What Scientists Actually Use AI For: Data Analysis and Writing Dominate LLM Usage in Research

A Landmark Study: What Scientists Actually Do with AI

Large language models have been present in scientific research for some time, but the question of “what scientists actually use AI for” has long lacked large-scale, quantifiable data. A recent report from an AI research organization provides the first comprehensive answer, based on an anonymized analysis of approximately 207,000 research-related conversational sessions.

The study spanned roughly two months and included about 129,000 assistant-style conversations and 78,000 programming-focused conversations. What makes the methodology notable is its restraint: the system retained only session-level metadata—discipline, subfield, and task classification labels—while extracting and storing no research data, hypotheses, experimental methods, or unpublished findings. All results were aggregated. This privacy-preserving design of “metadata only, no content access” is a model worth emulating for any technical team conducting user behavior analysis.

Discipline Distribution: Quantitative Fields Are Heavy Users

The data reveals a clear disciplinary skew in AI adoption within research:

• Computer science leads, accounting for 11.9% of dialogue-style research conversations, followed by engineering (8.4%) and medicine (7.7%).
• Programming conversations are even more concentrated: CS researchers represent just 4.3% of the global researcher population yet contribute 32.7% of all programming-related conversations.
• The AI/ML subfield is particularly notable: its practitioners make up only 1.2% of researchers but account for 5% of dialogue conversations and 15.2% of programming conversations.

Even after normalizing for researcher headcount across disciplines, computer science, mathematics, and statistics still show usage intensity significantly exceeding their population share. Conversely, the arts, humanities, and parts of the social sciences show markedly lower usage.

The logic behind this distribution is not hard to grasp: the match between research tasks and AI capability boundaries matters more than population size in determining adoption depth. Quantitative disciplines naturally work close to text and code—writing scripts, processing data, debugging models, drafting methodology sections—precisely the strengths of large models. In contrast, laboratory-intensive physical sciences and hands-on healthcare fields offer limited touchpoints for AI intervention. The core work of humanities and social sciences (fieldwork, close reading of primary texts, philosophical argumentation) remains difficult to replace with general-purpose models.

Task Structure: AI Concentrates at Two Extremes, Missing the Middle

This is the report’s most valuable finding. When broken down by research workflow stage, AI participation shows extreme imbalance:

Task Type	Share of Researchers’ Time	Share of Dialogue Conversations	Share of Programming Conversations
Data analysis	27.6%	33.7%	48.1%
Writing & communication	16.0%	32.0%	—
Data collection	28.6%	11.5%	11.9%
Research ideation & hypothesis generation	7.1%	4.2%	2.1%

Key interpretations:

1. Data analysis is the most “amplified” task by AI. Researchers spend about a quarter of their time on data analysis, yet it accounts for nearly half of all programming conversations. Writing data-cleaning scripts, running statistical tests, and generating visualizations are currently the smoothest integration points between large models and research—clear inputs and outputs, verifiable results, and low cost of trial and error.

2. Writing tasks account for nearly double their actual time share. 32% of conversations go to writing and communication, while researchers spend only 16% of their time on this activity. Paper polishing, responding to reviewer comments, and drafting grant proposals—high-frequency, formulaic, language-intensive work—fall squarely within the comfort zone of large models.

3. Data collection and research ideation are significantly bypassed. Data collection consumes nearly 30% of researchers’ time, yet AI conversations account for only about 11%; hypothesis generation drops even lower, to 2–4%. The reasons include: data acquisition often involves permissions, compliance requirements, and specialized systems that general-purpose models cannot access; and proposing original hypotheses is the core creative step of research, where researchers neither trust AI nor want to risk academic reputation.

Additionally, a survey of 528 researchers in life sciences and health sciences aligned closely with the behavioral session data, suggesting that observed behavior matches self-reported perceptions and strengthening the robustness of the conclusions.

Implications for Research Teams

Several recommendations emerge from these findings, from both engineering and management perspectives:

1. Prioritize integrating AI into data analysis and writing pipelines. These are currently the highest-ROI scenarios. A typical deployment pattern links “data → code → charts → text” into a single flow—for example, having models assist in generating reproducible analysis scripts:

# Example: Typical interaction pattern for AI-assisted statistical analysis script generation
prompt = """
Here is the dataset field description and cleaning rules.
Generate Python code that: handles missing values,
performs group-wise statistical tests,
and outputs standardized charts.
Requirements: code must be reproducible, with statistical test explanations.
"""

2. Build dedicated channels for the data collection stage. The low AI participation in data collection stems fundamentally from a lack of compliant, trustworthy data access. Research institutions should consider standardizing data acquisition workflows using cloud-based object storage, data integration services, and permission management frameworks—then layer AI capabilities on top of that standardized layer, rather than having AI directly touch raw sensitive data.

3. Stay cautious but open toward hypothesis generation. Although current usage is extremely low, AI’s potential as a supporting tool—analogical reasoning based on literature reviews, serving as a “hypothesis brainstorm” companion—remains promising. The key is establishing a clear human-AI division of labor: AI handles divergence and blind-spot coverage, while humans make the final judgments and convergence decisions.

4. Adopt the privacy-preserving research methodology as a baseline. The design principle of “retaining only classification labels, never touching raw content, and presenting only aggregated results” should be the minimum standard for any institution conducting AI usage analysis—especially when highly sensitive data like unpublished research is involved.

Conclusion

The core fact revealed by this study is that AI penetration in research is decidedly uneven: it concentrates in language-intensive, code-heavy, and verifiable stages (data analysis, writing), while showing significant gaps in data acquisition and the creative core of science. For technical teams, this represents a substantial opportunity—whoever can connect the full chain of “data access → automated analysis → writing assistance” will gain a decisive advantage in the next phase of AI-assisted research. And as model capabilities continue to integrate with domain-specific tools, even the high walls around research ideation and hypothesis generation may gradually be scaled.