rationale:

Current State and Strict Criteria The current count of qualifying discoveries is highly likely to be zero, driven by the strict requirement that discoveries must be genuinely new classes rather than subtypes, and must be explicitly claimed as such in a refereed paper. While anomaly-detection tools have found thousands of unusual objects, these have mostly failed to clear the bar. ESA's AnomalyMatch scanned roughly 100 million Hubble images, finding over 1,300 anomalies, but these are individual objects, not classes Researchers discover hundreds of cosmic anomalies with ... https://www.aanda.org/articles/aa/full_html/2025/12/aa55512-25/aa55512-25.html. The AI-flagged SN 2023zkd is an excluded supernova subtype AI Helps Astronomers Discover a New Type of Supernova https://arxiv.org/html/2502.19469. Even the strongest borderline case, SAURON, which was found without human targeting in MeerKAT data and called a "new class of radio galaxy" Making new discoveries with AI - ML Enabling New Discoveries with Machine Learning, could easily be ruled an excluded subtype. Hedged claims like those for Ansky/AT2020ukj A gallery of overlooked optical nuclear transients from the ... and XRT 200515 AI Discovers Hidden Extragalactic X-ray Flash in Chandra's ... may also fail to secure a definitive new-class designation.

The Case for Growth Despite this weak starting point, the opportunity for AI-driven discovery will expand considerably before 2030. The Rubin Observatory's ten-year LSST survey began in mid-2026 and already issues about 800,000 alerts a night, relying heavily on machine-learning brokers for triage and candidate selection NSF–DOE Vera C. Rubin Observatory Launches Real .... Combined with data from Euclid, the upcoming Roman Space Telescope Nancy Grace Roman Space Telescope, and next-generation radio surveys Revisiting the Unknown Unknowns in Transient ..., the volume and complexity of new observational territory make algorithmic first-pass selection essential. As AI capabilities improve—making cheap, catalogue-scale anomaly vetting increasingly routine—essentially any genuinely new transient class discovered in this era has a high probability of being first picked out by an AI system without human targeting.

Bottlenecks and Limitations However, the final count will be heavily constrained by the rarity of genuinely new astronomical phenomena and the slow pace of the scientific publication pipeline. Humans historically find only about 1–2 genuinely new classes a year across all of astronomy (such as fast radio bursts or long-period radio transients). The vast majority of promising anomalies will either turn out to be instrumental artifacts, be folded into known phenomena, or be classified as subtypes, which do not count. Furthermore, spectroscopic follow-up and human investigation are often required to distinguish true novelty from undersampled known classes Catching Disguised Transients with ASTRANet: Anomaly .... Taking a discovery from initial detection to follow-up observation and finally to a refereed publication typically takes 1–3 years. Consequently, many discoveries made by Rubin or Roman after 2028 are likely to miss the December 2030 cutoff. The timeline for the first AI-discovered new class suggests a high likelihood that no such discoveries are fully reported by the end of 2030, making the timing of even the first qualifying publication a central constraint.

Forecast Distribution These competing forces result in a median expectation of 1 new class, with substantial downside risk and a long right tail. Both the 10th and 25th percentiles are 0, reflecting the substantial possibility that strict criteria application excludes borderline cases like SAURON, and that the timeline to the first fully reported AI-discovered new class and the publication bottleneck prevent any post-2026 discoveries from clearing the hurdle in time. The median of 1 assumes that ML triage in the Rubin era and ongoing archive mining yield at least one refereed new-class claim by the deadline, rather than requiring a steady trickle of qualifying discoveries. The 75th percentile of 3 allows for several discoveries to clear both the classification and publication hurdles. The 90th percentile stretches to roughly 6, allowing for the possibility that referees accept provisional "new class" language more frequently, and that cross-survey anomaly detection pipelines successfully mine optical, infrared, and radio archives to produce multiple independent discoveries.