Why Did the AI Miss a Planet That Shouldn't Exist?

The most interesting part of the TOI-1899 b story isn’t that humans discovered the planet, but that the AI made a mistake. The algorithm did exactly what it was programmed to do: it filtered out a signal that didn’t fit the familiar picture of the universe.

TOI-1899 b is a gas giant nearly the size of Jupiter, but with a mass of about 0.66 times that of Jupiter. It orbits a red dwarf in 29.1 Earth days. This is a tricky case for TESS: the satellite observes a single sector for about 27 days, so the planet only managed to show a single transit on the graph.

A single signal is much harder to detect. But when combined with data on the red dwarf and accepted models of planet formation, the algorithm’s conclusion becomes predictable: such a large gas giant should not exist in this orbit. The archives of previous observations were not automatically checked—the case was deemed unlikely.

People didn’t notice proof, but rather a suspicious anomaly. Planet Hunters volunteers spotted a nearly textbook-perfect transit and took note of the star’s calm behavior. After that, astrophysicists checked the archives and requested additional data from other research groups.

Here’s the practical takeaway: AI is particularly effective when it comes to quickly identifying familiar signals within a massive dataset. But “doesn’t look like anything known” often means “can be discarded” to the AI, even though that’s exactly how a discovery sometimes begins.

There is also an important caveat. The author of the article does not have access to the source code for Planet Hunters TESS and describes how the algorithms work based on publications, presentations, and discussions with the project’s astrophysicists. Therefore, the details of the specific AI solution cannot be considered a fully verified technical reconstruction.

The method also has a fundamental limitation: the transit method can only detect systems oriented appropriately toward us, and TESS’s short observation window makes it difficult to search for planets with long orbital periods. Furthermore, the conclusion that TOI-1899 b “cannot exist” points to the limitations of current models, not to a violation of the laws of nature.

This leads to an interesting trade-off: we can trust automated systems to handle mass screening, but the decision of what to consider impossible is best left under human control. Otherwise, the system will flawlessly detect only what it is already capable of imagining.

In your work, would you prefer an algorithm that misses rare anomalies, or one that floods you with questionable signals for the chance to find something truly new? Source: All Articles in a Row / Artificial Intelligence / Habr