How 1,800 Scientists Use ChatGPT: Insights into Research Workflows and Productivity

An OpenAI survey of 1,800 scientists reveals ChatGPT as a high-frequency general assistant in research workflows.
OpenAI surveyed over 1,800 scientists to examine how ChatGPT is actually used in research, focusing on daily workflows, self-reported productivity, and expectations for AI-assisted discovery. ChatGPT is mainly used for fragmented, high-frequency tasks like literature review, code debugging, and text polishing—more a general assistant than a domain expert. The survey shifts the discussion from speculation to observation, showing ChatGPT holds a stable but modest place in research workflows.
OpenAI recently released a survey of over 1,800 scientists, attempting to answer a specific question: in real research scenarios, what exactly is ChatGPT being used for? This survey doesn't dwell on the abstract debate of "whether AI can do science," but instead focuses on daily workflows, self-reported productivity changes among researchers, and expectations for AI-assisted scientific discovery.

From "Can It Be Used" to "How Is It Used"
Over the past two years, most discussions about AI and scientific research have centered on capability boundaries—whether models can read papers, generate hypotheses, or design experiments. This survey shifts the perspective: it cares about usage behaviors that have already occurred. A sample of over 1,800 scientists is large enough to sketch a usage picture with reference value.
The survey focuses on three dimensions: usage in daily workflows, productivity changes perceived by researchers, and expectations for AI's involvement in scientific discovery. These three dimensions correspond to three different time scales: current habits, short-term efficiency perceptions, and long-term imagination. Only by looking at them together can we understand the research community's true attitude toward ChatGPT.
In the research context, a survey is an empirical method for systematically collecting data, typically obtaining behaviors and attitudes of a sample group through questionnaires, interviews, or log analysis. The survey mentioned in this article covers over 1,800 scientists and is a large-sample cross-sectional study. Its value lies in describing the status quo in a quantifiable way, rather than inferring trends from individual cases or anecdotes. Unlike randomized controlled trials, surveys cannot directly establish causality, but they can reveal correlations and distribution characteristics among variables. In AI application research, such surveys are often used to capture self-reported behavior, with the limitation that they may be affected by social desirability bias—that is, respondents tend to report usage patterns that align with expectations. Therefore, when reading such conclusions, one must distinguish between "what people say they did" and "what system logs show they did."
AI's Place in Daily Research Workflows
Judging from the survey's focus, ChatGPT's role in research is closer to a "general assistant" than a "domain expert." Research workflows contain many low-creativity but time-consuming tasks—literature review, code debugging, text polishing, organizing ideas, and quickly getting up to speed on interdisciplinary concepts. These are exactly the areas where conversational AI excels.
This also explains why the survey emphasizes "daily workflows." Scientists don't hand an entire paper to the model and ask it to produce conclusions; instead, they frequently invoke it for specific small tasks. This fragmented, high-frequency usage pattern stands in sharp contrast to many narratives claiming "AI will replace researchers." The value of the tool lies in compressing the time that doesn't produce core insights, freeing human attention for judgment and design.
A workflow refers to the series of ordered stages in research activities from problem formulation to output, including literature review, hypothesis formation, experimental design, data collection and analysis, code implementation, paper writing, and peer review. Conversational AI is typically embedded into workflows at the task level rather than the process level—that is, scientists invoke the model at a specific stage to complete short-cycle tasks, then integrate the results back into the overall process. This fragmented usage pattern is fundamentally different from end-to-end automation: the former preserves human control over key decisions, while the latter attempts to replace the entire process. Understanding this distinction helps assess AI's true role in research—it is more like an auxiliary tool embedded in existing workflows than an agent independently executing research processes.
Productivity Perception and Expectations for Scientific Discovery
The "reported productivity" covered in the survey deserves separate discussion. Self-reported productivity changes among researchers are a subjective indicator, but they reflect real feelings in the user experience rather than scores from benchmarks. In research—a long-cycle, high-uncertainty endeavor—efficiency gains are often difficult to quantify precisely, so subjective perception becomes a meaningful window of observation.
The focus on "expectations for discovery" points further ahead. Scientists' expectations of AI go beyond saving time; they also include whether it can provide clues that humans easily overlook in hypothesis generation, cross-domain association, and data pattern recognition. Such expectations are currently more directional, and the survey itself presents them as expectations rather than established facts.
Self-reported productivity is a subjective measurement commonly used in organizational behavior and information systems research, typically employing Likert scales to have users rate "the degree of efficiency improvement after using a tool." Its advantage is capturing experience changes in tasks that are difficult to quantify objectively (such as organizing ideas or understanding literature), but its drawbacks are also obvious: subjective feelings may be influenced by novelty, expectation effects, or selective memory. In research settings, objective productivity indicators (such as paper output or experiment cycles) are affected by multiple factors and difficult to attribute to a single tool, making subjective indicators a compromise. Academia generally recommends combining self-reported data with behavioral logs and controlled experiments to improve reliability through triangulation. This survey relies solely on self-reported data, so conclusions should be understood as a mapping of user experience rather than a strict measurement of efficiency.
A Signal Worth Noting
Perhaps the most valuable aspect of this survey is not any specific conclusion, but that it pulls the baseline of discussion from speculation back to observation. Usage data from over 1,800 scientists shows that ChatGPT has already secured a stable but modest position in research workflows: it is used frequently, applied to specific tasks, expected to take on more, but not yet widely regarded as a core decision-maker in the discovery process.
For readers interested in the real-world adoption of AI, the research setting is an excellent sample for observation—its users are highly specialized, sensitive to errors, and rigorous in judging a tool's actual value. How they weigh the boundaries of AI use in their daily work offers direct reference for other knowledge-intensive industries.
One-line Summary
An OpenAI survey of 1,800 scientists reveals a real-world picture of ChatGPT as a high-frequency general assistant in research workflows.
Paragraph Summary
OpenAI released a survey covering over 1,800 scientists, focusing on ChatGPT's usage behavior in real research scenarios rather than abstract discussions of capability boundaries. The survey examines three dimensions: usage in daily research workflows, self-reported productivity changes among researchers, and expectations for AI's involvement in scientific discovery. Results show that ChatGPT is mainly used for fragmented, high-frequency tasks such as literature review, code debugging, text polishing, and organizing ideas, with a role closer to a general assistant than a domain expert. Although self-reported productivity gains are subjective indicators, they reflect real user experience. Scientists' expectations of AI extend to discovery stages such as hypothesis generation and cross-domain association, but these remain directional expectations for now. The survey's core value lies in pulling the baseline of discussion from speculation back to observation, showing that ChatGPT has secured a stable but modest position in research workflows.
Related articles

Why You Shouldn't Run the Hermes Agent on Your Personal Computer
Why shouldn't you run autonomous AI agents like Hermes on your personal computer? This article analyzes why agents should be deployed to a dedicated VPS instead of your primary device, from four angles: security risks, resource consumption, credential leakage, and autonomy requirements—plus the practical logic of isolated deployment.

Ignite Keynote: Building a Frontier Ecosystem to Bring AI to Life
Microsoft CEO Nadella's Ignite keynote proposes a frontier ecosystem to turn AI advances into real-world impact. This article unpacks the strategy's ecosystem layers and deployment signals.
C++26 Standard Finalized: Key Takeaway…
C++26 Standard Finalized: Key Takeaways from the ISO C++ Meeting
C++26 wraps up at the ISO C++ meeting. This article explains the trip report's implications, the standardization process, and impact on developers and compiler ecosystems.