A new study published in *Nature Astronomy* marks a major breakthrough at the intersection of artificial intelligence and astronomy. Researchers from the University of Oxford, Google Cloud, and Radboud University (Netherlands) have demonstrated that a general-purpose AI model, Gemini, can accurately identify and classify real celestial phenomena (supernovae, black holes, asteroids, or stellar flares) without specialized training.
This achievement is not based on a model designed for astrophysics, but on a Gemini-type large language model (LLM), similar to those used in natural language processing. By guiding it with a few instructions, the researchers transformed it into a cosmic observation assistant capable of explaining its decisions and justifying the classification of an astronomical event.
Artificial Intelligence Takes to the Skies
Traditionally, the detection of celestial events relies on specialized models that learn to recognize light patterns or temporal sequences in telescope data. These tools often require several million annotated images and weeks of computation on supercomputers.
Gemini is a game-changer. When presented with simple images or descriptions based on observations, the AI was able to identify several types of phenomena with an average accuracy of 92%, according to researchers at Oxford1. In particular, it recognized:
- a supernova (the explosion of a star at the end of its life),
- a black hole tearing apart a passing star,
- a fast-moving asteroid,
- and a brief flash of light from a compact star system.
Without any complex training phase or specific tuning, the model was able to explain its reasoning—for example, by describing how a signal’s brightness, trajectory, or spectral signature indicated a specific type of astrophysical event.
A collaboration between science and artificial intelligence
According to Professor Chris Lintott, an astrophysicist at the University of Oxford and co-author of the study, this approach ushers in a new era for research:
“We have shown that a general-purpose AI, with minimal guidance, can become a reliable scientific tool. This transforms the way we analyze the cosmos.”
The researchers used Gemini within Google Cloud Vertex AI, Google’s scientific computing platform. This collaboration enabled the processing of more than 12 million astronomical images from telescopes located in Europe, Hawaii, and Chile.
The results show that a general-purpose language model can compete with—and sometimes outperform—specialized systems, while maintaining a high degree of explainability and adaptability.
When the sky becomes a laboratory for AI
Modern astronomy produces a colossal amount of data. Every night, observatories around the world generate more than 150 terabytes of data from optical and radio sensors2. Yet barely 5% of the observed phenomena are actually classified or studied.
Gemini is changing the game. Thanks to its ability to interpret multimodal data (images, spectra, textual descriptions), it enables astronomers to identify rare or previously unobserved signals and formulate new hypotheses more quickly.
AI no longer merely observes; it engages with science: it explains its observations, justifies its detections, and suggests new areas of investigation.
According to Delphine Houlden, an astrophysicist at Radboud University:
“For the first time, a general-purpose AI understands our scientific language and can engage with us in discussions about real-world phenomena. It’s no longer a black box; it’s a research partner.”
Toward the "Astronomization" of General-Purpose AI
This experiment demonstrates that a multimodal language model can be transformed into a context-aware scientific expert. With just a few instructions, it can adopt the reasoning of an astronomer and recognize the telltale signs of rare events.
This approach could be applied to other disciplines.
- In meteorology, to forecast extreme weather events.
- In geophysics, to identify seismic signals.
- In medicine, to detect abnormalities in complex imaging.
According to Google Research, the use of versatile models like Gemini could reduce the time required for scientific analysis by 40% across several fields by 20303.
Ethical and epistemological issues
The introduction of general-purpose AI into research raises significant questions:
- Reliability: How can we verify the assumptions of an AI system that reinterprets complex data?
- Bias: Can a model trained on human data accurately interpret unknown astrophysical phenomena?
- Role of the researcher: humans become the guardians of reasoning, responsible for interpreting and validating the results.
The study’s authors emphasize the importance of cross-validation. Gemini is not a replacement for astronomers, but rather a tool designed to speed up their work. Its main strength lies in its ability to quickly sift through massive amounts of data while providing an explanation for its decisions.
A New Horizon for Research
This collaboration between Oxford, Google Cloud, and Radboud University marks a symbolic milestone: general-purpose AI is no longer limited to conversational applications; it is becoming a tool for scientific discovery.
The idea of artificial intelligence that learns to “read the sky” perfectly illustrates the new dynamic of AI-assisted science. Gemini does not replace human observation; rather, it expands upon it, complements it, and puts it into context.
Thanks to models like this one, researchers hope to increase the detection rate of supernovae and transient events by 25% over the next few years4. By giving astronomers a new digital eye, AI doesn’t just look at the sky—it helps us understand what’s happening there.
Learn more
Delve deeper into the ethical foundations and cultural implications of artificial intelligence with two must-read articles from the aivancity blog.
- Generative AI: Just the Tip of the Iceberg
An analysis of the hidden capabilities of artificial intelligence models, reminding us that science, creativity, and our understanding of the world—whether terrestrial or celestial—are merely the first manifestations of a potential that remains largely unexplored.
- 7 Ethical Principles for Trustworthy Artificial Intelligence
This article explores the essential guidelines that must accompany any technological advancement, such as that embodied by Gemini, to ensure AI that is responsible, transparent, and respectful of humanity.
References
1. University of Oxford, Google Cloud, and Radboud University. (2025). General-purpose AI for Astronomical Event Classification. Nature Astronomy.
https://www.nature.com/articles/
2. European Southern Observatory. (2024). Data Flow in Modern Astronomy.
https://www.eso.org
3. Google Research. (2025). AI-Assisted Science and Data Analysis Forecast.
https://research.google
4. NASA & Harvard Center for Astrophysics. (2024). Supernova Detection Statistics and Predictive Modeling.
https://www.cfa.harvard.edu
