Astronomers have witnessed a groundbreaking shift in their research thanks to AI. A study involving Oxford University, Google Cloud, and Radboud University revealed how a general-purpose large language model (LLM), Google’s Gemini, can transform into a top-tier astronomy assistant with minimal training.
With just 15 images and a basic set of guidelines, Gemini accurately identified cosmic events versus imaging artefacts 93% of the time. Additionally, the AI provided clear, human-readable explanations for each classification. This leap is a significant step toward making AI-driven science more transparent and democratizing scientific tools without requiring large datasets or advanced AI knowledge.
Turan Bulmus, co-lead author from Google Cloud, shared his enthusiasm: This research proves that general-purpose LLMs can revolutionize scientific discovery by enabling people without formal training to engage in meaningful contributions. It highlights how accessible AI breaks down the barriers in research fields, opening the doors for non-experts to make impactful discoveries.
The Noise of the Universe and AI’s Role
Modern telescopes are constantly scanning the skies, generating millions of alerts nightly. While some of these signals, like exploding stars, are genuine, most are false alarms caused by satellite interference or cosmic ray hits.
In the past, astronomers relied on specialized machine learning models to sort through this data. However, these models often function as “black boxes,” offering no insight into how they arrive at their decisions. This lack of transparency leaves scientists with two options: trust the results blindly or spend excessive time manually reviewing thousands of potential signals. The upcoming Vera C. Rubin Observatory, which will process around 20 terabytes of data daily, will make manual review impossible.
The research team posed a crucial question: Can a general-purpose AI, such as Gemini, provide both high accuracy and a clear explanation of its reasoning?
Teaching Gemini to Identify Cosmic Events
The team provided Gemini with 15 labeled examples from three major sky surveys: ATLAS, MeerLICHT, and Pan-STARRS. Each example contained a small image of a new alert, a reference image, a “difference” image highlighting the change, and an expert note. These few-shot examples, along with clear instructions, enabled the model to classify new alerts, offering labels (real or bogus), priority scores, and explanations in simple language.
AI’s Self-Assessment and Human Collaboration
A critical element of the study was evaluating the AI’s ability to explain its decisions. The researchers formed a group of 12 astronomers who reviewed the AI’s explanations and found them to be coherent and insightful.
In a parallel experiment, Gemini assessed its own responses, assigning a “coherence score” to each one. The team discovered that the AI’s confidence level directly correlated with accuracy. Low-coherence responses were more likely to be incorrect, highlighting the importance of self-assessment in refining AI-driven decisions. This self-correction feature is vital for establishing a “human-in-the-loop” approach. It allows the system to flag uncertain results, directing human experts to the most promising leads. The researchers were able to enhance the AI’s performance on a dataset, increasing accuracy from approximately 93.4% to 96.7% by using this feedback loop.
A Glimpse Into the Future of AI in Science
The team envisions AI-powered systems as autonomous agents in scientific research. These “agentic assistants” would not only classify images but could also integrate various data types, evaluate their own confidence, and autonomously request additional observations from robotic telescopes. Most importantly, they would escalate only the most significant discoveries to human scientists.
Given that this method relies on a small set of examples and clear instructions, it can quickly adapt to new scientific instruments and research across diverse fields.
Turan Bulmus concluded, “We are entering a new era of scientific discovery, where transparent AI partners—not opaque algorithms—accelerate progress. This research paves the way for systems that learn with us, explain their logic, and empower researchers across all fields to focus on the essential questions.”
