AI in Scientific Research: Data Analysis and Academic Writing Emerge as Core Use Cases

In recent years, large language models (LLMs) have demonstrated strong capabilities in natural language processing, code generation, and logical reasoning. They are gradually reshaping traditional research workflows.

Recently, an AI research organization published a report based on a large volume of real-world conversations. The report analyzed more than 200,000 research-related sessions, including both general conversations and coding sessions conducted through AI programming tools. It examined how scientists use LLMs in practice, how usage varies across disciplines, and which tasks researchers are most likely to delegate to AI.

1. How AI Is Used in Research: Data Processing and Writing Lead the Way

The report clearly identifies the primary areas where LLMs create value in research. Data analysis and academic writing account for the largest shares of research-related tasks.

Data Analysis: A Major Opportunity for Productivity Gains

Data analysis represents 33.7% of general research conversations and as much as 48.1% of coding conversations. By comparison, researchers typically spend 27.6% of their working time on data analysis. This indicates that LLMs can have a significant leverage effect when assisting with research data processing.

Research data workflows often include data cleaning, missing-value handling, statistical analysis, and visualization. LLMs can generate Python or R code using libraries such as Pandas and Matplotlib, while also helping researchers refine statistical methods through interactive discussions.

The following example demonstrates how an LLM API can assist with exploratory research data analysis:

import pandas as pd
from llm_platform_sdk import ChatModel  # Example SDK for a general-purpose LLM platform
def generate_data_insight(data_path, user_query):
    # Load experimental data
    df = pd.read_csv(data_path)
    # Initialize the language model
    model = ChatModel(model="advanced-llm")
    # Combine a data summary with the user's request
    prompt = f"""
    You are a data science expert. Here is a summary of my experimental data:
    {df.describe()}
    First five rows of the data: {df.head()}
    My request is: {user_query}
    Please provide an analysis approach and generate the corresponding Python code.
    """
    response = model.chat(prompt)
    return response

Writing and Communication: From Literature Reviews to Publication

Writing and communication tasks account for 32% of research conversations, substantially exceeding the 16% of researchers’ working time typically devoted to these activities.

Researchers must prepare papers, grant proposals, experimental reports, and peer-review comments. LLMs can assist with language polishing, document organization, terminology standardization, and grammar checking. These capabilities can significantly shorten the time between completing an experiment and producing a publication-ready document.

Current Limits in Deeper Reasoning: Hypothesis Generation and Data Collection

Data collection accounts for only 11.5% of research conversations. Research ideation and hypothesis generation represent an even smaller share: 4.2% of general conversations and 2.1% of coding conversations.

This pattern reflects an important boundary in current LLM capabilities. AI performs well when operating within an existing logical framework, such as generating code or producing structured text. However, it cannot yet replace human insight in tasks that require cross-disciplinary intuition, original breakthroughs, or complex experimental design.

Similarly, physical data collection—including laboratory procedures and instrument-based measurements—falls outside the capabilities of software-only AI systems.

2. Differences Across Disciplines: Quantitative and Computational Fields Lead Adoption

AI adoption varies significantly across academic disciplines. These differences are influenced not only by the number of researchers in each field, but also by how easily research activities can be digitized and supported by AI.

High Usage in Computer Science and Quantitative Research

Researchers in computer science, mathematics, and statistics are among the most frequent users of AI tools. Computer science researchers account for just 4.3% of the global research population, yet they contribute 32.7% of coding sessions.

The gap is even more pronounced in artificial intelligence and machine learning. Researchers in AI and ML represent 1.2% of the research population but account for 5% of research conversations and 15.2% of coding sessions.

The reason is straightforward. Computer science and AI researchers generally have strong programming skills and can quickly integrate LLMs into their development environments. They use AI programming tools for algorithm validation, model debugging, and code refactoring, accelerating the research and development cycle.

Relatively Lower Usage in Experimental Sciences and the Humanities

AI usage is comparatively lower in fields such as medicine, the physical sciences, the arts, and the humanities.

Physical sciences rely heavily on physical instruments, materials, and laboratory experiments. Their core outputs—experimental measurements and observations—cannot be directly generated by an LLM.

The humanities and social sciences involve substantial amounts of text, but they also require deep contextual understanding and careful examination of historical sources. The risk of hallucination in current LLMs limits their reliability for this type of rigorous academic research.

3. Privacy and Compliance: The Baseline for Research AI

Data security is especially important in research environments. Unpublished hypotheses, experimental methods, and preliminary findings may contain valuable intellectual property.

The report emphasized that, when analyzing large-scale conversation data, the system retained only labels such as discipline, subfield, and task category. It did not extract or store the specific content of research discussions.

Organizations and research institutions should establish clear security and compliance boundaries before introducing AI capabilities. Enterprise LLM services from cloud providers such as Baidu AI Cloud may support private deployment or VPC-based network isolation, allowing research teams to process experimental data in controlled environments and keep sensitive information within the designated network boundary.

Logging, monitoring, and alerting mechanisms can also be used to audit LLM calls throughout the workflow and reduce the risk of data leakage.

4. Practical Recommendations for AI-Assisted Research

Based on these findings, researchers planning to integrate AI into their workflows can consider the following practices:

1. Start with high-impact productivity tasks. Prioritize data processing, such as writing analysis scripts, and document preparation, such as formatting experimental reports. This can reduce repetitive work and allow scientists to focus more on innovation and hypothesis development.

2. Build a research-specific knowledge base with RAG. General-purpose LLMs may lack specialized knowledge in a particular field. Retrieval-augmented generation (RAG) can connect a model with local literature collections and historical experimental records. By storing domain-specific documents in a vector database, research teams can provide evidence-based context and reduce the risk of hallucinations.

3. Develop AI-assisted programming skills. Researchers in experimental fields such as bioinformatics and computational chemistry should learn how to use intelligent programming tools. Even without a computer science background, natural-language instructions can be used to generate data-processing code and improve the efficiency of experimental analysis.

Conclusion

LLMs are evolving from general-purpose chat tools into specialized research copilots. Current usage data shows that AI has already demonstrated clear value in data analysis and academic writing.

As multimodal models develop and domain-specific fine-tuning matures, AI may play a greater role in deeper research activities, including experimental design optimization and hypothesis exploration. Researchers should actively explore AI-assisted workflows suited to their disciplines while maintaining strong standards for data privacy, security, and compliance.