AI in Science: Balancing Accessibility and Trust

Making informed decisions about health and policy requires a strong grasp of scientific evidence. While artificial intelligence (AI) plays a crucial role in improving access to science, it also raises concerns about the reliability of information. As AI reshapes how we communicate and produce scientific data, the questions surrounding understanding and trust in that data become more critical than ever.

The Evolution of Access to Scientific Literature

For decades, accessing scientific literature has been a significant barrier for many individuals seeking to make informed decisions. However, the rise of open-access publishing has transformed the landscape, allowing more researchers, patients, and policymakers to read and download literature without a paywall. Reports suggest that about 61% of journal articles published in 2025 were open access. Unfortunately, access alone doesn’t equate to comprehension or trust.

AI: Enhancing Comprehension Amidst New Challenges

Generative AI technologies like ChatGPT are increasingly facilitating comprehension. They assist researchers in literature navigation, help patients understand clinical papers, and aid policymakers in interpreting findings outside their fields. However, the efficiency of generating scientific-sounding content has outpaced the necessary verification of its accuracy. Rapid content generation is becoming cheaper while the verification process requires more time, expertise, and access to evidence.

Two Vulnerabilities in Research Integrity

AI-induced vulnerabilities can compromise trust in scientific literature. Firstly, AI can generate or manipulate research components that could potentially make their way into scientific records. Secondly, genuine studies can be misconstrued when findings are presented selectively. Research integrity hangs in the balance, where the challenge is not merely the proliferation of AI but also the authenticity and context of the information presented.

Recent studies highlight the gravity of reliance on AI in research:

• A study revealed that GPT-4o fabricated citations 78-90% of the time when unassisted by external information sources.
• Another study identified rising rates of suspected fabricated references in biomedical literature, indicating that some misleading citations pass through editorial checks unnoticed.

These findings underscore the urgency of ensuring that AI’s contributions to scientific knowledge are traceable and verifiable.

Recognizing Credibility Signals

Scientific readers typically look for familiar signs of credibility—manuscript structure, research methods, institutional affiliations, citations, and polished figures. While AI can reproduce these characteristics, it does not guarantee the underlying research’s authenticity. Compromised trust arises when AI-generated content appears credible merely by mimicking the patterns of scientific writing.

Moreover, AI’s ability to produce internal consistency can mask the manipulation of data—crafting narratives that appear credible but are fundamentally flawed. Authentic research demands reliable data, appropriate methodologies, transparency, and accountability.

The Danger of Distortion

Even when AI-generated summaries cite genuine studies, misinterpretation can occur. For example, an AI model may cite an observational study connecting supplement use to reduced dementia rates, presenting an association as evidence of causality. This misrepresentation occurs even when citations are accurate, undermining the original research’s integrity.

AI may also overlook methodological issues, conflating preliminary findings with established knowledge or creating a false consensus by ignoring conflicting evidence. Hence, scientific literacy needs to encompass critical questions about the origin, evidential support, peer-review status, and any changes made through AI.

The Limitations of Detection Tools

Detecting AI-generated text does not equate to verifying research reliability. A manuscript could incorporate AI-assisted language while still presenting robust research, whereas a human-written text could contain fabricated data or methods. Reliance on AI detection may lead to false positives, diverting attention from the actual verification needed to ensure valid research outcomes.

Editorial checks should not focus solely on identifying AI-generated content; they need to confirm the authenticity of the research itself regardless of how it was articulated. While AI can aid in integrity checks, human oversight remains crucial.

The Burden of Verification

Expecting readers to independently validate research findings is unrealistic. Verification demands access to sources, specialized knowledge, and time—resources not all readers possess. As AI drives down the cost of producing research, the cost of verifying it has not seen a proportional decrease, thereby placing a heavier burden on researchers, editors, and readers alike.

Expert review remains indispensable, but the expertise in specific fields is limited; the expectation that reviewers will uncover every instance of fabrication or manipulation without appropriate support puts significant strain on the system.

The Role of Publishers in Maintaining Integrity

Academic publishers play a critical role in safeguarding the scientific record. They must establish robust editorial processes, enforce transparency regarding AI usage, and maintain clear expectations for proper research reporting. This includes information about peer review status, corrections, and version histories that are accessible and machine-readable.

Determining when and what checks to implement depends on assessed risks. Enhanced scrutiny should focus on references, images, or datasets requiring closer examination. However, overly burdensome processes could hinder the publication of legitimate research.

The Shared Responsibility for Trust in Science

Trust in scientific findings is a collective responsibility that extends beyond publishers. Researchers must ensure the authenticity of their work and the appropriate application of AI while adhering to established research standards. Institutions should provide oversight, support training, and implement clear AI use policies.

AI providers must also be accountable—ensuring their systems retrieve, transform, and represent research responsibly. This includes linking claims to reliable sources, reflecting the evidence’s provenance, and distinguishing AI-generated interpretations from research findings.

Scientific literacy can thrive only when transparent, accessible evidence is provided, alongside an understanding of its limitations. The emphasis must shift from merely giving an impression of trustworthiness to enabling informed decisions based on robust evaluations of scientific claims.