As of late October 2026, no AI system has been credited with finding a new class of astronomical object or transient [S16]. While AI anomaly searches are routinely processing massive datasets, they have thus far identified rare known objects, instrumental artifacts, or novel subtypes rather than distinct new classes. For instance, AnomalyMatch found over 1,300 anomalies in the Hubble archive Researchers discover hundreds of cosmic anomalies with ... Researchers discover hundreds of cosmic anomalies with ..., and Gemini-driven agents vetted nearly 400,000 ASAS-SN variable stars [2608.23688] Agentic Active Learning Meets Visual ..., yet both predominantly recovered rare configurations of known morphologies https://www.aanda.org/articles/aa/full_html/2025/12/aa55512-25/aa55512-25.html. Even operational discoveries like the supernova SN 2023zkd AI discovers new type of supernova triggered by black hole or the SAURON radio galaxy Making new discoveries with AI - ML Making new discoveries with AI - ML represent subtypes or heavily human-steered finds, which the criteria explicitly exclude. Recent genuinely new classes, such as hypersoft X-ray sources and JWST's black hole star, were found by humans These strange celestial objects have been hiding in plain ... Astronomers discover a new kind of cosmic object.

The probability of an AI-first discovery will rise steeply over the next few years, driven primarily by expanding data streams and the fact that machine learning is becoming the default primary filter for observational astronomy. The Rubin Observatory's ten-year LSST survey, which began in June 2026, currently issues about 800,000 alerts per night using automated triage [S16, e95e3d]. Because its alert stream is public without a proprietary period https://rubinobservatory.org/for-scientists/data-products/alerts-and-brokers, multiple independent agentic searches can mine it simultaneously. Combined with falling inference costs and rapidly scaling autonomous agent capabilities, AI systems will soon evaluate raw data from Rubin, Euclid, and upcoming archive releases before humans ever see it. Once manual scanning of raw data is largely phased out, nearly every new class will have its first members selected by an AI system.

However, the pace of discovery is constrained not by data-processing capacity, but by the physical universe and the scientific community's rigorous classification standards. Historically, genuinely new astronomical classes—on the order of fast radio bursts or long-period radio transients—appear only once every one to three years. Establishing a distinct population requires extensive human follow-up observations and conceptual contextualization https://arxiv.org/pdf/2603.29511. These empirical verification bottlenecks also restrict the number of genuinely new phenomena likely to be established by 2030, limiting how quickly a growing pool of AI-selected candidates can produce a qualifying discovery . Furthermore, operational anomaly frameworks often struggle with true novelty; evaluations like the FLARE report indicate that blind detection recovers very few out-of-taxonomy objects at manageable false-alarm rates without a human review loop https://arxiv.org/abs/2610.30656. Because the resolution criteria demand a newly accepted class rather than a subtype, early claims will likely face intense scrutiny and potential reclassification, significantly delaying the time to a qualifying public report.

These conflicting pressures produce an estimate that features a near-term possibility but a substantial right tail. The 10th percentile falls on October 1, 2027, as initial Rubin data and aggressive archive mining offer the first strong opportunities for an AI-flagged breakthrough, while allowing time for empirical verification. The 25th percentile falls on September 1, 2028. The median falls on June 1, 2030, allowing roughly three to four years for an AI system to identify a genuinely distinct population, for human researchers to conduct the necessary follow-up, and for a public report to definitively establish it as a new class. The limited volume of discoveries that can clear these verification bottlenecks by 2030 keeps the median from moving earlier despite rapidly expanding search capacity. The strict exclusion of subtypes and the inherent scarcity of entirely new astrophysical phenomena create a long right tail, with the 75th percentile on June 1, 2033, and the 90th percentile on January 1, 2038.