OpenAI introduces GABRIEL, an AI tool

OpenAI has introduced GABRIEL, an open-source library that uses GPT to turn difficult-to-analyze text and images into measurable data. The tool is designed for economists, social researchers, and analysts who need to study thousands or millions of documents.

Much of the information about society exists in qualitative formats: interviews, syllabi, academic papers, social media posts, photographs, or customer reviews. The problem is that reviewing and classifying all that material by hand can take months or become practically impossible.

GABRIEL aims to address that bottleneck. The researcher describes in everyday language what they want to measure, such as whether a job posting appears compatible with family life. The tool then applies that criterion consistently across a large collection of documents and returns a score for each one.

The result does not replace the researcher’s work. It allows them to spend less time labeling repetitive data and more time deciding what to measure, checking whether the measurements are reliable, and drawing careful conclusions.

What GABRIEL can do

The tool is designed to work with information that normally does not come organized in rows and columns. Possible uses include:

• Analyzing academic papers to identify which methods they use and how those methods change over time.
• Reviewing syllabi and measuring how much space is devoted to different subjects or skills.
• Extracting historical data from small towns across Europe.
• Studying customer reviews to identify which aspects people value most.
• Classifying text excerpts and locating patterns in large collections of documents.
• Removing personal information from text to protect privacy.

It also includes practical features for combining databases even when their columns do not match, intelligently removing duplicate records, and helping propose new lines of research.

Accuracy still requires oversight

OpenAI says that, in a benchmark study included with the launch, GPT showed high accuracy when labeling qualitative data across different use cases. That does not mean the scores can be accepted without review.

A measurement depends on how the question is phrased, the examples used to define it, and the context of the documents. A system can classify thousands of texts in a few minutes, but researchers still need to validate the results and identify errors or bias.

That point matters: GABRIEL automates a costly part of the process. It does not automatically turn an interpretation into a scientific fact.

Available as a Python library

GABRIEL can now be downloaded as an open-source Python library and includes a tutorial notebook to help users get started. OpenAI says it is designed to require limited technical expertise and that it will continue refining the tool based on input from the academic community.

For you, the most relevant change is the type of information that can now enter an analysis. A collection of interviews, photographs, or comments that was previously too large to study manually can become a database with comparable scores.

The tool does not eliminate the need for human judgment. It reduces the mechanical work that prevented researchers from investigating many questions and leaves open an issue that will remain central: how to verify that the categories and measurements created with AI truly represent what people intended to study.