [4]; by December 2024, ChatGPT users sent more than one billion messages each day, just over two years after its public release[5]. By comparison, it took Google nearly eight years to reach this same one billion daily uses benchmark [6]. In July 2025, over 750 million accounts were using ChatGPT each week, which represents nearly 10% of the global adult population [7]. Rogers’ diffusion of innovation theory suggests that AI usage would be concentrated among a handful of early adopters [8], yet these tools are now being used widely among different demographic groups across the world [9]. This widespread integration of AI tools runs contra to early forecasts of how AI adoption would manifest among workers generally and librarians in particular [10, 11].

However, adoption of AI tools in the workplace is not only a reflection of individual workers’ preferences. Businesses across many industries, including healthcare, are partnering with vendors to embed AI within their own networks. Firms investing in AI infrastructure are increasingly expecting their workers to demonstrate how they are using AI, with some technology firms evaluating individual worker productivity based on their quantity of AI tool usage. For those of us familiar with Goodhart’s Law, this has led to the predictable result of workers performatively using AI in order to appear more productive [12, 13]. Although affirmative adoption of AI remains contested within librarianship [14, 15], AI applications have become so thoroughly integrated into enterprise software that refusal to use AI is no longer tenable for many information professionals [16]. These tools’ rapid ascension has prompted us to revisit generative AI on the three-year anniversary of our previous editorial on AI.

EDUCATIONAL CHALLENGES CREATED BY AI TOOLS

The most advanced AI models appear superintelligent, responding accurately and rapidly to complex questions designed to assess clinician reasoning skills [17]. Although the most robust models often require subscription fees to access, even free versions of AI tools can interpret and answer quantitative and qualitative questions alike [18].

Yet the ubiquity of AI has proven double-edged: AI’s presence within workplaces and classrooms invites workers and students to become overly reliant upon it [19]. Early evidence on the integration of AI into workplaces and educational settings suggests that using these tools does not facilitate permanent skill development [20], and in fact may inhibit learning new skills [21].

That AI tools could inhibit learning follows from our understanding of how humans learn: deep thinking is essential for learning new things [22]. Repeated exposure to concepts and intentional practice applying skills improves retention and transfer [23], which are essential for promoting lifelong learning [24]. However, thinking is a cognitively intensive task that can be psychologically unpleasant in the short term, so learners often seek to avoid cognitively challenging tasks when possible. Educators account for this by designing “desirable difficulties,” assignments that push students to persevere past these difficult thresholds and to achieve real learning [25]. AI tools can disrupt this model of learning, as they present a ready means for students to solve assignments that instructors use to create desirable difficulty.

AI tools’ capacity to trivialize major assignments has led to concerns that AI will promote maladaptive, excessive cognitive offloading behaviors among students. Excessive use of these tools within classroom environments may limit students’ ability to internalize knowledge and to think critically [26]. Excessive cognitive offloading is most likely to occur when novices are allowed to use AI tools without guardrails. Large language models have a handful of critical limitations baked into their probabilistic design that ensure that they can never be unfailingly correct in all instances [27]. However, novices learning about a new topic lack the prior knowledge needed to recognize when an AI is offering mistakes in facts or interpretations [28]. This portends a vicious cycle: students who become reliant upon AI will fail to gain the lasting, broad-based knowledge of the world needed for critical thinking, which will in turn push the learners to use AI for even more tasks.

Yet cognitive offloading is not inherently maladaptive [29]. Because the human’s brain capacity to store information within long-term, short-term, and working memory is limited [30], without cognitive offloading, humans’ ability to perform analytical reasoning is limited by the knowledge they can store in their mind at one time [31]. The development of literacy enabled humans to write down and physically store knowledge elsewhere until it was needed, leading to the creation of libraries and librarianship [32]. The human brain’s ability to use literacy to offload and onramp knowledge is why libraries eventually became essential infrastructure for modern health care and the scientific research enterprise [27]. As curators of cognitively offloaded knowledge, librarians have a clear role to play in guiding novices’ use of AI tools.

RESEARCH CHALLENGES CREATED BY AI TOOLS

While the term “novices” often connotes students, researchers publishing outside of their narrow band of expertise can cognitively behave like novices. The downstream effects of researchers using AI tools to publish outside their areas of expertise are no longer hypothetical. Hallucinated citations to non-existent papers are now a common occurrence within published papers [33]. AI tools have been linked to a massive increase in low-quality manuscripts [34], a trend so intense within computer science that the preprint server arXiv suspended accepting review article submissions in early 2026 [35]. Perhaps most concerningly for health information professionals, intentionally fake diseases loaded into large language model’s training data are now cited as true conditions [36]. Because AI tools are correct in most instances, students and researchers alike can become accustomed to trusting the validity of their outputs, even if they know that these tools can facilitate fraud and make mistakes that are obvious to a human expert.

Yet despite these negative externalities, the scientific research community remains split on the role AI should have within research workflows. These opinions range from viewing AI as antithetical to the humanistic enterprise of genuine scientific inquiry [37], believing AI will transform every aspect of research [38], and everywhere in between [39]. Given these deltas, consensus among researchers on the appropriate roles for AI in research is unlikely. This creates another opportunity librarians and information scientists to take a leading role in guiding ethical adoption of AI.

IMPLICATIONS FOR HEALTH SCIENCES LIBRARIANS

As domain experts in information systems, librarians and information professionals are poised to become leaders in how AI is implemented into research workflows. But to earn a space at the leadership table, librarians must be able to articulate clear guidance on AI use that is grounded in our broad conceptual knowledge about information architectures, our deep expertise in navigating information systems, and our own professional ethics. We suggest three criteria for information professionals to foreground during discussions over whether to use AI tools in research workflows: compliance, verifiability, and disclosure. Though not comprehensive of all ethical considerations, the Compliance, Verifiability, and Disclosure (CVD) Framework provides pragmatic, domain-agnostic guidance for weighing whether and when to use AI tools within research. These criteria are summarized in Table 1.

Compliance recognizes that practical or legal constraints limit tasks that AI can be used for. Local institutional review boards, and journal editorial staffs, have wide ranging policies that span from extremely restrictive to very permissive. Prior to using AI, researchers should ascertain whether their data or other materials are permitted to be uploaded into a large language model in the first place.

Verifiability recognizes that a researcher’s prior knowledge and existing expertise should dictate what tasks an AI performs within a research process. This consideration suggests that researchers should only use AI tools to complete tasks they already know how to do. If a researcher cannot perform a given task using other means, then they cannot be confident in the validity of the AI’s output. Moreover, verifiability requires researchers to critically evaluate AI outputs against trusted, gold standard sources of evidence in addition to relying upon their own expertise [40].

Disclosure recognizes that researchers have an intuitive, personal sense of what tasks an AI should or should not perform on their behalf. If a researcher feels hesitant to acknowledge their use of AI to perform a task, it indicates their own recognition that they ought not to have used AI to do it.

GUIDANCE FOR PROSPECTIVE JMLA AUTHORS

We remain skeptical of unrestricted use of AI tools within research. The JMLA editorial team wants its authors to grow and learn through process of conducting practitioner-based research. We see every task offloaded to an AI as a lost opportunity to potentially learn something new. To that end, we view writing as a unique and valuable form of thinking [41], through which a researcher decides which information to share with readers and how best to communicate it. Engaging deeply in the writing process is an essential step for understanding your own findings and for identifying which components of your work are worth highlighting. Authors who elect not to use AI when writing their manuscripts will likely have a richer understanding of their own work and may be better equipped to respond to reviewers’ comments and critiques.

For authors who wish to use AI tools as part of their workflows, we encourage JMLA authors to carefully consider the CVD Framework described in Table 1. Consult your local institutional review board prior to instigating a study. Be careful about offloading tasks to an AI that you do not already know yourself. Be prepared to proactively disclose, in detail, the full extent of how AI was used during your study planning, data collection, data analysis, and manuscript preparation.

Authors that elect to use AI must ensure that, as the human authors assuming responsibility for the content of their manuscript, they retain authorial control of their research outputs. Table 2 offers a comparison of Not Recommend Practices and Recommended Practices. While opinions on these specific examples will vary, we believe these are emblematic of our broader belief that research within health information science is a human-centered process: performed by humans for the benefit of humans.

For JMLA authors who plan to use AI tools as their research assistants, copyeditors, and internal reviewers, we close with one final note to consider: large language models lack an analytic, human mind [42]. While AI tools invite anthropomorphism through their use of natural language and their impressive capabilities across domains [42], large language models cannot read [44]. Their ability to use word sequences to detect patterns within text is remarkable, but it is categorically a different process than how humans learn to read [45]. Likewise, although large language models can rapidly produce volumes of text, the language model has no understanding of what these words mean; the text they create is best described as what Shakespeare’s King Claudius called “words without thoughts” [46]. Though AI can produce literary text rich with simile and metaphor [47], the sui generis figures of speech created by AI lack the grounded human meaning that made The Bard’s dialogue so inimitable [48].

Conducting original research is an opportunity for librarians to make their mark on the research record, and to make the case that humans are central to the work of information and data. Researchers who offload their brainstorming, hypothesizing, and interpreting onto a large language model are taking ideas from a machine that cannot think; authors who offload their editing process onto a large language model are taking advice from a machine that can neither read nor write. We urge caution against removing ourselves out of these processes entirely.

REFERENCES

1. Boruff JT, Kraft M, Carroll AJ. Introducing the Journal of the Medical Library Association’s policy on the use of generative artificial intelligence in submissions. J Med Libr Assoc. 2023 Oct 2;111(4):4. DOI: http://dx.doi.org/10.5195/jmla.2023.1826
.

2. Deming D. How People Use ChatGPT [Substack newsletter]. Forked Lightning [Internet]. 2025 Sep 15 [cited 2026 Apr 14]. Available from: https://forklightning.substack.com/p/how-people-use-chatgpt.

3. Deming D. ChatGPT really does offer mundane utility [Substack newsletter]. Forked Lightning [Internet]. 2025 Sep 16 [cited 2026 Apr 14]. Available from: https://forklightning.substack.com/p/chatgpt-really-does-offer-mundane.

4. Bick A, Blandin A, Deming DJ. The Rapid Adoption of Generative AI [Working Paper] [Internet]. National Bureau of Economic Research; 2024 [cited 2026 Apr 14]. (Working Paper Series). Available from: https://www.nber.org/papers/w32966. DOI: 10.3386/w32966.

5. Roth E. The Verge [Internet]. 2024 [cited 2026 Apr 14]. ChatGPT now has over 300 million weekly users. Available from: https://www.theverge.com/2024/12/4/24313097/chatgpt-300-million-weekly-users.

6. Comscore, Inc. [Internet]. Reston, VA; 2007 [cited 2026 Apr 14]. 61 Billion Searches Conducted Worldwide in August. Available from: https://www.comscore.com/Insights/Press-Releases/2007/10/Worldwide-Searches-Reach-61-Billion?utm_source=chatgpt.com.

7. Roth E. The Verge [Internet]. 2025 [cited 2026 Apr 14]. OpenAI says ChatGPT users send over 2.5 billion prompts every day. Available from: https://www.theverge.com/news/710867/openai-chatgpt-daily-prompts-2-billion.

8. Rogers EM. Diffusion of Innovations. 5th ed. New York: Free Press; 2003. 576 p.

9. Chatterji A, Cunningham T, Deming DJ, Hitzig Z, Ong C, Shan CY, et al. How People Use ChatGPT [Working Paper] [Internet]. National Bureau of Economic Research; 2025 [cited 2025 Sep 29]. (Working Paper Series). Available from: https://www.nber.org/papers/w34255. DOI: 10.3386/w34255.

10. Humlum A, Vestergaard E. The unequal adoption of ChatGPT exacerbates existing inequalities among workers. Proc Natl Acad Sci. 2025 Jan 7;122(1):e2414972121. DOI: http://dx.doi.org/10.1073/pnas.2414972121
.

11. Lund BD, Omame I, Tijani S, Agbaji D. Perceptions toward Artificial Intelligence among Academic Library Employees and Alignment with the Diffusion of Innovations’ Adopter Categories. Coll Res Libr. 2020 Jul 1;81(5):865. DOI: http://dx.doi.org/10.5860/crl.81.5.865

12. Rosner-Uddin R. Amazon staff use AI tool for unnecessary tasks to inflate usage scores. Financial Times [Internet]. 2026 May 12 [cited 2026 May 12]. Available from: https://www.ft.com/content/8ee0d3ef-9548-422d-8ff1-ebd48ad4b2ca?syn-25a6b1a6=1.

13. Goodhart CAE. Problems of Monetary Management: The UK Experience. In: Goodhart CAE, editor. Monetary Theory and Practice: The UK Experience [Internet]. London: Macmillan Education UK; 1984 [cited 2023 Apr 26]. p. 91–121. Available from: https://doi.org/10.1007/978-1-349-17295-5_4. DOI: 10.1007/978-1-349-17295-5_4.

14. Pasiak G. AI Fatigue and Vocational Awe in Academic Libraries. The Scholarly Kitchen [Internet]. 2026 Mar 4 [cited 2026 Mar 30]. Available from: https://scholarlykitchen.sspnet.org/2026/03/04/guest-post-ai-fatigue-and-vocational-awe-in-academic-libraries/.

15. Slater K “Kay.” Against AI: Critical Refusal in the Library. Libr Trends. 2025;73(4):588–608.

16. Estlund K, Hudson Vitale C. Strategic Implications of AI Futures for Research Libraries: Workshop Report [Internet]. Association of Research Libraries and Coalition for Networked Information; 2026 Mar [cited 2026 Mar 30]. Available from: https://www.arl.org/resources/strategic-implications-of-ai-futures-for-research-libraries-workshop-report/. DOI: http://dx.doi.org/10.29242/report.futurescape2026
.

17. Kung TH, Cheatham M, Medenilla A, Sillos C, Leon LD, Elepaño C, et al. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digit Health. 2023 Feb 9;2(2):e0000198. DOI: http://dx.doi.org/10.1371/journal.pdig.0000198
.

18. Brynjolfsson E, Li D, Raymond L. Generative AI at Work. Q J Econ. 2025 May 1;140(2):889–942. DOI: http://dx.doi.org/10.1093/qje/qjae044
.

19. Dell’Acqua F, McFowland E, Mollick E, Lifshitz H, Kellogg KC, Rajendran S, et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organ Sci. 2026 Mar;37(2):403–23. DOI: http://dx.doi.org/10.1287/orsc.2025.21838
.

20. Wiles E, Krayer L, Abbadi M, Awasthi U, Kennedy R, Mishkin P, et al. GenAI as an Exoskeleton: Experimental Evidence on Knowledge Workers Using GenAI on New Skills. SSRN; 2024. DOI: https://dx.doi.org/10.2139/ssrn.4944588
.

21. Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proc Natl Acad Sci. 2025 Jul;122(26):e2422633122. DOI: http://dx.doi.org/10.1073/pnas.2422633122
.

22. Willingham DT. Why don’t students like school?: a cognitive scientist answers questions about how the mind works and what it means for the classroom. San Francisco, CA: Jossey-Bass; 2009.

23. Ericsson KA, Krampe RT, Tesch-Römer C. The Role of Deliberate Practice in the Acquisition of Expert Performance. Psychol Rev. 1993 Jul;100(3):363–709.

24. Bjork RA. Memory and Meta-memory Considerations in the Training of Human Beings. In: Metcalfe J, Shimamura AP, editors. Metacognition: Knowing about Knowing. MIT Press; 1994. p. 185–205.

25. Bjork EL, Bjork RA. Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In: Psychology and the real world: Essays illustrating fundamental contributions to society. New York, NY, US: Worth Publishers; 2011. p. 56–64.

26. Gerlich M. AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies. 2025 Jan;15(1):6. DOI: http://dx.doi.org/10.3390/soc15010006
.

27. Carroll AJ, Borycz J. Integrating large language models and generative artificial intelligence tools into information literacy instruction. J Acad Librariansh. 2024 Jul 1;50(4):102899. DOI: http://dx.doi.org/10.1016/j.acalib.2024.102899
.

28. Pressley L. The Prior Knowledge Problem. Libraries as Epistemic Institutions [Internet]. 2026 Mar 9 [cited 2026 Mar 11]. Available from: https://laurenpressley.com/library/2026/03/09/the-prior-knowledge-problem/.

29. Grinschgl S, Papenmeier F, Meyerhoff HS. Consequences of cognitive offloading: Boosting performance but diminishing memory. Q J Exp Psychol. 2021 Sep 1;74(9):1477–96. DOI: http://dx.doi.org/10.1177/17470218211008060
.

30. Cowan N. What are the differences between long-term, short-term, and working memory? In: Progress in Brain Research [Internet]. Elsevier; 2008 [cited 2026 Apr 15]. p. 323–38. Available from: https://www.sciencedirect.com/science/chapter/bookseries/pii/S0079612307000209. DOI: 10.1016/S0079-6123(07)00020-9.

31. Halford GS, Cowan N, Andrews G. Separating cognitive capacity from knowledge: a new hypothesis. Trends Cogn Sci. 2007 Jun 1;11(6):236–42. DOI: http://dx.doi.org/10.1016/j.tics.2007.04.001

32. Strevens M. The knowledge machine: how irrationality created modern science. First edition. New York: Liveright Publishing Corporation; 2020. 350 p.

33. Naddaf M, Quill E. Hallucinated citations are polluting the scientific literature. What can be done? Nature. 2026 Apr 1;652(8108):26–9. DOI: http://dx.doi.org/10.1038/d41586-026-00969-z
.

34. Naddaf M. AI linked to explosion of low-quality biomedical research papers. Nature. 2025 May 21;641(8065):1080–1. DOI: http://dx.doi.org/10.1038/d41586-025-01592-0
.

35. Jones N. ArXiv preprint server clamps down on AI slop. Science. 2026 Jan 23. DOI: http://dx.doi.org/10.1126/science.z66g9k7
.

36. Stokel-Walker C. Scientists invented a fake disease. AI told people it was real. Nature. 2026 Apr 7;652(8110):559–61. DOI: http://dx.doi.org/10.1038/d41586-026-01100-y
.

37. Hao Q, Xu F, Li Y, Evans J. Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature. 2026 Jan;649(8099):1237–43. DOI: http://dx.doi.org/10.1038/s41586-025-09922-y
.

38. Hahnel M. What parts of the academic knowledge creation & dissemination pipeline can we automate with AI? OpenResearch.wtf [Internet]. 2026 Mar 23 [cited 2026 Apr 17]. Available from: https://www.openresearch.wtf/what-parts-of-the-academic-knowledge-creation-dissemination-pipeline-can-we-automate-with-ai/.

39. Shao E, Wang Y, Qian Y, Pan Z, Liu H, Wang D. SciSciGPT: advancing human–AI collaboration in the science of science. Nat Comput Sci. 2026 Mar;6(3):301–15. DOI: http://dx.doi.org/10.1038/s43588-025-00906-6
.

40. Cornelio C, Ito T, Cory-Wright R, Dash S, Horesh L. The need for verification in artificial intelligence-driven scientific discovery. Philos Trans R Soc Math Phys Eng Sci. 2026 Apr 9;384(2317):20240591. DOI: http://dx.doi.org/10.1098/rsta.2024.0591
.

41. Oatley K, Djikic M. Writing as Thinking. Rev Gen Psychol. 2008 Mar 1;12(1):9–27. DOI: http://dx.doi.org/10.1037/1089-2680.12.1.9
.

42. Carroll AJ, Borycz J. Generative Artificial Intelligence. In: Essentials of STEM librarianship. Chicago: ALA Neal-Schuman; 2026. p. 143–60.

43. Placani A. Anthropomorphism in AI: hype and fallacy. AI Ethics. 2024 Aug 1;4(3):691–8. DOI: http://dx.doi.org/10.1007/s43681-024-00419-4
.

44. Warner J. Writers Write [Substack newsletter]. The Biblioracle Recommends [Internet]. 2026 Apr 5 [cited 2026 Apr 21]. Available from: https://biblioracle.substack.com/p/writers-write.

45. Foster J. Language Models. Ingeniously Simple [Internet]. 2021 Oct 14 [cited 2023 Oct 6]. Available from: https://medium.com/ingeniouslysimple/language-models-15e45dce0805.

46. Shakespeare W. 3.3.97-98. In: Hamlet [Internet]. Washington, DC: Folger Shakespeare Library; [cited 2026 Apr 21]. Available from: https://www.folger.edu/explore/shakespeares-works/hamlet/read/3/3/.

47. Mollick E. Choosing to Stay Human [Internet]. 2026 Feb 18 [cited 2026 May 28]. Available from: https://www.oneusefulthing.org/p/choosing-to-stay-human.

48. Anderson H. BBC [Internet]. 2023 [cited 2026 May 28]. First Folio at 400: The Shakespeare words you don’t know you know. Available from: https://www.bbc.com/culture/article/20140527-say-what-shakespeares-words.

Alexander J. Carroll, AHIP, 1 alexander.j.carroll@vanderbilt.edu, Associate Editor, Journal of the Medical Library Association, Associate Director, Stevenson Science and Engineering Library, Vanderbilt University, Nashville, TN, United States

Jill T. Boruff, AHIP, 2 jill.boruff@mcgill.ca, Co-Lead Editor, Journal of the Medical Library Association, Associate Librarian, Schulich Library of Physical Sciences, Life Sciences, and Engineering, McGill University, Montreal, QC, Canada

Michelle Kraft, AHIP, FMLA, 3 kraftm@ccf.org, Co-Lead Editor, Journal of the Medical Library Association, Medical Library Director, Cleveland Clinic Foundation, Cleveland, OH, United States

© 2026 Alexander J. Carroll, Jill T. Boruff, Michelle Kraft

This work is licensed under a Creative Commons Attribution 4.0 International License.

Journal of the Medical Library Association, VOLUME 114, NUMBER 4, October 2026