Why feeding AI with documents is not enough for high-quality scientific review

Researcher recommends an incremental approach to analyzing papers, enabling greater control, traceability, and critical thinking

Juliana Protásio

Stacks of books piled haphazardly on a wooden surface, photographed from a perspective with selective focus and warm ambient light

Platforms designed for academic research, such as Consensus, Elicit, and SciSpace, have streamlined processes that once required weeks of manual work. Generative artificial intelligence (AI) has become an increasingly important tool in scientific research, but its use requires methodological rigor to ensure the integrity of analyses.

The main risk lies in what is known as “lazy prompting”—creation of imprecise prompts that delegate critical thinking to the machine. In literature review, the most common manifestation of this problem is uploading large volumes of PDF files at once without providing the model with sufficient context for accurate processing.

In an interview with Science Arena, neurologist João Brainer, a clinical researcher at Einstein Hospital Israelita and professor at the Federal University of São Paulo (UNIFESP), warned about the risks of this approach.

“People believe AI is capable of developing critical thinking, but that doesn’t exist,” said Brainer.

Hallucination and loss of context

The problem stems from the architecture of large language models (LLMs). As the volume of information included in a prompt increases, the system’s accuracy tends to decline.

“The more information I feed in, the greater the chance that the AI system will hallucinate,” warns Brainer.

Bulk document processing can compromise the analytical consistency of AI models. The result is often a low-reliability analysis characterized by three major flaws:

• Mixing authors and references from different studies;
• Loss of essential information diluted across large volumes of text;
• Generic conclusions.

“If you simply ask the system to take what already exists and replicate it, you’ll end up with a lot of nonspecific texts. In science, you need to understand and identify where the problem lies, understand the reality, talk to other people, and read similar articles,” explains Brainer.

According to the researcher, this lack of depth undermines the rigor required by cutting-edge science.

There is also a frequently overlooked technical limitation: most tools limit direct uploads to about ten articles at a time.

Attempting to circumvent this restriction through general-purpose platforms or overly broad contexts can overwhelm the model’s ability to process information accurately.

An incremental approach as a solution

To preserve the quality of a literature review, Brainer recommends an incremental strategy: providing the AI with small batches of articles, for example, five PDFs at a time.

The proposed workflow is outlined below. When applied systematically, it enables cross-checking and helps reduce both model errors and inconsistencies in human analysis.

“Computational intelligence is still, and should continue to be, subservient to human intelligence; it should not be viewed as a substitute,” notes Brainer.

In his view, the researcher’s critical thinking and judgment remain essential for ensuring ethics, transparency, and meaningful scientific contributions.

How to process scientific articles with AI in incremental steps

Small-group upload: Upload a limited set of articles (e.g., five PDFs) and analyze the response generated by the tool.
Incremental addition: Add the next group of articles and ask the AI model to identify what new information the additional set contributes compared with the previous one.
Traceability of claims: Require the tool to justify its conclusions and indicate the specific passages in each article that support every claim.

Watch the full livestream with João Brainer below:

* This article may be republished online under the CC-BY-NC-ND Creative Commons license.
The text must not be edited and the author(s) and source (Science Arena) must be credited.