Modern academia often rewards productivity, conformity, and short-term impact, sometimes at the expense of intellectual risk-taking. Within journal-based peer review systems, the evaluation of research is typically shaped by disciplinary norms, methodological expectations, and the perspectives of individual reviewers. This can create challenges for unconventional or paradigm-shifting work, particularly when it departs significantly from established frameworks or lacks easily recognizable benchmarks for quality. While this does not apply uniformly across all fields or reviewers, and some areas, especially in the Humanities, may be more receptive to innovative approaches, there remains a structural tendency in parts of the system to favor work that aligns with existing standards and literatures.

At the same time, an unprecedented volume of publications, driven in part by performance metrics, can limit opportunities for careful reflection, synthesis, and creative exploration. These pressures are often especially pronounced for early-career researchers, who may feel incentivized to prioritize safer, more conventional projects over riskier but potentially transformative ideas.

This article explores how artificial intelligence could help rebalance this system. Rather than replacing human judgment, AI can act as an assistive, pluralistic tool—lowering structural barriers to contribution, accelerating routine tasks, and opening space for deeper conceptual innovation. When thoughtfully designed, AI-assisted evaluation could identify high-risk, high-reward research, recognize structural novelty beyond citation metrics, and counteract entrenched disciplinary biases. The real promise of AI lies not in automating decisions, but in reshaping how scientific ideas are assessed, debated, and allowed to mature—potentially transforming peer review from a gatekeeping mechanism into a catalyst for creativity.

Structural Constraints in Contemporary Research Systems

Contemporary academic research environments tend to penalize intellectual risk-taking, particularly when new ideas question dominant frameworks or entrenched assumptions (Azoulai & Greenblatt, 2025). Evaluation mechanisms, especially peer review, are generally effective when research proceeds incrementally within accepted paradigms, but they often become restrictive when confronted with unconventional or transformative proposals (Liu & Liu, 2025). In such cases, editorial assessments may hinge on loosely defined criteria, including presumed audience interest, rather than on explicit methodological or empirical grounds, leading to the systematic exclusion of research with high uncertainty but potentially significant impact (Wang et al. 2017). At the same time, the scientific enterprise is characterized by an unprecedented expansion in output, with millions of papers produced annually, frequently motivated by quantitative performance indicators rather than substantive intellectual progress (Hanson et al. 2024). This emphasis on volume over depth constrains the time available for reflection, synthesis, and critical reasoning, disproportionately affecting early-career researchers who must navigate intense productivity pressures while still developing their scientific judgment (Liu et al. 2024).

AI as an Enabling Infrastructure for Research and Dissemination

In this context, artificial intelligence can serve as a corrective and enabling tool rather than a threat. AI systems can lower structural barriers to contribution by assisting researchers in data analysis, literature synthesis, hypothesis generation, and manuscript preparation, allowing scientists to focus on conceptual innovation rather than procedural overhead (Mabirizi et al. 2025). By accelerating routine tasks, AI creates intellectual space for deeper engagement with ideas and encourages exploration beyond mainstream research trajectories (Krenn et al. 2022; Webb et al. 2023). Moreover, AI-assisted tools can support alternative dissemination models by lowering practical barriers and improving coordination across research communities. For example, in preprint ecosystems, they can streamline manuscript preparation through language editing, formatting, and basic clarity checks, which facilitates earlier sharing of work without relying on journal-mediated processes; at the same time, they can enhance visibility by matching papers with relevant audiences based on content. In addition, within open peer commentary and post-publication review models, AI can help organize large volumes of feedback by clustering similar critiques, highlighting recurring concerns, and summarizing discussion threads, thereby making collective evaluation more accessible and usable. Furthermore, such tools can assist in identifying potential reviewers beyond traditional editorial networks by analyzing expertise and publication patterns, which may broaden participation and diversify perspectives; consequently, these combined functions can reduce dependence on centralized editorial gatekeeping while still supporting a degree of structure and quality control through AI-assisted filtering, prioritization, and synthesis of feedback. In doing so, AI has the potential not merely to increase the volume of scientific output in a quantitative sense, but to reshape what that volume represents by enabling a broader range of contributions to be produced and shared. Rather than reinforcing metric-driven proliferation, this expansion can support greater diversity, originality, and accessibility by lowering barriers to participation, facilitating the dissemination of niche or unconventional ideas, and enabling more iterative and open forms of scholarly exchange; in this sense, volume becomes meaningful insofar as it reflects a wider spectrum of voices and approaches rather than redundant or strategically optimized outputs. Consequently, AI may help foster a more pluralistic and innovation-friendly research ecosystem (Tennant et al. 2017; Fraser et al. 2021).

AI in Peer Review: Capabilities and Limitations

Beyond assisting researchers in their work, AI also holds potential to transform the peer-review and editorial process itself (e.g., Checco et al. 2021). By leveraging machine learning and natural language processing, AI could support the rapid screening of manuscripts for aspects such as methodological structure, statistical reporting, and logical coherence, as well as enable comparisons across large bodies of prior literature (see Kousha & Thelwall, 2024). At the same time, the use of such systems raises important risks that warrant careful consideration, including the possibility of embedded biases in training data, limitations in accurately interpreting complex or unconventional methodologies, and the risk of false positives or negatives in automated assessments. As a result, while AI may contribute to more consistent preliminary evaluations by avoiding fatigue or certain social influences, its outputs cannot be assumed to be neutral or error-free and therefore require critical scrutiny (Joaquim et al. 2025).

Bias, Innovation, and the Design of AI Evaluation Systems

In this context, AI would not replace human judgment but rather function as a first-line evaluator whose role is to assist rather than decide. For instance, it could flag submissions that meet baseline standards of clarity or highlight work that appears novel or interdisciplinary based on patterns in the literature, thereby helping editors and reviewers allocate attention more effectively. However, responsibility would remain with human experts to interpret these signals, assess the substantive quality and significance of the work, and account for nuances that AI systems may overlook, particularly in the case of innovative or paradigm-challenging research. This collaborative model positions AI as a tool for augmenting, rather than substituting, editorial judgment, while also underscoring the need for transparency, oversight, and reflexivity in how such tools are integrated into evaluation processes (Bai et al. 2025; Teixeira, 2025).

Bias, Novelty, and the Risk of Algorithmic Conservatism

However, as noted earlier, the promise of AI-based evaluation must be contextualized carefully, particularly with respect to its limitations and risks. While machines do not possess personal opinions or career interests, they are not neutral by default (Mavrogiorgos et al. 2024), and this caveat is important to foreground alongside their potential benefits. AI systems inherit the assumptions, values, and blind spots embedded in their training data and design; consequently, if they are trained predominantly on mainstream, high-impact journals, existing citation networks, or historically dominant paradigms, they may reproduce and even amplify the very conservatism that constrains innovation. This reinforces concerns raised earlier regarding biases and potential errors in AI-assisted evaluation, highlighting that such systems can misinterpret unconventional approaches or disproportionately favor familiar methodologies. In such cases, AI would not eliminate bias but rather automate and scale it, encoding prevailing norms into algorithmic form and reinforcing what is already widely recognized or institutionally validated, which underscores the need for critical oversight and careful integration throughout the evaluation process.

 AI, Creativity, and the Future of Scientific Evaluation

This concern becomes especially salient when considering creative and paradigm-shifting research. Scientific innovation often involves conceptual leaps, methodological hybrids, or unconventional framings that initially appear incomplete, speculative, or poorly aligned with established disciplinary standards. Historically, many transformative ideas were initially rejected or ignored precisely because they did not fit existing evaluative frameworks. An AI system optimized to recognize patterns, statistical regularities, and similarity to prior work may struggle to assess contributions that deliberately depart from these patterns. In other words, if novelty is defined too narrowly—such as deviation measured only within known conceptual neighborhoods—AI systems may systematically undervalue research that lies genuinely outside the norm.

Designing AI for Innovation and Pluralistic Evaluation

That said, AI also offers underexplored opportunities to expand, rather than restrict, the space of innovation, provided it is trained and deployed thoughtfully (e.g., Bianchini et al. 2022; Aldoseri et al. 2024). In practical terms, this depends on how such systems are designed, what data they are exposed to, and how their outputs are interpreted within research workflows (Wang et al. 2023). For instance, training AI models on more diverse and interdisciplinary corpora, including preprints, negative results, and work from underrepresented regions or fields, can broaden the range of patterns and ideas they are able to recognize and surface (Toto et al. 2025). Similarly, AI tools can be used to identify overlooked connections between distant areas of research, suggest unconventional combinations of methods or theories, and highlight emerging topics that have not yet been consolidated within dominant paradigms (Krenn et al. 2022). In evaluative contexts, they could be calibrated not only to assess alignment with established standards but also to detect novelty or divergence from typical patterns, thereby making space for less conventional contributions (Sourati & Evans, 2023). At the same time, these possibilities depend on maintaining transparency about how systems operate and ensuring that human researchers critically engage with, rather than defer to, AI-generated suggestions (Messeri & Crockett, 2024). Under such conditions, AI can function as a tool that helps widen the scope of inquiry and supports more exploratory and diverse forms of knowledge production, rather than narrowing them (Aldosari et al. 2024). Unlike human reviewers, AI systems can be explicitly designed to evaluate manuscripts along multiple, independent dimensions rather than collapsing judgment into a single accept/reject recommendation (Checco et al. 2021). For example, separate assessments could be generated for methodological soundness, conceptual originality, interdisciplinary integration, and long-term speculative potential. A manuscript that scores modestly on immediate empirical completeness but highly on conceptual novelty could be flagged as “high-risk, high-reward” rather than dismissed outright. Such multidimensional evaluation is difficult to implement consistently within traditional peer review but is well suited to algorithmic frameworks.

Training strategies will be decisive. To foster openness to unconventional ideas, AI models would need exposure not only to successful, mainstream publications but also to historically important papers that were initially controversial, interdisciplinary, or poorly received (Garg et al. 2018). Training datasets could deliberately include rejected manuscripts later recognized as influential, early-stage theoretical proposals, and research emerging from peripheral fields or non-dominant institutions. Moreover, AI systems could be trained to recognize structural novelty—defined as new combinations of concepts, methods, or datasets—rather than relying solely on semantic similarity or citation-based indicators, which are known to undervalue genuinely innovative research ((Uzzi et al. 2013; Foster et al., 2015). Achieving this would require moving beyond surface-level text similarity toward deeper representations of scientific reasoning that encode conceptual and relational structure (Bzdok et al., 2018). At the same time, such an approach raises important practical considerations related to copyright and intellectual property, particularly regarding the data used to train and evaluate these systems. Developing models capable of capturing underlying scientific reasoning would likely require access to a large and diverse corpora of full-text publications, including potentially restricted or proprietary materials, which introduces questions about permissions, licensing, and fair use. Clear frameworks for data governance, including agreements between publishers, institutions, and researchers, would therefore be necessary to ensure that training practices are both legally compliant and ethically justified. In addition, transparency around how texts are used, represented, and transformed within AI systems would be important for maintaining trust, especially in contexts where authors may be concerned about appropriation, attribution, or the reuse of their intellectual contributions. Addressing these issues is a necessary condition for developing more conceptually sophisticated AI systems that can engage meaningfully with scientific knowledge without undermining the rights and incentives of its producers.

Equally important is the role of humans in the loop. AI should not function as an autonomous gatekeeper but as an assistive and exploratory tool that augments editorial judgment. Editors and reviewers could use AI-generated assessments to identify overlooked strengths, challenge their own intuitions, or detect implicit biases in their evaluations (Amershi et al. 2017). In this sense, AI could act as a counterweight to entrenched disciplinary norms, rather than a replacement for human discernment (Green & Chen, 2019). Transparency in model design, training data, and decision logic would be essential to ensure accountability and trust (Kahneman et al. 2011).

Ultimately, the question is not whether AI can judge creativity in the same way humans do—it cannot—but whether it can help create evaluation environments that are more pluralistic, reflective, and self-aware than current systems allow. If designed conservatively, AI will simply mirror existing practices and reinforce intellectual inertia. If designed conservatively, AI will tend to mirror existing practices and reinforce intellectual inertia. Conversely, if developed with explicit attention to uncertainty, diversity, and long-term scientific value, it could contribute to shifting evaluation away from narrow, short-term performance metrics and toward a broader understanding of contribution. However, this distinction depends critically on how “experimental” design is defined and governed. Not all claims of experimentation are necessarily aligned with careful or responsible innovation, and without appropriate oversight, such approaches may reflect a willingness to deploy insufficiently tested systems rather than a genuine commitment to improving evaluative practices. For this reason, any effort to use AI in more exploratory ways would need to be accompanied by clear standards for validation, transparency, and accountability, ensuring that experimentation enhances, rather than undermines, the reliability and integrity of scientific assessment. In that sense, the creative potential of AI in science lies less in its capacity to generate ideas, and more in its ability to reshape the structures through which ideas are recognized, debated, and allowed to mature.