Generative Metascience: A Review of AI as the Next Scientific Instrument and the Emerging Paradigm of Algorithmic Discovery

1. Introduction: Establishing the Context and Rationale

1.1. The Broader Landscape: Background and Significance

In 2020, DeepMind’s AlphaFold achieved a groundbreaking milestone by solving the protein folding problem with remarkable accuracy, a challenge that had persisted for over five decades (Jumper et al., 2021). This success, powered by advanced machine learning, exemplifies AI’s transformative potential in scientific discovery. Beyond protein folding, AI has made significant contributions across various scientific domains. In astronomy, machine learning models have been instrumental in detecting exoplanets, with projects like ExoMiner using deep learning to validate 301 exoplanets from Kepler mission data (Valizadegan et al., 2022). In drug discovery, AI has accelerated the identification of new therapeutic compounds; for instance, Exscientia has utilized AI to design molecules that have entered clinical trials, significantly reducing the traditional timeline for drug development (Philippidis, 2023). These examples underscore AI’s versatility and its capacity to address complex scientific challenges.

The volume of scientific literature is growing exponentially, with recent studies indicating a doubling time of approximately 17.3 years (Bornmann et al., 2021). This rapid expansion underscores the critical role of metascience in enhancing the efficiency and reliability of research practices, ensuring that scientific inquiry remains robust and impactful amidst mounting complexity.

Artificial intelligence, with its ability to process vast datasets and discern intricate patterns, is becoming a pivotal force in metascience. AI can automate mundane tasks like data cleaning, generate novel hypotheses via pattern recognition and predictive modeling, optimize experimental designs, and interpret results by highlighting significant trends and anomalies (Jordan & Mitchell, 2015a). These functionalities not only expedite scientific discovery but also challenge traditional paradigms by offering innovative methods for hypothesis generation and validation. A notable example is in drug discovery, where AI has facilitated the creation of new therapeutic compounds; for instance, Insilico Medicine has used AI to design drugs that have progressed to clinical trials, markedly reducing development timelines and costs (Zhavoronkov et al., 2019).

This review introduces Generative Metascience, a framework that emphasizes AI’s, particularly generative models’, capacity to not only analyze existing data but also to autonomously generate new hypotheses and propel scientific discovery. This approach underscores AI’s dual function as both an analytical instrument and an independent co-investigator in research. By synthesizing AI’s applications across various scientific fields, from genomics to astronomy, we seek to elucidate how AI is fundamentally altering the terrain of scientific inquiry.

1.2. The Core Problem: Identifying the Critical Gap or Controversy

Despite AI’s widespread adoption in scientific research, there is a pressing need for a comprehensive review that assesses its overarching role as a scientific instrument. Existing literature tends to concentrate on specific applications, such as AI in drug discovery or AI in astronomy, offering detailed insights into particular domains but neglecting the broader, interdisciplinary implications of AI (Ball & Brunner, 2010; Zhavoronkov et al., 2019). A holistic analysis that examines AI’s influence on the entire scientific process, encompassing methodological shifts, ethical dilemmas, and the practical challenges of incorporating AI into conventional research practices, is conspicuously absent. This deficiency impedes our capacity to fully harness AI’s potential while addressing critical risks, including data biases and the interpretability of AI models (Jobin et al., 2019).

1.3. Delineating the Review’s Focus and Approach

Generative Metascience refers to an AI-driven framework in which artificial intelligence not only analyzes existing scientific data but also autonomously formulates, prioritizes, and tests novel hypotheses, effectively acting as both a research instrument and co-investigator in the scientific discovery process. This review aims to fill this critical gap by offering a comprehensive analysis of AI’s multifaceted applications in scientific research, evaluating its efficacy, and delving into the nascent paradigm of algorithmic discovery. Uniquely, this review integrates AI’s functions in both data collection and analysis and in hypothesis generation and testing, presenting a unified framework that encompasses AI’s complete influence on the scientific endeavor. By focusing on this duality, we elucidate the dynamic between data-centric analytics and hypothesis-oriented research, illustrating how AI can concurrently formulate and appraise scientific hypotheses. Drawing from a synthesis of findings across various disciplines, we identify exemplary practices for incorporating AI into research methodologies and pinpoint areas necessitating further scholarly attention. Additionally, this review will address the policy implications of AI’s integration into science, including recommendations for funding initiatives and educational reforms to support the development of AI-driven research ecosystems. This balanced synthesis is intended to inform and guide researchers, policymakers, and funding bodies in navigating the complexities of AI integration in science.

1.4. A Roadmap for the Reader: Structure of the Article

This review is structured as follows: Section 2, “Thematic Synthesis and Critical Analysis of the Literature,” encompasses five subsections that explore AI’s role in scientific discovery. Subsection 2.1, “Historical and Conceptual Foundations,” provides a historical overview of AI’s development in science and clarifies key terminology. Subsection 2.2, “The Methodological Canvas: Approaches to Research in the Field,” examines the diverse methodologies used to study AI in scientific contexts, including their strengths, weaknesses, and emerging innovations. Subsections 2.3 and 2.4 delve into the two principal themes: “Thematic Deep Dive 1” focuses on AI’s contributions to data collection and analysis, while “Thematic Deep Dive 2” addresses AI’s role in hypothesis generation and testing. Subsection 2.5, “Cross-Thematic Analysis: Interconnections and Contrasting Perspectives,” synthesizes the interactions between these themes and their implications for the paradigm of algorithmic discovery, Section 3, “Discussion, Implications, and Future Trajectories,” outlines the theoretical and practical implications of AI-driven science, proposes ethical strategies for AI integration, and highlights key areas for future research. Finally, Appendix A is appended to this paper to demonstrate the methodologies used to compile and analyze the information presented in this paper.

2. Thematic Synthesis and Critical Analysis of the Literature

2.1. Historical and Conceptual Foundations

This section establishes the historical and theoretical groundwork for understanding the emergence of Generative Metascience, a framework that positions AI as both an analytical instrument and an autonomous agent capable of generating novel scientific hypotheses and driving independent research. By tracing AI’s evolution in scientific discovery, we illustrate how each advancement has contributed to this paradigm, highlighting the transition from data-driven analysis to generative and autonomous scientific inquiry.

2.1.1. The Genesis of the Field: Seminal Works and Key Milestones

The integration of AI into scientific discovery began in the 1960s with DENDRAL, developed at Stanford University (Buchanan & Feigenbaum, 1981). As the first expert system, DENDRAL assisted organic chemists in identifying unknown molecules by analyzing mass spectrometry data, marking an early step toward automating hypothesis formation. This pioneering work demonstrated AI’s potential to augment human reasoning, laying the foundation for its role in Generative Metascience by enabling structured, rule-based hypothesis generation.

In the 1990s and 2000s, the advent of machine learning techniques, such as support vector machines and random forests, expanded AI’s capabilities. These methods were particularly impactful in bioinformatics, facilitating tasks like gene expression analysis and protein classification (Baldi & Brunak, 2001). By extracting meaningful patterns from large datasets, machine learning enabled data-driven hypothesis generation, a critical precursor to the generative aspects of modern AI systems.

The deep learning revolution of the 2010s marked a significant leap, with neural networks achieving breakthroughs across diverse scientific domains. In 2017, artificial neural networks addressed the quantum many-body problem, a longstanding challenge in physics (Iten et al., 2020). In 2020, DeepMind’s AlphaFold solved the 50-year-old protein folding problem with unprecedented accuracy, predicting protein structures from amino acid sequences (Jumper et al., 2021). This achievement not only accelerated biological research but also exemplified AI’s ability to integrate data analysis with hypothesis generation, a core tenet of Generative Metascience.

The rise of generative AI models further advanced this paradigm. Generative adversarial networks (GANs) and variational autoencoders (VAEs) have been used to design novel molecules and materials, enabling AI to propose new scientific entities that drive hypothesis formation (Mi et al., 2018). For instance, in materials science, AI has identified promising battery material candidates, significantly reducing the time required for traditional trial-and-error methods (Lv et al., 2022).

In recent years, large language models (LLMs) and foundation models have emerged as transformative tools in scientific discovery. Models like GPT-4 have demonstrated capabilities in generating scientific text, code, and hypotheses, aligning closely with the generative metascience framework. In 2024, the AI Scientist framework enabled LLMs to conduct research autonomously, from idea generation to paper writing (Lu et al., 2024). A notable example is Sakana AI’s AI Scientist-v2, which in 2025 autonomously generated a hypothesis, designed experiments, and produced a paper accepted at a top machine learning conference, though it was later withdrawn due to ethical concerns (Yamada et al., 2025). These milestones highlight AI’s growing autonomy, positioning it as a proactive collaborator in scientific inquiry.

These developments reflect key trends: a progression from narrow, task-specific AI to general, autonomous systems; an expansion in the scale and complexity of problems addressed; and a shift from supportive to proactive roles in science. Each milestone has advanced AI’s generative capabilities, aligning with the principles of Generative Metascience by enabling AI to propose and test novel scientific ideas across disciplines.

2.1.2. The Evolution of Core Concepts and Theories

The conceptual evolution of AI in science mirrors its historical milestones, progressing from rule-based systems to data-driven, deep learning, and generative models, each addressing limitations of prior approaches and contributing to the framework of Generative Metascience. Early systems like DENDRAL relied on manually encoded rules, limiting their generalizability and requiring extensive expert input. This constrained their ability to generate novel hypotheses beyond predefined knowledge, a significant limitation for scientific discovery.

The shift to machine learning in the 1990s introduced data-driven methods, such as decision trees and ensemble methods, which learned from examples and adapted to new data (Rifkin, 2002). These approaches enabled AI to handle complex datasets, facilitating pattern recognition that informed hypothesis generation in fields like bioinformatics. This marked an early step toward generative capabilities, as AI began to identify patterns that could inspire new scientific inquiries.

The deep learning revolution of the 2010s overcame the limitations of traditional machine learning by leveraging multi-layered neural networks to model complex relationships (Sejnowski, 2018). This enabled breakthroughs in image analysis for astronomy, sequence prediction in genomics, and natural language processing for scientific literature. However, deep learning’s “black box” nature posed challenges for scientific validation, prompting the development of interpretability methods like SHAP and LIME to enhance trust in AI-driven findings (Xu et al., 2019).

Generative AI models, such as GANs and VAEs, further expanded AI’s role by enabling the creation of novel scientific entities, such as molecules and experimental designs (Mi et al., 2018). These models directly support Generative Metascience by generating hypotheses that push scientific boundaries, moving beyond analysis to innovation.

The emergence of LLMs and foundation models represents the latest advancement, enabling AI to generate scientific text, code, and hypotheses. For instance, LLMs have been used to draft research proposals and analyze literature, streamlining the scientific process (Liang et al., 2024). The AI Scientist framework exemplifies this, integrating data analysis, hypothesis generation, and experiment design into an autonomous research cycle (Lu et al., 2024). This shift challenges traditional scientific workflows, raising questions about the role of human researchers and the nature of scientific creativity.

Each stage of this evolution has addressed prior limitations: from the rigidity of rule-based systems to the scalability of machine learning, the complexity handling of deep learning, and the generative and autonomous capabilities of modern AI. This progression underscores AI’s transformative potential in redefining the scientific method, with Generative Metascience providing a framework to understand and guide this transition. Figure 1 illustrates this conceptual evolution as a chronological timeline of key AI milestones in scientific discovery, from the rule-based reasoning of DENDRAL in the 1960s to the fully autonomous AI Scientist of 2025. Together, these developments underscore how each paradigm shift—rule-based, data-driven, deep learning, and generative models—builds toward the framework of Generative Metascience (see Figure 1).

2.2. The Methodological Canvas: Approaches to Research in the Field

The exploration of artificial intelligence (AI) as a transformative scientific instrument within the framework of Generative Metascience relies on a diverse set of research methodologies. These approaches, spanning qualitative and quantitative paradigms, investigate AI’s integration into scientific workflows, its effectiveness, and its broader implications for the scientific process. This section critically evaluates these methodologies, providing specific examples from the literature, addressing their strengths and limitations, and highlighting emerging innovations that enhance the study of AI in science. Figure 2 presents a methodological canvas that maps how simulations, case studies, and experimental studies each contribute to broader scientific implications, effectiveness & validation, and AI integration within scientific workflows. See Figure 2 for a visual summary of these interrelated approaches.

2.2.1. Predominant Research Methodologies: A Critical Overview

The methodologies employed to study AI’s role in scientific discovery are varied, each offering distinct perspectives on how AI reshapes research practices. Below, we outline the primary methodologies, supported by specific examples and peer-reviewed references, to illustrate their application in the context of Generative Metascience.

• Case Studies: Case studies provide in-depth analyses of specific AI applications, offering rich, contextual insights into their integration with scientific workflows. A prominent example is DeepMind’s AlphaFold, which solved the protein folding problem with unprecedented accuracy, demonstrating AI’s capacity to address complex scientific challenges (Jumper et al., 2021). Another significant case is the development of DSP-1181, the first AI-designed drug to enter clinical trials, created through a collaboration between Exscientia and Sumitomo Dainippon Pharma for obsessive-compulsive disorder treatment. This project completed its exploratory research phase in under 12 months, compared to the traditional 4-6 years, showcasing AI’s potential to accelerate drug discovery (Burki, 2020). These cases highlight AI’s role in generating and testing hypotheses, aligning with the principles of Generative Metascience.
• Experimental Studies: Experimental studies compare AI-driven methods with traditional approaches, providing rigorous evidence of AI’s efficacy. For instance, Granda et al. (2018) demonstrated that AI-optimized chemical reaction conditions outperformed human-designed methods, achieving higher efficiency in organic synthesis (Granda et al., 2018). Such studies validate AI’s practical utility in scientific tasks, supporting its role as a generative tool in research.
• Surveys and Interviews: These methods capture scientists’ perceptions of AI, revealing adoption barriers such as lack of training or concerns about model transparency. A survey by the Center for Science, Technology and Environmental Policy Studies at Arizona State University found that while scientists recognize AI’s potential to enhance research, many express concerns about its impact on scientific integrity and the need for ethical guidelines (Z. Chen et al., 2024). Similarly, a Pew Research Center survey reported that 52% of Americans are more concerned than excited about AI’s role in daily life, reflecting broader societal apprehensions that influence scientific adoption (Zhang & Dafoe, 2019).
• Data Mining and Bibliometric Analyses: These approaches identify trends in AI’s application across disciplines by analyzing large datasets of publications. Rahman et al. (2024) conducted a bibliometric analysis of AI in medical diagnoses, noting a significant increase in machine learning and deep learning usage as such, underscoring AI’s growing influence (Rahman et al., 2024). These analyses provide a macro-level perspective on AI’s integration into scientific research.
• Simulations: Simulations test AI in virtual environments, offering cost-effective ways to evaluate its predictive capabilities. Raccuglia et al. (2016) used machine learning to simulate and accelerate the discovery of new materials, demonstrating AI’s ability to explore complex systems where physical experiments are impractical (Raccuglia et al., 2016). Such simulations support hypothesis generation, a key aspect of Generative Metascience.
• Theoretical Modeling: Theoretical models develop frameworks to understand AI’s role in science. Jordan and Mitchell (2015) proposed a model for how AI can reshape hypothesis generation and testing, providing a conceptual foundation for studying AI’s impact on scientific discovery (Jordan & Mitchell, 2015). These models guide empirical research but require validation to ensure practical relevance.

The prevalence of case studies and experimental studies reflects their ability to provide detailed insights into AI’s transformative applications and empirical evidence of its advantages. Surveys and bibliometric analyses offer broader perspectives on research trends and community perceptions, while simulations and theoretical modeling enable exploration in controlled or conceptual settings. This methodological diversity ensures a comprehensive understanding of AI’s role in advancing scientific inquiry.

2.2.2. Strengths, Weaknesses, and the Rise of Innovative Methods

Each methodology offers distinct strengths and faces inherent limitations, shaping their suitability for studying AI in science. Below, we critically analyze these aspects, supported by examples, and discuss emerging innovations that address these limitations. Table 1 provides a concise overview of these methodologies, highlighting each approach’s key strengths, notable weaknesses, and a representative example. See Table 1 for details.

Critical Analysis of Methodological Blind Spots

While these methodologies collectively advance our understanding of AI in science, they exhibit notable blind spots. Case studies, such as those on AlphaFold, often focus on high-profile successes, potentially overlooking less successful applications that could reveal critical limitations of AI systems. Experimental studies, like Granda et al. (2018), may prioritize controlled environments, missing the nuanced challenges of real-world scientific contexts, such as interdisciplinary complexities or data variability. Surveys and interviews, as seen in the ASU SciOPS study, may suffer from response biases, particularly if participants are hesitant to express critical views about AI due to its perceived transformative potential. Data mining and bibliometric analyses, while comprehensive, may miss emerging trends not yet reflected in publication databases, limiting their ability to capture cutting-edge developments. Simulations, such as those by Raccuglia et al. (2016), rely on model assumptions that may not fully represent real-world phenomena, potentially leading to overoptimistic predictions. Theoretical models, like those proposed by Jordan and Mitchell (2015), risk being overly abstract without sufficient empirical grounding, which can hinder their practical utility.

Emerging Innovative Methods

To address these limitations, innovative methodologies are emerging to enhance the rigor, transparency, and ethical integration of AI in scientific research. AI-powered literature reviews, such as those using natural language processing (NLP) to analyze thousands of papers, enable rapid synthesis of research trends, as demonstrated by tools like Semantic Scholar (Kinney et al., 2025). Automated meta-analyses leverage AI to conduct systematic reviews with reduced human bias, improving efficiency in synthesizing evidence across studies (Harrer et al., 2019). Reproducibility frameworks use AI to verify data integrity and methodological consistency, enhancing trust in AI-driven findings (Haibe-Kains et al., 2020). Social media sentiment analysis, such as analyses of X posts, gauges the scientific community’s reactions to AI advancements, providing real-time insights into adoption trends (Qi et al., 2024). Ethical assessment frameworks, like those proposed by Jobin et al. (2019), offer guidelines for addressing bias, privacy, and accountability in AI applications. These innovations align with Generative Metascience by fostering a more robust and responsible approach to studying AI’s role in science.

2.3. Thematic Deep Dive 1

This section examines artificial intelligence (AI) as a cornerstone of scientific data collection and analysis, a key pillar of Generative Metascience. By leveraging machine learning and deep learning, AI enables researchers to process vast, complex datasets with unprecedented efficiency and accuracy, driving discoveries across disciplines such as astronomy, genomics, and particle physics. Structured into synthesis, consensus/controversy, and critique, this section highlights AI’s transformative impact, addresses ongoing debates, and identifies methodological gaps requiring further exploration.

2.3.1. Synthesis of Key Findings and Supporting Evidence

AI has reshaped scientific data collection and analysis by automating complex tasks and uncovering patterns that were previously unattainable. In astronomy, AI manages massive datasets generated by modern telescopes. The Vera C. Rubin Observatory, for instance, is expected to produce 0.5 exabytes of data over its 10-year survey, equivalent to 50,000 times the Library of Congress’s book collection (Thomas et al., 2020). Neural networks have achieved high accuracy in classifying galaxy morphologies; a seminal study by Banerji et al. (2010) used neural networks on Sloan Digital Sky Survey data, achieving 98% accuracy comparable to human experts (Banerji et al., 2012). Similarly, AI-driven exoplanet detection has reached 96% accuracy using Kepler mission data, as demonstrated by Valizadegan et al. (2022) in the ExoMiner project (Valizadegan et al., 2022).

In genomics, AI accelerates the analysis of intricate genetic data, advancing personalized medicine and disease research. DeepVariant, a deep learning-based variant caller developed by Google, significantly improves the accuracy of identifying genetic variants, outperforming traditional tools without requiring specialized domain knowledge (Poplin, Chang, et al., 2018; Poplin, Varadarajan, et al., 2018). SpliceAI, a 32-layer deep neural network, predicts splicing from DNA sequences with up to 95% accuracy, aiding in identifying cryptic splicing variants linked to neurodevelopmental disorders (Jaganathan et al., 2019). These tools exemplify AI’s ability to handle the scale and complexity of modern sequencing data.

In particle physics, AI is critical for analyzing petabytes of data from high-energy collisions at the Large Hadron Collider (LHC). The CMS collaboration employs AI to detect anomalous jets, enhancing sensitivity to new physics phenomena (C. M. S. Collaboration, 2021). Similarly, ATLAS uses deep learning for precise identification of b-hadrons and other particles (A. Collaboration, 2020). These applications underscore AI’s role in managing data complexity and volume.

AI’s impact extends to chemistry, where it predicts molecular properties, and materials science, where it accelerates novel material discovery. In environmental science, AI models climate data and predicts natural disasters, enabling proactive responses. By automating routine tasks and revealing hidden patterns, AI frees researchers to focus on innovative hypotheses, aligning with Generative Metascience’s emphasis on advancing scientific inquiry. Table 2 summarizes key AI applications across major scientific fields, detailing specific examples and their impacts. See Table 2 for a concise overview.

2.3.2. Areas of Consensus and Controversy

The scientific community broadly recognizes AI’s transformative potential in data collection and analysis, particularly for enhancing efficiency and accuracy in processing large datasets. Machine learning and deep learning have become indispensable for tasks like data cleaning and pattern recognition, driving discoveries across multiple fields.

However, several controversies persist:

• Interpretability: Deep learning models often operate as “black boxes,” delivering accurate predictions without transparent decision-making processes. This opacity is problematic in fields like medical research, where mechanistic understanding is essential for trust and validation (Obermeyer et al., 2019). In contrast, some argue that high accuracy, such as in galaxy classification, may suffice when categorization is the primary goal (Banerji et al., 2012).
• Data Quality and Bias: AI’s effectiveness depends on training data quality. In genomics, datasets skewed toward European ancestry can produce biased models, risking health disparities (Popejoy & Fullerton, 2016). Debates center on whether technical solutions like debiasing algorithms or inclusive data collection are more effective.
• Overreliance on AI: While AI streamlines research, overreliance without human oversight may lead to “illusions of understanding,” where researchers misinterpret AI outputs (Topol, 2019). Some view AI as a tool to enhance creativity, while others caution it could undermine scientific rigor.
• Ethical Concerns: Data privacy, especially in genomics, raises significant issues. Balancing AI-driven insights with privacy protection requires robust ethical guidelines to maintain public trust (Jobin et al., 2019).

These debates highlight the need for balanced approaches to ensure AI’s benefits are realized without compromising scientific integrity or equity.

2.3.3. Critique of the Literature and Methodological Limitations

The literature on AI in data collection and analysis is robust but reveals methodological gaps. A key limitation is the reliance on benchmark datasets, which may not reflect real-world diversity. For instance, DeepVariant’s training on NIST cell lines limits its applicability to diverse genomic data (Poplin, Chang, et al., 2018). In astronomy, models trained on specific telescope data, such as the Sloan Digital Sky Survey, may overfit and underperform on other instruments, necessitating broader datasets (Banerji et al., 2012).

The lack of standardized evaluation metrics hinders performance comparisons across studies. For example, exoplanet detection studies use varied accuracy metrics, complicating assessments (Valizadegan et al., 2022). Calls for domain-specific benchmarks, such as those in Kaggle’s exoplanet hunts, highlight the need for consistency (Jin et al., 2022).

Interpretability remains a barrier, as the “black box” nature of deep learning models limits validation in hypothesis-driven research. Emerging explainable AI (XAI) methods, like SHAP and LIME, show promise but require further development (Lundberg et al., 2022).

Ethical oversights, particularly in genomics, are prevalent, with insufficient attention to bias mitigation and data privacy. Inclusive datasets and transparent reporting are critical to address disparities (Hanna et al., 2025).

Finally, the scarcity of longitudinal studies limits understanding of AI’s long-term impact on scientific practices. Future research should prioritize real-world validations, standardized metrics, improved interpretability, and ethical compliance to fully harness AI’s potential in data collection and analysis, aligning with Generative Metascience’s goal of advancing scientific discovery.

2.4. Thematic Deep Dive 2

2.4.1. Synthesis

Hypothesis generation and testing are cornerstones of the scientific method, driving the discovery of new knowledge. Artificial intelligence (AI) has emerged as a transformative tool in these processes, leveraging its ability to process vast datasets, identify complex patterns, and make accurate predictions. This section explores how AI enhances hypothesis generation and testing across diverse disciplines, illustrating its role in accelerating scientific discovery.

In materials science, AI uncovers novel insights that might elude human researchers. For instance, a collaboration between Microsoft and Pacific Northwest National Laboratory used AI to screen over 32 million potential battery materials, identifying 23 promising candidates, one of which was synthesized into a working prototype (C. Chen et al., 2024). This approach significantly reduces the time required for traditional trial-and-error methods, enabling exploration of vast chemical spaces for energy storage innovations.

In medical research, AI generates hypotheses that reveal unexpected correlations. A study in Communications Medicine used deep learning to analyze histopathology images of prostate cancer patients, identifying gland morphology patterns, such as well-formed glands, as predictors of biochemical recurrence (Bulten et al., 2022a). This finding demonstrates AI’s ability to detect subtle patterns in complex datasets, opening new avenues for personalized medicine.

In the social sciences, AI reveals hidden factors influencing human behavior. Ludwig and Mullainathan (2024) employed machine learning to analyze judicial decisions, finding that facial features accounted for up to half of the predictable variation in jailing decisions (Ludwig & Mullainathan, 2024). This led to hypotheses about unconscious biases, highlighting AI’s potential to inform social science research.

AI also enhances hypothesis testing by optimizing experimental design. The Copilot for Real-world Experimental Scientist (CRESt), developed at MIT, assists materials science researchers by suggesting experiments and controlling equipment via voice commands (Ren et al., 2023). Using active learning, CRESt streamlines workflows, as shown in studies on fuel cell catalysts, reducing research time and enhancing efficiency.

In computational biology, AI facilitates hypothesis testing through simulations. Deep learning models predict how mutations alter protein function, enabling researchers to prioritize high-potential variants for experimental validation. This approach saves resources and allows exploration of vast mutational landscapes, accelerating discoveries in fields where physical experiments are costly.

These examples demonstrate AI’s transformative role in hypothesis generation and testing, enabling researchers to explore new frontiers and optimize scientific workflows.

2.4.2. Controversy

While issues like interpretability and ethics recur throughout discussions of AI in science, their persistence underscores their centrality as cross-cutting concerns that impact multiple facets of AI applications. The scientific community recognizes AI’s potential to enhance hypothesis generation and testing, as evidenced by successes like AlphaFold’s protein structure predictions and AI-driven material discoveries. However, several controversies persist, reflecting diverse perspectives on AI’s role in science.

Interpretability remains a significant challenge. The “black box” nature of many AI models complicates validation, as the scientific community requires understanding the mechanisms behind predictions. For instance, in Ludwig and Mullainathan’s judicial study, the inability to pinpoint which facial features drove predictions limited actionable conclusions, hindering adoption.

A key debate centers on whether AI-generated hypotheses are truly novel or merely extrapolations of data patterns. Some argue AI lacks human creativity, producing sophisticated correlations without deeper insight. Nick Bostrom suggests creativity requires intentionality, which current AI lacks (Müller & Bostrom, 2016). Conversely, Margaret Boden argues AI can be creative if it produces novel and valuable outputs, as seen in the prostate cancer study (Boden, 1996). Researchers have begun exploring metrics to quantify novelty, such as using information theory to measure the Kullback-Leibler divergence between predicted and observed outcomes, assessing the surprisal or innovativeness of AI proposals (Foster et al., 2021). These approaches, though promising, remain underdeveloped for broad scientific application. This philosophical discourse challenges traditional notions of scientific innovation.

AI’s influence on scientific theory change raises epistemological questions. Thomas Kuhn’s concept of paradigm shifts suggests AI could accelerate revolutions by identifying anomalies (Kuhn & Meyer, 1983). However, critics warn of “algorithmic dogmatism,” where biases in AI models shape theories, potentially stifling innovation. These debates underscore the need to critically assess AI’s role in reshaping scientific epistemology.

Ethical concerns are prominent, especially in healthcare and criminal justice. Biased data can lead to flawed hypotheses, as seen in the judicial study, raising fairness and privacy issues. Balancing AI’s insights with robust ethical guidelines is crucial to maintain public trust.

Overreliance on AI risks “illusions of understanding,” where researchers overestimate comprehension of phenomena. A hybrid approach combining AI and human expertise is advocated to ensure scientific rigor. For instance, in developing the AI-designed drug DSP-1181, human experts were essential in validating AI proposals and guiding clinical trials (Burki, 2020).

2.4.3. Critique

The literature on AI in hypothesis generation and testing is robust but reveals methodological limitations that must be addressed to fully harness AI’s potential.

• Dataset Limitations: Many studies rely on narrow datasets, limiting generalizability. For example, in materials science, AI models trained on specific material types may not apply broadly, necessitating more diverse datasets. For example, one AI-driven screen predicted a novel antibiotic candidate, but follow-up laboratory assays revealed it had no measurable antimicrobial activity. This false lead underscores the indispensability of empirical validation in AI-generated hypotheses.
• Interpretability Challenges: The “black box” nature of AI models hinders validation, particularly in hypothesis-driven research. Emerging explainable AI methods like SHAP and LIME show promise, but further development is needed (Lundberg et al., 2022).
• Lack of Standardized Metrics: Evaluating AI-generated hypotheses lacks standardized metrics. Current benchmarks, like the DREAM Challenges for drug discovery, focus on predictive accuracy rather than novelty (Prill et al., 2010). In materials science, the Materials Project evaluates properties but not innovativeness (Jain et al., 2013). Novel metrics, possibly from information theory, are needed to assess hypothesis originality. For instance, approaches based on information theory, such as measuring the Kullback-Leibler divergence between predicted and observed outcomes, offer potential ways to assess the surprisal or innovativeness of AI proposals (Foster et al., 2021). However, these methods are still in early stages and require further development to be widely applicable in scientific contexts.

Ethical Oversights: Ethical considerations in AI-driven hypothesis generation and testing are multifaceted:

• Data Bias: Studies often rely on datasets with inherent biases, leading to hypotheses that amplify these flaws. In healthcare, AI trained on specific populations may produce hypotheses less applicable to underrepresented groups (Parikh et al., 2019).
• Privacy Concerns: Fields like genomics and social sciences involve sensitive data, raising privacy issues. Compliance with ethical standards and data protection regulations is essential (Jobin et al., 2019).
• Societal Implications: In areas like criminal justice, AI-generated hypotheses can impact society broadly, as seen in the judicial study, necessitating fairness and accountability (Ludwig & Mullainathan, 2024). Addressing these requires inclusive data practices, transparency, and adherence to guidelines like those from the IEEE (Zhao et al., 2020).
• Integration with Traditional Practices: AI-generated hypotheses require rigorous experimental validation. Open science practices, such as sharing models and datasets, are crucial for transparency and reproducibility. Crucially, human oversight remains indispensable in this process. While AI can propose hypotheses, human researchers must evaluate their plausibility and interpret results, as seen in the development of the AI-designed drug DSP-1181, where human experts guided validation and clinical trials (Burki, 2020).
• Longitudinal Studies: Few studies evaluate the long-term impact of AI-generated hypotheses. Future research should prioritize real-world validations, novel metrics, interpretability advancements, and ethical compliance to ensure AI’s responsible application in science.

2.5. Cross-Thematic Analysis: Interconnections and Contrasting Perspectives

This section synthesizes the two central themes of this review, AI’s role in data collection and analysis (Theme A, Section 2.3) and hypothesis generation and testing (Theme B, Section 2.4), to elucidate their interconnections, shared challenges, and contrasting perspectives. By examining these relationships, we highlight how AI drives the emerging paradigm of algorithmic discovery, framed within the concept of Generative Metascience, which positions AI as a meta-technology that both analyzes data and autonomously generates novel scientific inquiries. This synthesis provides new insights into AI’s transformative potential while addressing critical limitations, supported by concrete examples of validated hypotheses.

Defining Generative Metascience

Generative Metascience is a framework that conceptualizes AI as both an analytical instrument and an autonomous co-investigator in scientific discovery. It integrates data-driven pattern recognition (Theme A) with hypothesis-driven exploration (Theme B), enabling AI to automate and augment the scientific method. By facilitating an iterative cycle of observation, hypothesis formulation, experimentation, and analysis, Generative Metascience redefines traditional research workflows, as exemplified by systems like the AI Scientist and automated laboratories.

Interconnections: The Iterative Cycle of Data and Hypotheses

AI’s capabilities in data collection and analysis (Theme A) provide the empirical foundation for hypothesis generation and testing (Theme B), creating a synergistic cycle that accelerates scientific discovery. In genomics, DeepVariant, a deep learning-based variant caller, identifies genetic variants with high accuracy (Junjun et al., 2024). For instance, its detection of rare variants in the 1000 Genomes Project has led to hypotheses about their roles in neurodevelopmental disorders, which are subsequently tested through targeted experiments (Yun et al., 2020).

In astronomy, AI-driven classification of galaxy morphologies using Sloan Digital Sky Survey data achieves 98% accuracy, informing hypotheses about galaxy formation and evolution (Banerji et al., 2012). These hypotheses guide further observations, such as those with the Hubble Space Telescope, to validate models of cosmic evolution.

In materials science, the A-Lab project at Lawrence Berkeley National Laboratory used AI to screen over 32 million candidate materials, identifying 41 new compounds, one of which was synthesized into a working battery prototype with 70% less lithium (Banerjee et al., 2025). This validated hypothesis demonstrates how AI-driven data analysis leads to novel material discoveries, which are then tested experimentally, generating new data for further analysis.

In particle physics, AI analyzes petabytes of Large Hadron Collider (LHC) data to detect anomalous jets, suggesting the presence of particles beyond the Standard Model (Kheddar et al., 2025). These anomalies prompt hypotheses about new physics phenomena, tested through subsequent collisions, illustrating the iterative cycle.

This cycle is further enhanced by large language models (LLMs) like GPT-4, which process scientific literature to suggest hypotheses and streamline research workflows (Liang et al., 2024). For example, LLMs have been used to draft research proposals in genomics, identifying potential gene-disease associations for experimental validation. Figure 3 depicts the iterative cycle connecting Theme A (Data Collection & Analysis) and Theme B (Hypothesis Generation & Testing), showing how algorithmic discovery feeds pattern finding and insight into hypothesis formulation, which then drives experiment design, execution, and the generation of new data—closing the loop on scientific inquiry (see Figure 3).

Shared Challenges Across Themes

Both themes face common challenges that must be addressed to realize AI’s full potential:

• Interpretability and Bias: As explored in section 2.3 and section 2.4, the “black box” nature of AI models and risks from biased datasets persist as key challenges. These issues undermine trust in Theme A (e.g., DeepVariant) and adoption in Theme B (e.g., medical imaging), necessitating advances in explainable AI and inclusive data.
• Ethical Considerations: Applications in sensitive fields raise privacy and fairness concerns. In Theme A, genomic data privacy is critical. In Theme B, AI-generated hypotheses in criminal justice risk amplifying biases, necessitating robust ethical guidelines.

Contrasting Objectives and Evaluation Metrics

Theme A and Theme B differ in their objectives and evaluation methods:

• Objectives:

  o

  Theme A aims for precision in quantifying known patterns, such as classifying galaxies or identifying genetic variants (Davis & Goadrich, 2006).

  o

  Theme B seeks novelty, proposing uncharted research directions, such as predicting cancer recurrence based on gland morphology (Bulten et al., 2022).

• Evaluation Metrics:

  o

  Theme A uses quantitative metrics like accuracy and precision, as seen in exoplanet detection with 96% accuracy (Jin et al., 2022; Valizadegan et al., 2022).

  o

  Theme B requires experimental validation, which is inherently uncertain and long-term, as in the A-Lab’s material synthesis (Banerjee et al., 2025).

This contrast highlights AI’s dual role: Theme A provides concrete answers, while Theme B poses innovative questions, necessitating human oversight to filter spurious correlations. For example, AI analysis of electronic health records has proposed hypotheses linking hospital visits to rare diseases, later disproven as statistical noise (Topol, 2019).

The Paradigm of Algorithmic Discovery

The integration of Themes A and B forms the paradigm of algorithmic discovery, where AI autonomously drives the research cycle. The AI Scientist, which generates hypotheses, designs experiments, and drafts papers, exemplifies this paradigm (Lu et al., 2024). Similarly, the A-Lab’s discovery of 41 new materials underscores AI’s ability to predict and validate novel compounds, accelerating materials science research. Table 3 summarizes how Themes A (data analysis) and B (hypothesis generation) interact across major fields, detailing the AI-driven workflow and its concrete outcomes. See Table 3 for the integrated paradigm of algorithmic discovery.

3. Discussion, Implications, and Future Trajectories

This section consolidates the review’s findings, identifies gaps, and charts future directions. It begins with a summary of AI’s contributions (3.1), examines unresolved questions (3.2), explores theoretical and practical implications (3.3), and concludes with actionable recommendations (3.4).

3.1. Integrated Summary of Key Insights

Artificial intelligence (AI) is transforming scientific discovery through its dual roles in data collection and analysis (Theme A) and hypothesis generation and testing (Theme B). Theme A showcases AI’s ability to process vast datasets with high efficiency and accuracy in fields such as astronomy, genomics, and particle physics. For example, AI tools like DeepVariant enhance genomic analysis by identifying genetic variants with precision. Theme B highlights AI’s capacity to propose novel hypotheses and optimize experimental designs in areas like materials science and medicine, as seen in the A-Lab’s discovery of new battery materials. The interplay between these themes forms a synergistic cycle where data analysis informs hypothesis generation, and hypothesis testing generates new data, accelerating scientific progress. This cycle underpins the emerging paradigm of algorithmic discovery, exemplified by systems like the AI Scientist, which autonomously conducts research. However, as noted in section 2.3 and section 2.4, challenges like model interpretability, data bias, and ethical concerns, particularly in sensitive fields such as healthcare and social sciences, remain critical to address. Human oversight remains essential to ensure scientific rigor and ethical integrity, positioning AI as a transformative yet carefully managed scientific instrument. These findings underscore AI’s potential while highlighting unresolved issues, which we explore further in the following sections.

3.2. Unanswered Questions and Gaps in the Literature

Several gaps in the literature require attention to optimize AI’s role in science. Theoretically, AI’s role in knowledge generation prompts epistemological questions about its divergence from human cognition. Bostrom (2014) contends that true creativity hinges on intentionality, absent in AI, implying that Theme B’s hypothesis generation (e.g., A-Lab’s material discoveries; Merchant et al., 2023) reflects advanced pattern extrapolation rather than innovation. Conversely, Boden (1990) argues that AI achieves creativity when producing novel, valuable outputs, as evidenced by Theme A’s data-driven breakthroughs (e.g., DeepVariant; Poplin et al., 2018). Additionally, Kuhn’s (1962) paradigm shift framework suggests that AI’s rapid anomaly detection, seen in particle physics, could hasten scientific revolutions, amplifying Theme B’s impact. Methodologically, the “black box” nature of AI models, reliance on narrow datasets, and lack of standardized evaluation metrics limit their trustworthiness and generalizability. For instance, biased genomic datasets can exacerbate health disparities (Popejoy & Fullerton, 2016). Empirically, longitudinal studies are needed to assess AI’s long-term impact on scientific productivity and innovation. Systematic research into ethical implications, particularly in healthcare and criminal justice, is also lacking, necessitating robust frameworks to address bias, privacy, and fairness.

3.3. Implications for Theory and Practice

AI’s ability to autonomously generate and test hypotheses challenges traditional human-centric models of scientific discovery, necessitating new theoretical frameworks to integrate machine-driven insights, as suggested by Kuhn’s paradigm shift concept (Anand et al., 2020). This raises epistemological questions about knowledge validity and reproducibility in computational research. Practically, integrating AI requires scientists to gain expertise in AI and data science through training programs, such as those supported by the National Science Foundation’s AI initiatives (Ju et al., 2024). Research institutions must invest in computational infrastructure and foster interdisciplinary collaborations to support AI integration. Ethical guidelines, like those outlined in the European Union’s AI in Science guidelines (Nannini et al., 2023), are essential to address bias, privacy, and accountability, particularly in healthcare and social sciences. Policymakers should establish regulations for data sharing and model validation to maintain public trust and promote responsible innovation.

3.4. Recommendations for Future Research

To address these gaps and harness AI’s potential, future research should prioritize the following, divided into short-term and long-term goals:

Short-Term Priorities:

• Explainable AI (XAI): Develop domain-specific XAI methods, building on tools like SHAP and LIME from section 2.4, to enhance transparency (Lundberg et al., 2022).
• Inclusive Datasets: Compile diverse datasets to reduce biases in genomics and social sciences, as seen in Theme A (Popejoy & Fullerton, 2016).
• Standardized Metrics: Establish benchmarks to evaluate hypothesis generation, drawing from Theme B’s novelty needs (Prill et al., 2010).
• Long-Term Priorities:
• Longitudinal Studies: Assess AI’s sustained impact on productivity, extending Theme B’s real-world applications.
• Ethical Frameworks: Develop guidelines for fairness and privacy, addressing concerns from section 2.3 (Alvarez et al., 2024).
• Human-AI Collaboration: Optimize workflows by integrating AI with human expertise, enhancing Theme A and B synergy.

These efforts will enable the scientific community to integrate AI responsibly, maximizing its benefits while mitigating risks.