How Researchers Use Generative AI: Data Analysis and Writing Lead the Way

In recent years, generative AI has gradually evolved from a general-purpose question-answering tool into an auxiliary system embedded in research workflows. It can process natural language, assist with programming, interpret statistical results, organize materials, and improve text expression, which gives it strong versatility in scientific research. However, actual usage data shows that researchers do not apply AI evenly across the entire research process. Instead, they concentrate it on relatively standardized, verifiable, and iteratively refinable tasks—primarily data analysis and writing/communication.

1. Overall Patterns in Research AI Usage

A research study focused on scientific research scenarios analyzed approximately 207,000 research-related conversations, covering general dialogue, collaborative work sessions, and code-assistance sessions. The study retained only classification information such as discipline, research subfield, and task type; it did not store research data, experimental protocols, unpublished conclusions, research hypotheses, or manuscript text. Results were presented in aggregated statistical form.

The findings show that researchers most commonly use AI for two types of tasks:

1. Data processing and analysis
   Including data cleaning strategy, statistical method selection, result interpretation, visualization design, code debugging, and documentation of analysis workflows.

2. Paper writing and research communication
   Including language polishing, paragraph restructuring, abstract generation, explanation of research results, preparation of peer communication materials, and translating technical content into plain language.

These two task types share a common trait: inputs and outputs are relatively well-defined, the process can be reviewed by researchers, and results can be verified against data, literature, or experiments. AI is therefore better suited as an “accelerator” and “review assistant” rather than a replacement for researchers’ final judgment.

2. Significant Differences Across Disciplines

AI usage frequency in research is closely tied to discipline size, task structure, and degree of digitalization.

Computer science, mathematics, statistics, and some engineering fields show relatively high usage rates. These disciplines typically share the following characteristics:

• Research materials consist mainly of code, datasets, formulas, and text;
• Work can be completed in digital environments;
• Tasks can be broken down into well-defined steps;
• Results can be verified through testing, computation, or formal derivation;
• Researchers themselves have strong tool-usage skills.

In research conversations, computer science tasks accounted for roughly 11.9% of sessions, while engineering and medicine accounted for about 8.4% and 7.7% respectively. When the number of researchers per discipline is factored in, computer science, mathematics, and statistics still show AI usage levels well above their share of the overall research population.

The concentration is even more pronounced in code-assistance sessions. Computer science researchers make up about 4.3% of the global research workforce but contribute roughly 32.7% of research code-assistance conversations. Researchers in AI and machine learning account for only about 1.2% of researchers yet contribute approximately 15.2% of code-assistance sessions.

This indicates that AI usage frequency depends not just on how many people work in a field, but on whether tasks can be converted into digitalized, structured, and verifiable operations.

By contrast, the arts, humanities, parts of the social sciences, and lab- or field-dependent physical sciences show relatively low usage. This does not mean AI is worthless in these areas. Rather, their core work often involves complex contextual judgment, long-term observation, on-site operations, materials handling, or deep interpretation—activities that cannot be fully accomplished through conversation and code.

3. Why Data Analysis Is the Leading Use Case

Data analysis accounts for about 33.7% of general research conversations and as much as 48.1% of code-assistance sessions. This result makes strong sense.

First, research data analysis typically involves substantial repetitive work: checking for missing values, unifying field formats, identifying anomalous records, converting data types, writing statistical scripts, and adjusting visualizations. These tasks are not trivial, but they follow fairly regular patterns—well suited to having AI generate an initial approach that professionals then review and refine.

Second, AI lowers the barrier to using analytical tools. Researchers no longer need to write complex statements from memory; they can describe goals in natural language, such as “compare mean differences between two sample groups and output confidence intervals with visualizations.” The system can generate an analysis approach or sample code, helping researchers quickly reach an executable starting point.

Third, many data analysis problems require cross-disciplinary interpretation. A researcher may understand experimental design but not a particular statistical method, or may be skilled in programming but uncertain about a model’s applicability conditions. AI can provide method explanations, concept comparisons, and risk warnings, shortening knowledge-retrieval time.

That said, AI-generated analysis code is not the same as a correct research conclusion. Researchers should focus their review on:

• Whether the data satisfies the method’s assumptions;
• Whether samples carry selection bias;
• Whether variable definitions are accurate;
• Whether statistical significance is mistaken for practical significance;
• Whether visualizations exaggerate or obscure differences;
• Whether the code contains data leakage, indexing errors, or double counting;
• Whether results can be independently reproduced.

A more robust workflow is therefore: “AI generates the approach → researchers review → code execution → result verification → documented analysis process,” rather than directly copying generated content.

4. Writing and Communication: The Second-Highest-Usage Scenario

Writing and communication tasks account for about 32% of general research conversations—noticeably higher than the share of researchers’ daily working time devoted to them. This shows AI has a clear efficiency advantage in language organization and information expression.

Scientific writing is rarely a matter of simply putting words on a page. It requires multi-level transformations:

• Converting experimental procedures into clear methods descriptions;
• Turning complex results into understandable conclusions;
• Organizing scattered notes into complete sections;
• Compressing specialized content into abstracts, figure captions, or presentation materials;
• Adjusting depth and terminology density for different audiences.

AI is particularly well suited to first-draft organization, structural suggestions, language proofreading, and style conversion. For example, researchers can supply already-confirmed facts and ask the system to reorganize them into a “background—methods—results—limitations” structure. The goal is not to let AI replace the author, but to reduce mechanical text processing and free up more time for research judgment and argumentation.

However, factual accuracy is critical in scientific writing. AI can produce concept confusion, fabricated citations, exaggerated causal claims, and omitted caveats. For papers, grant proposals, and technical reports, the following review mechanisms should be established:

1. All data, conclusions, and citations must be checked against original sources;
2. Unverified generated content must never be treated as experimental fact;
3. Literature summaries produced by the model must be checked item by item;
4. The stages where AI was involved must be clearly documented;
5. For sensitive data, use only organizationally authorized, secure environments.

5. Why Data Collection and Hypothesis Generation See Less Use

Compared with data analysis and writing, data collection tasks account for only about 11.5% of general research conversations and 11.9% of code-assistance sessions—yet researchers actually spend roughly 28.6% of their time on data collection. This shows AI has not yet replaced real-world sampling, experiments, interviews, equipment operation, and field observation at scale.

The main reasons include:

• Data acquisition depends on physical equipment, experimental materials, and site conditions;
• Data quality depends on experimental design and operational rigor;
• Much data involves ethics approval, privacy protection, and access management;
• Automatically generated collection plans may not fit specific environments;
• Anomalies during collection require immediate judgment by professionals.

Research ideation and hypothesis generation are used even less—about 4.2% of general research conversations and 2.1% of code-assistance sessions, below the share of time researchers spend on such work.

This reflects the fact that scientific innovation is not just “generating a few possible directions.” High-quality hypotheses are built on long-term domain expertise, problem awareness, experimental experience, and sensitivity to anomalies. AI can help organize existing literature, propose variable relationships, and list possible mechanisms, but it cannot guarantee that a question is original, testable, or scientifically valuable.

AI is therefore better suited to assist after a hypothesis takes shape, for example by:

• Checking whether the hypothesis is sufficiently well-defined;
• Mapping potentially relevant variables;
• Designing control groups and validation pathways;
• Flagging potential confounders;
• Comparing the evidence requirements of competing explanations.

6. How Research Teams Should Design AI Workflows

To improve efficiency while controlling risk, research AI workflows can be divided into five stages.

1. Define task boundaries

First determine whether a task involves information organization, coding assistance, or language expression—or core scientific judgment. The closer a task gets to experimental decisions, medical judgments, ethics review, and final conclusions, the less generated output should be relied on directly.

2. Prepare structured inputs

Provide the system with clear objectives, data field descriptions, constraints, and output formats. The vaguer the input, the more likely the output is ambiguous. For sensitive data, anonymize first, reduce fields, or test workflows with simulated data.

3. Demand explainable processes

Don’t just ask for answers—require the system to explain the methods used, underlying assumptions, possible failure modes, and steps that need human review. This helps researchers uncover hidden errors.

4. Establish independent verification

Run and test code, cross-check statistical results with alternative methods, and verify written content against original sources. For important conclusions, keep version records and review trails to ensure traceability.

5. Create reproducible records

Research projects should document when and where AI was involved, the task, input types, human modifications, and the final adopted results. These records need not contain sensitive original text, but they should be sufficient to show what role automated tools played in the research process.

7. From “Replacing Researchers” to “Augmenting Research Capability”

Current usage data shows that generative AI’s most mature value is not performing research independently, but reducing repetitive labor, lowering the cost of expression, accelerating code writing, and improving knowledge organization.

For researchers, sensible positioning includes:

• Treating AI as a discussion partner for analysis approaches;
• Using it as a first-draft generator for code and text;
• Using it as an aid for spotting omissions and checking logic;
• Using it as a translation layer for cross-disciplinary communication.

For research institutions, the priority is not maximizing usage counts but establishing systems for data security, attribution, process documentation, result review, and accountability. Only when AI outputs can be verified, traced, and reproduced will efficiency gains not turn into new research risks.

Overall, AI in research is settling into a clear division of labor: data analysis and writing are currently the easiest entry points, code assistance is a key growth area for digitally oriented disciplines, while data collection, experiment execution, and original hypothesis generation still depend heavily on human expertise. In the future, truly valuable research AI will not make all decisions for researchers—it will help them process information faster, verify results more rigorously, and devote more energy to genuinely innovative scientific questions.