AI Exoplanet Searches: Follow the Evidence Behind the Count

A headline announcing thousands of AI-found planets sounds like an extraordinary haul of distant worlds. Before sharing it, look at what was counted. A list of promising signals, a catalogue of objects needing investigation and a set of validated planets describe different achievements.

The published abstract of the 2025 ExoMiner++ study provides a useful example. Among 147,568 unlabelled threshold crossing events in TESS data, the model classified 7,330 as planet candidates. That is a substantial result for organising further investigation. Understanding the unit behind that number is part of understanding the achievement.

A dimming star gives astronomers something to investigate

When a planet passes in front of its star from our viewpoint, it can block some starlight. Astronomers plot brightness over time to examine the resulting dip. NASA’s account of the original ExoMiner work explains this transit method: researchers study the changing light rather than a close-up photograph of a distant world.

The NASA Exoplanet Archive’s FAQ describes a threshold crossing event, or TCE, as a sequence of transit-like features strong enough to be passed on for further analysis. The term names a stage in the search. Further analysis determines how well a planetary explanation fits.

Other phenomena can resemble a planet’s transit. NASA’s ExoMiner++ explanation, published on 22 January 2026 and updated on 8 October, identifies eclipsing binary stars as one alternative. The model helps classify observations of possible transits, distinguishing promising planetary signals from other explanations.

Read the catalogue’s unit of measurement

ExoMiner++ uses several diagnostic inputs, including brightness trends, difference images and spacecraft pointing information. The published study combined labelled Kepler and TESS data for training, addressing the noisier, more ambiguous labels in TESS. Its analysis therefore draws on several pieces of diagnostic evidence.

The abstract also maps the 7,330 candidate classifications to existing catalogue objects and lists 50 newly introduced Community TESS Objects of Interest. These are different counting units and statuses. Neither a candidate label nor a newly introduced community entry, by itself, establishes a newly confirmed planet.

For a reader, this changes the question to ask of a large number. Does it count events processed, entries given a particular model label, distinct objects selected for follow-up, or planets accepted after validation? Preserve the paper’s unit when describing its result.

Statistical validation can establish a planet

It would also be a mistake to dismiss all machine-assisted results as permanently provisional. NASA’s 23 November 2021 account describes 301 planets statistically validated using the original ExoMiner. Those signals had already been detected by the Kepler processing pipeline and promoted to candidate status before that validation step.

NASA distinguishes statistical validation, which evaluates the likelihood of a planetary explanation, from confirmation through different observational techniques. That distinction describes routes to establishing a planet; it should not be turned into a rule that statistical work can never establish one.

The archive makes its own naming convention explicit: planets accepted under its criteria are categorised as confirmed, and it does not maintain a separate disposition for validated planets. Consequently, read the supporting paper as well as the catalogue label to understand how planetary status was established.

The useful question is which evidence supports this particular object. A model score belongs inside that enquiry. It should not be treated as interchangeable with every other kind of evidence, or dismissed merely because a statistical method contributed to the result.

Follow the work after classification

NASA’s January 2026 explanation describes ExoMiner++ working from a supplied list of possible transit signals. It separately describes identifying signals directly from raw data as work the team was pursuing. Those statements concern the workflow described in that release, rather than every subsequent version of the software.

The same account emphasises publicly available mission data and open-source code. Access lets other researchers investigate the work and develop it further. My interpretation is that a well-constructed candidate catalogue can be valuable research infrastructure: it gives investigators a more focused starting point for deciding what to examine next.

Consider a fictional science newsletter preparing a post about this research. Its editor initially writes that AI has discovered thousands of new worlds. Reading the abstract reveals that the central number describes signal classifications. The editor revises the post to explain the candidate catalogue and the follow-up work it supports.

That revision preserves the interest of the research while making the claim traceable. Before publishing, the editor records:

• Which paper and software version produced the reported result.
• What each entry represents and whether existing catalogue objects are included.
• What status the authors assign to the entries.
• What additional evidence would justify the next claim.

Keep habitability as a further scientific question

Even an established planet has more to reveal. In a 5 June 2024 explanation from NASA’s Webb project scientists, the habitable zone concerns an orbital region where surface liquid water could be possible. Being in that zone does not establish that a world can support life, still less that life is present.

The scientists also describe the difficulty of detecting small planets’ atmospheres and interpreting possible biological signatures. Planetary status, atmospheric characterisation and evidence of life each require their own supporting observations and analysis. A transit-classification result should not silently acquire all three conclusions as it travels into a headline.

Illustrations need the same care. An imagined planet with blue seas and clouds can make a story engaging, but those details must not be presented as observations unless the evidence supports them. Explain what an image depicts and what the instruments actually measured.

Read the discovery story as a sequence of claims

For creators explaining AI research, dates and verbs carry useful information. A new analysis of archived telescope data can make a real scientific contribution. Its publication date does not mean the underlying observations were made that day. Likewise, detecting, classifying, validating and characterising describe different work.

MaryChuks.com’s guide to reading a research paper critically offers a broader route through the question, method, finding and limitations. For this example, a good first step is to keep the count, its unit and its scientific status together in your notes.

AI-assisted astronomy is worth following because it can help researchers work through difficult evidence at scale. The most informative story explains where that contribution enters the discovery process, what has been established and what remains to investigate. Readers can appreciate the achievement more fully when they can see how the conclusion was reached.

Sources

Primary sources checked on 25 October 2026. The newsletter example is fictional; historical results are identified by their publication dates.

• Valizadegan and colleagues: ExoMiner++ — published abstract and metadata; online 28 October 2025, journal issue 3 November 2025; DOI: 10.3847/1538-3881/ae03a4.
• NASA: New Deep Learning Method Adds 301 Planets to Kepler’s Total Count — 23 November 2021.
• NASA: ExoMiner++ and TESS data — 22 January 2026; updated 8 October 2026.
• NASA Exoplanet Archive: Frequently Asked Questions — TCE definitions and treatment of validated planets.
• NASA Webb: Reconnaissance of Potentially Habitable Worlds — 5 June 2024.