Tired of “black box” algorithms in science? Tune in to hear about a remarkable AI breakthrough co-led by the University of Oxford and Google Cloud.

Modern telescopes scan the sky relentlessly, generating millions of alerts every night about potential changes. While some of these alerts are genuine discoveries, such as an exploding star, a black hole tearing apart a passing star, or a fast-moving asteroid, the vast majority are “bogus” signals caused by things like satellite trails or instrumental artefacts. Traditionally, astronomers have relied on specialized machine learning models to filter this data. However, these systems often operate like a “black box,” providing a simple label without explaining their logic. This forces scientists to spend countless hours manually verifying candidates—a task that will become impossible with the next generation of telescopes.

A new study demonstrates that a general-purpose large language model (LLM)—Google’s Gemini—can be transformed into an expert astronomy assistant with minimal guidance. The research team provided the multimodal AI with just 15 labeled examples and concise instructions. Guided by these few-shot examples, Gemini learned to distinguish real cosmic events from imaging artefacts with approximately 93% accuracy.

This approach is considered a total game changer for the field because of its transparency. Crucially, the AI provided a plain-English explanation for every classification, moving away from traditional, opaque systems. Astronomers reviewing the AI’s descriptions rated them as highly coherent and useful.

This accessibility shows how general-purpose LLMs can democratize scientific discovery, empowering anyone with curiosity to contribute meaningfully, even without a deep expertise in AI programming or formal astronomy training.

Furthermore, the system is reliable because it knows when to ask for help. The model reviews its own answers and assigns a “coherence score,” demonstrating a self-assessment capability that is critical for building a reliable “human-in-the-loop” workflow. By automatically flagging its own uncertain cases for human review, the system focuses astronomers' attention where it is most needed. Using this self-correction loop, the team improved the model's performance on one dataset from about 93.4% to about 96.7%.

The team envisions this technology as the foundation for autonomous “agentic assistants” that could integrate multiple data sources, autonomously request follow-up observations from robotic telescopes, and escalate only the most promising discoveries to human scientists. This work shows a path toward transparent AI partners that accelerate scientific discovery.