1. Introduction
The diffusion of scientific knowledge hinges on academic endorsement. Researchers review abstracts or manuscripts of their peers to determine suitability for publication in scientific journals or presentations at scientific conferences. Reviewers may also provide feedback on the elements that could be improved, ultimately, validating the rigor of the work conducted. However, as scientific production continues to expand, the quality and efficiency of the peer review system suffer. Moreover, with the rapid advancements in technology, the peer review system faces new challenges. One disruptive development is the integration of artificial intelligence (AI) in scientific writing, which introduces additional layers of complexity in reviewing and validating research. The complexity of the scientific work coupled with the mass production of text generated by AI challenges the reviewers’ ability to offer constructive feedback (Liang et al., Reference Liang, Zhang, Cao, Wang, Ding, Yang, Vodrahalli, He, Smith, Yin, McFarland and Zou2024) and delays the publication process (Dwivedi et al., Reference Dwivedi, Kshetri, Hughes, Slade, Jeyaraj, Kar, Baabdullah, Koohang, Raghavan, Ahuja, Albanna, Albashrawi, Al-Busaidi, Balakrishnan, Barlette, Basu, Bose, Brooks, Buhalis and Wright2023).
In light of these challenges, recent advancements in large language models (LLMs) prompted a discussion about whether they could support the academic review process, making it more efficient and fairer (e.g., Dwivedi et al., Reference Dwivedi, Kshetri, Hughes, Slade, Jeyaraj, Kar, Baabdullah, Koohang, Raghavan, Ahuja, Albanna, Albashrawi, Al-Busaidi, Balakrishnan, Barlette, Basu, Bose, Brooks, Buhalis and Wright2023; Guo et al., Reference Guo, Zhang, Wang, Jiang, Nie, Ding, Yue and Wu2023; Hosseini and Horbach, Reference Hosseini and Horbach2023; Liang et al., Reference Liang, Zhang, Cao, Wang, Ding, Yang, Vodrahalli, He, Smith, Yin, McFarland and Zou2024; Liu and Shah, Reference Liu and Shah2023; Schulz et al., Reference Schulz, Barnett, Bernard, Brown, Byrne, Eckmann, Gazda, Kilicoglu, Prager, Salholz-Hillel, ter Riet, Vines, Vorland, Zhuang, Bandrowski and Weissgerber2022). Although some argue that AI could provide valuable assistance during the review and editorial process (e.g., for copy editing), they are similarly skeptical about its ability to engage in an in-depth evaluation of academic writings (e.g., Lindsay, Reference Lindsay2023). However, we already see some early evidence of reviewers resorting to AI assistance in evaluating conference submissions. Latona et al. (Reference Latona, Ribeiro, Davidson, Veselovsky and West2024) analyzed the reviews submitted for the International Conference on Learning Representations (ICLR) and found that at least 15% of them were fully or partially AI-assisted, thereby leading to roughly half of the submissions receiving at least 1 AI-assisted review. Although LLMs are already integrated into the peer review process, we currently lack empirical evidence of their proficiency as academic reviewers.
We aim to contribute to this discussion by systematically evaluating the ability of LLMs’ to differentiate between AI- and human-written conference abstracts and comparing the evaluations of all abstracts given by human reviewers to those given by AI. To this end, we conducted a large-scale field experiment, relying on the abstracts and human peer reviewers of a medium-sized scientific conference. Our research helps to elucidate to what extent easily accessible LLMs could support the scientific review process.
1.1. Related literature
Researchers from different fields have highlighted that the current peer review process is cumbersome and, often, not objective. Scientific output has become increasingly more complex, including extensive and sometimes very technical supplementary materials, which places the burden of scrutiny on the typically non-incentivized reviewers (Petrescu and Krishen, Reference Petrescu and Krishen2022). As such, writing a peer review typically takes 4–8 h (Huisman and Smits, Reference Huisman and Smits2017; Kovanis et al., Reference Kovanis, Porcher, Ravaud and Trinquart2016); the associated overall costs for peer-reviewing are estimated to sum up to approximately 1.5 billion USD per year (Aczel et al., Reference Aczel, Szaszi and Holcombe2021). The mass production of AI-generated text could tax the peer review system even further and undermine the credibility of scientific research (Dwivedi et al., Reference Dwivedi, Kshetri, Hughes, Slade, Jeyaraj, Kar, Baabdullah, Koohang, Raghavan, Ahuja, Albanna, Albashrawi, Al-Busaidi, Balakrishnan, Barlette, Basu, Bose, Brooks, Buhalis and Wright2023; Sabel et al., Reference Sabel, Knaack, Gigerenzer and Bilc2023).
The increasing burden of scientific reviewing has prompted scientists to explore ways in which AI can be used as a prescreening tool to assess the general journal fit, compliance with submission and reporting guidelines, risk of plagiarism, conflict of interest, and inappropriate content (Dwivedi et al., Reference Dwivedi, Kshetri, Hughes, Slade, Jeyaraj, Kar, Baabdullah, Koohang, Raghavan, Ahuja, Albanna, Albashrawi, Al-Busaidi, Balakrishnan, Barlette, Basu, Bose, Brooks, Buhalis and Wright2023; Foltýnek et al., Reference Foltýnek, Meuschke and Gipp2019). But can AI tools, such as GPT-4, tell the difference between human-authored research output and AI or evaluate the quality of scientific research as reported in abstracts? The recent body of literature suggests that humans are generally capable of differentiating human-written from AI-generated scientific text (e.g., Gao et al., Reference Gao, Howard, Markov, Dyer, Ramesh, Luo and Pearson2023; Ma et al., Reference Ma, Liu, Yi, Cheng, Huang, Lu and Liu2023). However, LLM systems like GPT-4 are not far behind (Gao et al., Reference Gao, Howard, Markov, Dyer, Ramesh, Luo and Pearson2023). When it comes to evaluating the quality of scientific output, the peer review system aims to impartially and constructively evaluate the robustness and reliability of the scientific work submitted to a conference or journal. However, recent studies suggest that the review system has not been immune to gender, racial, seniority, and demographic biases (e.g., D’Andrea and O’Dwyer, Reference D’Andrea and O’Dwyer2017; Haffar et al., Reference Haffar, Bazerbachi and Murad2019; Laycock and Bailey, Reference Laycock and Bailey2019; Lee et al., Reference Lee, Sugimoto, Zhang and Cronin2013; Murray et al., Reference Murray, Siler, Larivière, Chan, Collings, Raymond and Sugimoto2019; Smith et al., Reference Smith, Davis, Pizza, Waterman, Dobson, Foster, Jarvey, Jones, Leuenberger, Nourn, Conway, Fiser, Hansen, Hristova, Mack, Saunders, Utley, Young and Davis2023; Wicherts, Reference Wicherts2016). While LLMs can potentially help in reducing reviewer bias and presenting feedback in a neutral tone, we also need to ascertain that its evaluation can distinguish between different levels of scientific quality. There is encouraging evidence that LLMs are quite accurate in spotting errors and assessing papers using a set of checklist criteria. According to Liu and Shah (Reference Liu and Shah2023), using AI help has the potential to mitigate or eliminate established biases in the review process, such as assessments influenced by author identifiers, null findings, or certain buzzwords. However, it remains unclear how aligned human and AI reviewers are when it comes to evaluating the overall quality of presented scientific work and whether this alignment is comparable to the existing peer review system (e.g., the agreement between different reviewers when evaluating the same scientific work).
To add more empirical evidence to this discussion, our field experiment seeks to make 2 contributions. First, we investigate the ability of human and AI reviewers to accurately detect the source of authorship, that is, whether a scientific conference abstract was written by a human author, reporting on actual research, or whether it was written by ChatGPT, reporting on a made-up study. Second, we compare the quality assessments of abstracts by human and AI reviewers.
Based on the literature summarized above, we preregistered the following hypotheses:Footnote 1
-
• H1. Reviewers will rate fabricated research reported in AI-generated scientific abstracts as more likely to be generated by AI than the authentic research reported in human-generated abstracts by researchers.
-
• H2. The AI reviewer will have a higher accuracy rate in identifying AI-generated scientific abstracts than human reviewers.
-
• H3. The research quality as reported in AI-generated scientific abstracts is rated lower than the research quality as reported in original submissions by researchers.
-
• H4. Human reviewers will evaluate the research as reported in abstracts written by researchers as having higher quality compared to the research as reported in abstracts generated by AI, when compared to the evaluations made by the AI reviewer.
2. Method
Participants were not deceived in this study. The study has been approved by the Departmental Review Board of Occupational, Economic, and Social Psychology (DRB number 2023/W/004). This study was preregistered (https://osf.io/zaj6p). Additional analyses are flagged as exploratory. The ethics approval, study materials, data, and analysis code are publicly accessible online (https://osf.io/nje23).
2.1. Experimental design and participants
This study took place as part of the abstract submission and evaluation of the 29th Biennial Subjective Probability Utility and Decision Making conference (SPUDM) in 2023, hosted at the Vienna University of Economics and Business. We employed a
$2\times 2$
mixed factorial design with the author type as the within-participants factor (human vs. AI author) and the reviewer type as the between-participants quasi-experimental factor (human vs. AI reviewer).
Regarding author type, the conference received 429 abstract submissions by 404 unique corresponding authors, of which 305 authors agreed to participate in the experiment, that is, having their abstract evaluated by AI. Participants could choose 1 subfield to which their abstract relates most (see below). In addition, there were 20 AI-generated abstracts, 1 in each subfield. The final abstract sample includes 245 abstracts (20 AI-generated), with reviews conducted by both humans and AI.Footnote 2
Regarding reviewer type, of the 404 submitting authors, 289 agreed to review other abstracts, and 217 scientists agreed to participate in the experiment by additionally reviewing AI-generated abstracts (119 female, 122 male, 1 diverse, 4 prefer to not say; 92 hold a PhD). Each human-generated abstract was evaluated by 1–5 (
$M = 2.10, SD = 0.92$
) human reviewers; each AI-generated abstract was evaluated by 1–35 (
$M = 11.47, SD = 9.44$
) human reviewers.
2.2. Experimental factors
Abstract type. The submission guidelines for human authors were as follows: Participants were instructed to submit abstracts, limited to a maximum of 700 words, for the review process. The emphasis was placed on including details regarding hypotheses, methods, and analyses. Additionally, participants were requested to specify their preferred presentation format (warm-up online talk, flash talk, or extended talk) and select the most applicable subfield for their abstract.Footnote 3 They were informed that the abstracts would undergo a double-blind review by the JDM community.
For AI-generated abstracts, ChatGPT (based on GPT-3.5 in February 2023) was fed with the following prompt: ‘Create a paper title and scientific abstract in the subfield of “X” in the broader research area on judgment and decision-making. Use past tense. The abstract should have a maximum of 700 words and include details on the hypotheses, methods, and analyses.’ X was filled with the respective subfields (see Footnote Footnote 3). Accordingly, ChatGPT generated 20 abstracts, 1 for each subfield.Footnote 4
Reviewer type. In the human reviewer condition, all submitting authors were asked to evaluate other abstract submissions and submit their reviews within 2 weeks via an online survey conducted with Qualtrics (https://www.qualtrics.com/). They could decide to take part in the experiment by reviewing 4 instead of 3 abstracts, of which one would have been generated by AI (without knowing which one). They only received abstracts that were within their stated area of expertise based on the 20 subfields.
In the AI reviewer condition, ChatGPT (based on GPT-4 in June 2023) was queried with the following prompt prior to posing the same evaluative questions that human reviewers faced (see below): ‘Act as a reviewer for a conference abstract. The conference represents an international and interdisciplinary forum for scientists dealing with modeling, analyzing, and aiding decision processes. It covers fundamental as well as applied research, attracting contributions from various disciplines such as psychology, economics, medicine, law, management science, philosophy, and computer science. You will be provided with an abstract to review, and you will rate it according to the following 6 criteria. Your review should be in the given format, with no need for additional explanation:’
Measures. Human and AI reviewers were asked to answer the following questions: (1) ‘What is the significance or importance of this research for the field of judgment and decision-making?’, response scale 0–10, from ‘not important at all’ to ‘very important’; (2) ‘What is your evaluation of the quality of the presented research (e.g., please consider the rigor and appropriateness of design, methods, data analysis, etc.)’, response scale 0–10, from ‘very poor quality’ to ‘very high quality’; (3) ‘What is your assessment of the robustness of the findings presented?’, response scale 0–10, from ‘very unlikely to replicate’ to ‘very likely to replicate’ and additional ‘not applicable’ response option; (4) ‘Would the presented research benefit more from focused, in-depth discussions or from broader, conceptual comments?’, response scale 0–10, from ‘detailed discussion’ to ‘broader conceptual comments’; and (5) ‘How likely is it that this abstract was created by an AI-model?’, response scale 0–10, from ‘very unlikely’ to ‘very likely’.
3. Results
3.1. Confirmatory analyses
As preregistered, we employed mixed effects regressions to predict (i) the probability of an abstract being identified as AI-generated and (ii) the perceived quality of the research. This analysis utilized 2 dummy variables: the first representing the author type (human vs. AI author) and the second representing the reviewer type (human vs. AI reviewer). The subfield of the research topic, with 22 distinct categories, as well as the reviewer (given that human reviewers evaluated several abstracts) were treated as random effects to account for the potentially interrelated error variance.
Figure 1 Research likelihood to be generated by AI (panel A) and research quality (panel B) of abstracts written by human or AI authors as evaluated by human or AI reviewers. Barplots represent mean scores in each group, and error bars represent 95% confidence intervals. Dots represent individual observations, and the size of the dots illustrates the number of observations.
Table 1 Regression models predicting the perceived likelihood of an abstract to be AI-generated (Models 1 and 2) and the perceived quality of an abstract (Models 3 and 4) using human and ChatGPT ratings
Note: We excluded the random effect of the research topic from Models 1 and 2, as they did not contribute to explaining any variance in the outcome.
*
$p < 0.05,$
**
$p < 0.01,$
***
$p < 0.001.$
As shown in Figure 1A, the mean rating for human-generated abstracts to be generated by AI on a scale from 0 (‘very unlikely’) to 10 (‘very likely’) was
$M = 2.73$
(
$Md = 2, SD = 2.19$
), which was significantly lower when compared to AI-generated abstracts (
$M = 6.69, Md = 8, SD = 2.92$
), Mann–Whitney U test:
$U = 26,216, z = 15.96, p < 0.001$
. However, as indicated by the significant interaction effect of author type and reviewer type in Models 1 and 2 (see Table 1), this difference is entirely driven by human reviewers (simple slope: unstandardized
$B = 4.15, SE = 0.67, p < 0.001$
), supporting H1. In contrast, the AI reviewer does not discriminate between human- and AI-generated abstracts (simple slope:
$B = 0.28, SE = 0.49, p = 0.563$
). This rejects H2 that the AI reviewer would be more likely to identify AI-generated abstracts as AI-generated than human reviewers; actually, the data suggest that the opposite is true. Regarding the quality of the research, rated from 0 (‘very poor quality’) to 10 (‘very high quality’), research made-up and reported by AI (
$M = 4.01, Md = 4, SD = 2.42$
) was rated as significantly lower in quality when compared to actual research reported in human-written abstracts (
$M = 6.88, Md = 7, SD \,{=}\, 1.82$
), Mann–Whitney U test:
$U = 29,865, z = 15.00, p < 0.001$
. Although this finding is in support of H3, the mixed effects regression analyses reported in Models 3 and 4 (Table 1) again show that this difference is qualified by a significant interaction effect between author type and reviewer type (for a visualization, see Figure 1B). Specifically, human reviewers evaluated the research reported in human-written abstracts more positively than the research reported in AI-generated abstracts (simple slope:
$B = 2.79, SE = 0.13, p < 0.001$
). In contrast, the AI reviewer did not discriminate in the research quality reported in human- versus AI-generated abstracts (simple slope:
$B = 0.12, SE = 0.38, p = 0.752$
). This finding is in support of H4.
3.2. Exploratory analyses
GPTZero. The results indicate that ChatGPT is poor at discriminating between human- and AI-generated abstracts. However, there are commercially available LLM detector tools, such as GPTZero (Tian and Cui, Reference Tian and Cui2023), that have been designed to identify the extent to which texts were written with the assistance of LLM. To explore whether GPTZero is indeed better at detecting AI-generated abstracts, each abstract was fed into GPTZero one by one, and each received a score between 0 and 100, indicating the likelihood that the text is entirely AI-generated.
To test H2 using GPTZero as the AI reviewer, we conducted the same mixed effects regressions with the perceived AI likelihood as the outcome as reported above. Prior to the analysis, we rescaled the GPTZero score from 0–100 to 0–10, to make it comparable to the ratings by human reviewers. As shown in Table 2, we find a main effect of author type, indicating that both human and AI reviewers were able to accurately distinguish between human and AI-generated abstracts. In fact, the significant interaction between author type and reviewer type suggests that GPTZero was even better than human reviewers in discerning authorship (see Figure 1A). Hence, when using GPTZero instead of ChatGPT, we find support for H2.
Table 2 Regression models predicting the perceived likelihood of an abstract to be AI-generated using human and GPTZero ratings
Note:
*
$p < 0.05,$
**
$p < 0.01,$
***
$p < 0.001.$
Rating agreement. Human reviewers often show variability in their assessments, indicating a lack of agreement in the evaluation of scientific work. Exploring the agreement among human reviewers when evaluating the same abstract, we found the quality scores showed a medium-sized positive correlation (
$r = 0.38, 95\% CI [0.26, 0.48], t(229) = 6.14, p < 0.001$
).Footnote 5 Similarly, the AI likelihood scores from 2 human reviewers exhibited a medium-sized positive correlation (
$r = 0.33, 95\% CI [0.21, 0.44], t(229) \,{=}\, 5.23, p < 0.001$
). Additionally, using Cohen’s Kappa for human reviewer scores, we found a 20.8% agreement in quality assessments (N Abstracts = 231) and a 16% agreement in AI likelihood assessments. So, indeed, agreement between human reviewers was rather low.
Investigating the association between AI and human reviewers’ assessments, we observed a positive small-to-medium correlation in the quality scores assigned by both AI and human reviewers (
$r \,{=}\, 0.23, 95\% CI [0.11, 0.35], t(243) = 3.68, p < 0.001$
). The AI likelihood scores from AI and 1 randomly selected human reviewer exhibited a positive but very small and statistically insignificant correlation (
$r = 0.08, 95\% CI [-0.04, 0.21], t(243) = 1.29, p = 0.200$
), which corresponds to the findings reported in Table 1 (Models 1 and 2). Furthermore, utilizing Cohen’s Kappa for AI reviewer and 1 randomly selected human reviewer, we identified a 22% agreement in quality assessments and a 17.6% agreement in AI likelihood assessments for the same set of abstracts and rater. Notably, these figures are very similar to the agreement between the 2 human reviewers reported above. Taken together, the results indicate that reviewer agreement is low but comparable between human–human and human–AI reviewer pairs.
To further explore whether LLMs may be useful for screening conference abstracts, and if so, whether there is agreement with human reviewers in the classification of very good or very bad abstracts, we utilized Bland–Altman plots (Bland and Altman, 1986) to visually assess the agreement between human and AI reviewers for different quality levels. Specifically, these plots depict the differences between the 2 methods of measurement in comparison to their average evaluation. Figure 2 illustrates that the AI quality ratings are biased, and the mean human-perceived quality is lower than AI ratings, see panels A and B (AI vs. human reviewers) in comparison to panel C (only human reviewers). Nevertheless, as the differences between human and AI quality ratings become smaller with larger (average) quality ratings in panels A and B, current LLMs may be more reliable in prescreening high-quality research abstracts than low-quality research abstracts.
Figure 2 Bland–Altman plots of agreement in research quality ratings between AI and the mean rating of human and AI reviewers (A), between AI and the mean rating of several human reviewers (B), and between 2 randomly paired human reviewers (C).
4. Discussion
Our study contributes to the growing discussion on the utilization of LLMs for improving the taxed academic review process (e.g., Liang et al., Reference Liang, Zhang, Cao, Wang, Ding, Yang, Vodrahalli, He, Smith, Yin, McFarland and Zou2024; Liu and Shah, Reference Liu and Shah2023). We focused on 2 main aspects of the review process: authorship detection and quality evaluation.
First, our results corroborate previous findings on probabilistic detection of AI-generated abstracts, suggesting that researchers are good, and in our case even better than ChatGPT, at differentiating human- versus AI-generated content. However, our exploratory analysis revealed that AI classifier tools like GPTZero can accurately detect AI-generated content, even exceeding human reviewers’ discriminatory ability.
Second, we find that human reviewers evaluate the real abstracts by human authors more positively than the made-up abstracts by AI authors, whereas the AI reviewer failed to differentiate between them. While the evaluations of human and AI reviewers are positively but only weakly related, the same is true for the evaluations of different human reviewers. ChatGPT offers inflated quality evaluations of both original and generated work, failing to distinguish high- and low-quality output. The findings further suggest that human and AI reviewers are more aligned when it comes to evaluating the very best abstracts than the low-quality abstracts.
4.1. Implications
The results suggest that at this point LLMs cannot substitute human reviewers in the peer review process. Nevertheless, there might be opportunities for human–AI collaboration. For instance, there may be an increasing need to differentiate whether certain texts were generated by humans or AI. Our results suggest that AI-generated scientific abstracts can well be identified by GPTZero. Such specialized software may serve journals or editors as a screening device, but it is important to bear in mind that its accuracy is highly dependent on text length and training data (OpenAI et al., Reference Achiam, Adler, Agarwal, Ahmad, Akkaya, Aleman, Almeida, Altenschmidt, Altman, Anadkat, Avila, Babuschkin, Balaji, Balcom, Baltescu, Bao, Bavarian, Belgum and Zoph2023). While it is true that the review process is time-consuming and that some cases take more time to review than others, utilizing LLMs for preprocessing and initial screening can provide valuable insights across all types of submissions. Although it is unlikely for an extended paper to be purely AI-generated, such scenarios are becoming more common. Many conferences only require an abstract for submission, and the sheer volume of submissions makes high-quality and fast reviews more challenging. Utilizing LLMs to reduce the burden of filtering and preprocessing submissions can significantly improve the overall review quality by freeing up more time and resources for reviewers to focus on in-depth evaluations for the remaining submissions. Future research should investigate the detection accuracy in the case of full-length scientific papers, with only parts being written by LLMs and in different scientific disciplines.
In the context of quality evaluations, human researchers were much better at indicating gradation of content quality and clearly ranked AI-generated content as much lower quality than human-written text. As human reviewers knew that 1 abstract was AI-generated, it could be that such lower quality ratings result from the human reviewers’ belief that those abstracts were generated by AI, reflecting a general competence bias against AI-generated content (e.g., Böhm et al., Reference Böhm, Jörling, Reiter and Fuchs2023). Our results should be interpreted with caution when generalizing to settings in which reviewers are not aware that content could be generated by AI (which may be less affected by AI aversion).
Moreover, we confirm the tendency of AI-assisted evaluations to present inflated quality ratings for both human-generated and AI-assisted abstracts, significantly favoring its own output. In the work of Latona et al. (Reference Latona, Ribeiro, Davidson, Veselovsky and West2024), this had a positive effect on the paper acceptance rate, whereby a submission that received an AI-assisted review was 4.9 percentage points more likely to be accepted. Combined, our results stress the importance of examining the impact that LLMs have on the review process and introducing guidelines for the integration of LLMs.
Importantly, in the present study, the AI reviewers had good agreement with human reviewers when it came to identifying high-quality research abstracts. As such, AI could support the review process as a prescreening tool, before human reviewers engage in more fine-grained evaluation (e.g., for selecting conference presentations). Agreement between human and AI reviewers could be further increased when providing more specific and less subjective evaluation criteria (e.g., Liu and Shah, Reference Liu and Shah2023). In the present study, reviewers were simply asked to evaluate the ‘quality’ of the presented research, which is quite broad and subjective.
AI-based support in scientific reviewing has the potential to reduce or even eliminate well-known biases in the reviewing system, such as evaluations based on author identifiers, null results, or certain buzzwords (Liu and Shah, Reference Liu and Shah2023). However, this potential should be viewed cautiously and critically, given the documented instances of AI models inheriting and even amplifying biases from their training data (Bernhardt et al., Reference Bernhardt, Jones and Glocker2022; Hartmann et al., Reference Hartmann, Schwenzow and Witte2023; Scherrer et al., Reference Scherrer, Shi, Feder and Blei2024). Future research should investigate under what conditions and with what prompts about evaluation criteria such biased reviews are likely to disappear when evaluated by AI reviewers.
4.2. Limitations
There are several limitations to our study. Our sample of participants suffers from self-selection bias since the results are based on responses from conference participants who consented to review AI-generated work.Footnote 6 While the call for participation was open to all conference participants, it is quite likely that participants who opted into the study were already familiar with and interested in LLMs. Such reviewers may have been particularly knowledgeable in distinguishing between actual human-generated abstracts and made-up AI-generated abstracts. Furthermore, we only compared purely AI-generated and human-written abstracts. Our primary intention was to isolate the effects of LLMs in the abstract generation and review processes, thereby limiting the potential impact of human intervention. Researchers commonly employ LLMs to improve many parts of their writing, such as language refinement, clarity, and idea development (Liang et al., 2024a,b). We did not have data on the extent or method by which researchers used LLMs to improve their submissions. Therefore, our study cannot account for the complex ways AI may have contributed to the quality of human-written abstracts. However, we did evaluate the extent to which authors have potentially used ChatGPT using GPTZero, and the overall prediction was rather low (8.9%, see Model 2 in Table 2). Moreover, we have used the same prompt for both human reviewers and ChatGPT for evaluation. As described above, some prompts were quite abstract and lacked specific evaluation criteria, which could have been an issue for ChatGPT evaluations. A precise and unambiguous prompt may better demonstrate the limits and capabilities of LLMs, and some researchers (e.g., Liang et al., Reference Liang, Zhang, Cao, Wang, Ding, Yang, Vodrahalli, He, Smith, Yin, McFarland and Zou2024; Liu and Shah, Reference Liu and Shah2023) have either given very specific criteria or dynamically changed the prompt to probe for the most accurate answer. Finally, our results are naturally limited to the current technology (i.e., GPT-3.5 and GPT-4 as used in the present study) and knowledge of the sample of reviewers.Footnote 7 Improved LLMs, some of which are fine-tuned on scientific data (e.g., SciBERT Beltagy et al., Reference Beltagy, Lo and Cohan2019), may soon overcome some of the weaknesses of AI-based reviews.
5. Conclusion
LLMs will undoubtedly impact the scientific peer review process, and our results suggest that the impact is likely to be positive and useful for the scientific reviewing system. While we should not expect to outsource scientific review to LLMs such as ChatGPT in the nearest future, we already know which tasks can be delegated. Prescreening papers can identify and elevate highly robust scientific work, bypassing known reviewer biases. In sum, at its current state, LLMs have a lot to offer to the scientific review process, both in terms of publishing efficiency and diversity of the author pool.
Data availability statement
The ethics approval, study materials, data, and analysis code are publicly accessible in online (https://osf.io/nje23).
Author contributions
A.S. and H.H. have contributed equally to this work.
Competing interests
The authors declare no competing interests in this work.
Open access
