How Researchers Use Generative AI: Data Analysis and Writing Lead Adoption

1. Generative AI in Research Is Moving from Experimentation to Workflow Collaboration

After generative AI entered research settings, initial attention focused on whether it could propose novel hypotheses, independently complete experimental design, or even replace parts of research work. In practice, however, researchers tend to place AI in supporting roles within existing workflows rather than letting it determine research direction directly.

A study of research usage analyzed about 207,000 research-related sessions, including general conversations, collaborative work sessions, and code-assistance sessions. The researchers retained only labels such as discipline, research subfield, and task category. They did not extract specific research data, experimental protocols, unpublished results, or draft papers, and results were presented in aggregate form.

The results show that researchers most often use generative AI for two types of work:

1. Cleaning, transforming, analyzing, and visualizing existing data;
2. Helping write, revise, summarize, and organize research text.

This indicates that the practical role of AI in research has shifted from “replacing how scientists think” toward “improving researchers’ information-processing efficiency.”

2. Usage Varies Significantly by Discipline

The frequency of AI use in research is not evenly distributed. Computer science, mathematics, statistics, and some engineering fields are more active, while healthcare, experimental physics, and other fields that rely heavily on physical experiments remain more cautious overall.

In the disciplinary distribution of research sessions, computer science accounts for about 11.9%, engineering about 8.4%, and medicine about 7.7%. When the overall population of researchers in each field is taken into account, the intensity of AI use in computer science, mathematics, and statistics remains clearly higher.

Code-assistance sessions are even more concentrated. Computer science researchers make up about 4.3% of the global researcher population but contribute 32.7% of code-assistance research sessions. AI and machine learning researchers make up about 1.2% of global researchers but contribute about 5% of overall research sessions and 15.2% of code-assistance sessions.

This variation is driven by several factors:

2.1 Different Degrees of Task Digitization

Research objects in computer science, mathematics, and statistics are often based on code, data, formulas, and text, which are naturally suited to digital tools. Researchers can directly provide data structures, algorithm descriptions, or code snippets and ask AI to participate in analysis and modification.

By contrast, experimental sciences often depend on instrument operation, sample preparation, field observation, and control of experimental conditions. These steps cannot be completed through text interaction alone, so the parts AI can directly address are relatively limited.

2.2 Different Feedback Speeds

In program debugging, statistical computation, and text revision, researchers can usually judge quickly whether AI output is useful. For example, whether code runs, whether a chart is correct, or whether a formula satisfies constraints can often be verified quickly through tools or manual checks.

In biological experiments, materials experiments, or clinical research, however, conclusions may take days, weeks, or longer to verify, and the cost of incorrect advice is higher. Users therefore tend to be more cautious.

2.3 Different Data and Compliance Requirements

Some research fields involve patient information, unpublished experimental data, sensitive samples, or intellectual property. Researchers cannot simply copy raw materials into external systems; they need de-identification, access control, and security review, which also reduces direct AI usage.

3. Data Analysis Is the Most Mature Research Use Case

The data shows that data analysis accounts for about 33.7% of general research sessions and about 48.1% of code-assistance sessions. For reference, researchers typically spend about 27.6% of their working time on data analysis. This means AI use in data analysis is already higher than the share of time this task occupies in traditional research work.

Research data analysis usually includes multiple steps:

• Checking field types, missing values, and outliers;
• Designing data cleaning and transformation workflows;
• Writing statistical analysis or machine learning code;
• Generating charts and interpreting results;
• Refactoring and optimizing the analysis process;
• Helping locate program errors;
• Organizing analysis results into Methods and Results sections.

AI is best suited to work that is repetitive, relatively rule-based, and verifiable through results. For example, researchers can ask the system to generate an initial data-cleaning script based on a data dictionary, and then review variable meanings, boundary conditions, and exception-handling logic themselves.

However, AI-generated analysis code is not equivalent to correct research conclusions. Common risks include:

1. Misunderstanding variable meanings;
2. Confusing correlation with causation;
3. Choosing inappropriate statistical methods;
4. Ignoring sample bias and data leakage;
5. Producing logical errors even when the code runs;
6. Generating plausible but non-reproducible charts.

A more reliable workflow should therefore be:

“Data dictionary confirmation — analysis objective definition — code generation — local execution — result validation — human interpretation — version logging.”

AI can improve execution efficiency, but researchers remain responsible for the research question, method selection, and final conclusions.

4. Writing and Communication Are the Second Most Frequent Use Case

Writing and communication tasks account for about 32% of general research sessions, noticeably higher than the roughly 16% of time researchers spend on such work. This gap suggests that researchers may view generative AI as an efficient language-processing and structuring tool.

Common uses in research writing include:

• Outlining paper structure;
• Turning scattered notes into an outline;
• Revising grammar, expression, and paragraph transitions;
• Condensing abstracts or expanding method descriptions;
• Rewriting complex content for different audiences;
• Helping organize meeting notes and project updates;
• Generating figure captions based on existing results;
• Checking terminology consistency.

The advantage of AI in these tasks is not creating unverified scientific facts for researchers, but reducing the cost of expression. Researchers can devote more effort to research design, evidence evaluation, and argumentation.

However, research writing must guard against three types of problems:

4.1 Content Distortion

The system may add information not provided in the original text to make the article flow better, causing research conclusions to be overstated. All content involving sample sizes, experimental conditions, statistical significance, and study limitations should be checked item by item.

4.2 Citation Errors

Generative AI may provide nonexistent papers, incorrect author information, or inaccurate research conclusions. References must be verified through formal databases or original literature and should not be adopted directly from system-generated citations.

4.3 Blurred Responsibility Boundaries

Authors of research papers must be responsible for the content. AI can serve as an auxiliary tool, but it cannot replace authors’ review of data authenticity, methodological soundness, and conclusion reliability.

5. Data Collection and Research Ideation Remain Lower

Compared with data analysis and writing, data collection accounts for about 28.6% of researchers’ working time but only 11.5% of general AI sessions and about 11.9% of code-assistance sessions.

Research ideation and hypothesis generation account for an even smaller share: about 4.2% of general sessions and about 2.1% of code-assistance sessions, while researchers spend about 7.1% of their time on such work.

This does not mean AI cannot help researchers generate ideas. Rather, it indicates that these tasks involve higher uncertainty and responsibility risk.

Research hypotheses usually require combining:

• Domain knowledge;
• Existing literature;
• Real-world problems;
• Testability;
• Research ethics;
• Resource and time constraints;
• The ability to explain anomalous results.

AI can quickly generate many seemingly reasonable directions, but “seemingly reasonable” does not equal “scientifically valuable.” Without literature verification and experimental validation, automatically generated hypotheses may simply be recombinations of existing views or may even contain factual errors.

A more appropriate approach is to use AI for divergence and organization during hypothesis development rather than directly accepting its conclusions. For example, researchers can ask the system to list multiple possible mechanisms, judge which ones are testable, and then screen them through literature search and experimental design.

6. Practical Recommendations for Research Teams

6.1 Define Usage Boundaries by Task Risk

Low-risk tasks can be handled by AI first, such as format conversion, text polishing, code comments, and meeting notes. Medium-risk tasks include statistical analysis, data cleaning, and chart interpretation, which require human review. High-risk tasks include clinical judgment, experimental conclusions, research ethics, and final paper conclusions, which must be independently confirmed by qualified professionals.

6.2 Establish a Data Classification Mechanism

Research teams should classify data into public data, internal data, sensitive data, and data prohibited from external transmission. For internal or sensitive data, teams should prefer de-identified samples, field mappings, or summary information, and avoid directly entering content that can identify individuals, institutions, or unpublished results.

6.3 Keep Process Records

Teams should record which stages AI participated in, what inputs were used, what code or text was generated, and who reviewed the final output. For important research projects, teams should also retain prompts, output versions, revision records, and validation results for later reproduction and audit.

6.4 Adopt “Machine First Draft, Expert Final Review”

AI is responsible for increasing the speed of initial output, while researchers are responsible for fact verification, method selection, logical judgment, and final sign-off. This model can leverage automation while avoiding transferring research responsibility to an unexplainable system.

6.5 Make Reproducibility a Core Metric

Whether for code, statistical results, or paper text, other researchers should be able to reproduce the results using the same data, parameters, and steps. AI-generated content must be included in version management and testing workflows rather than remaining in one-off conversations.

7. Conclusion

From the actual usage structure, the most valuable directions for generative AI in research remain data analysis, code assistance, text processing, and information communication. It is good at accelerating execution, organizing complex information, and lowering the barrier to expression. But in research hypotheses, experimental design, and final conclusions, professional knowledge, empirical validation, and strict responsibility mechanisms are still required.

In the future, competition in research AI will not be only about whether models can generate longer answers. It will be about whether they can integrate into real research workflows and support access control, process logging, result reproducibility, and human review. For research teams, the most reliable path is not to pursue full automation blindly, but to start with verifiable, reversible, low-risk tasks and gradually build a human-AI collaborative research workflow.