NASA's volunteers found 3,000 brown dwarfs. Kids can help too

Two pictures of the same patch of sky, taken years apart, flicking back and forth on a laptop. Almost everything sits exactly where it was. Then one faint smudge shifts, just slightly, against the fixed stars behind it. Somebody clicks on it.

That is the whole job. And in May 2026 NASA announced what the job had added up to: volunteers on a project called Backyard Worlds: Planet 9 have found more than 3,000 brown dwarfs, roughly doubling the number known, over the ten years the project has run. Brown dwarfs, as NASA describes them, are “balls of gas the size of Jupiter, less massive than stars”. About 200,000 people have taken part. The paper, published in the Astronomical Journal and led by the astronomer Adam Schneider of the U.S. Naval Observatory, carries 75 authors. Sixty one of them are volunteers.

[[photo:1]]

Why a person, and not a program

The obvious question in 2026 is why anyone is clicking at all. Surely a model could scan those images faster than 200,000 people blinking through sixteen years of data from NASA’s retired WISE and NEOWISE-R missions.

Some of it can, and plenty of the sky has been swept that way. But the project’s own page is blunt about the limit: “There’s still no substitute for the human eye when it comes to recognizing subtle motions in astronomical images.” When NASA’s Jet Propulsion Laboratory launched the project in 2017, it described the problem more precisely. Machines “are often overwhelmed by image artifacts, especially in crowded parts of the sky”, meaning the brightness spikes around bright stars and the blurry blobs of light scattered inside the instrument. The project, JPL wrote, “relies on human eyes because we easily recognize the important moving objects while ignoring the artifacts”.

The method is not even new. The project page shows the paired 1930 photographs in which Pluto was found by exactly this trick, and JPL called the effort a 21st-century version of the technique the astronomer Clyde Tombaugh used to find Pluto that year. Two images of one patch of sky, a careful eye, a dot that had shifted.

A billion clicks

Backyard Worlds runs on Zooniverse, the platform that hosts most of this kind of work, co-founded by the Adler Planetarium in Chicago and the University of Oxford. At the end of June 2026 Zooniverse announced that its volunteers had passed one billion classifications: more than three million people, across 160 countries, on over 550 projects. The team put that at roughly 2,000 years of full time effort. Even at one second per classification, a billion of them would take 31.7 years.

NASA alone has sponsored 31 projects there since 2020, drawing 120 million classifications from 324,000 volunteers, which have fed 96 scientific papers. Fifty six of those papers list a NASA citizen scientist as a co-author. Laura Trouille, the Zooniverse principal investigator, called the milestone “one billion moments of curiosity transformed into meaningful contributions”.

The leftover pile is the interesting pile

Here is the part worth explaining to a child, because it is the opposite of what they will assume. AI has not taken this work away. It has taken the boring nine tenths and left the strange bit.

A 2024 study in Citizen Science: Theory and Practice looked at this, using camera trap images from Snapshot Safari in the Serengeti. The AI model on its own had an overall error rate of 34.89%. Put a human in the loop to check the model’s calls and the overall error rate fell to 8.73%, and 42 of the 44 species classes improved. The worst failure is the telling one: for eleven rare species with fewer than 100 training images, the model scored zero. Not poor. Zero. The authors concluded that even improved models “require ongoing human training, validation, and supervision to return sufficiently accurate results”.

So the animal the machine has barely seen is the one a nine year old gets handed. Worth knowing at nine.

[[photo:2]]

What the platform now has to tell you

On 9 September 2026 Zooniverse published its first AI ethics framework, after a year of workshops and more than a thousand volunteer survey responses. It asks project teams to answer five plain questions (who manages data quality, what techniques are used, where in the process the AI appears and where volunteers fit, and why the AI is needed at all) and to say clearly on the project homepage where a model is involved.

Two lines in it matter for a parent. Volunteers’ classifications may be used to train models, and projects now have to disclose that up front. And after a project’s proprietary period, typically two years, the data becomes open, which means it can be picked up for uses nobody listed at the start. None of that is alarming. It is worth a child knowing that their clicks teach a machine as well as a scientist.

How to start, tonight

You do not need an account to classify. Zooniverse states plainly that “anyone can participate” and “you do not need any specialized background or expertise”. If your child wants an account so they can keep a tally of their own clicks, the guidance asks under 16s to read the youth policy with a parent first, which is a sensible ten minutes.

Pick an animal project before an astronomy one for younger children: naming a zebra is easier than judging a smudge. Guess out loud before you click. Argue with each other. When the two of you disagree, remember that this is exactly why several people are shown every image. For the version that runs on ears rather than eyes, we wrote about AI learning animal languages in August.

The Singapore version of the hobby is mostly outdoors, and better for it. NParks runs Heron Watch, Dragonfly Watch, Butterfly Watch, Garden Bird Watch, Intertidal Watch and a nationwide BioBlitz through its Community in Nature initiative, all built on ordinary people writing down what they saw. The machines are arriving here too. NParks has funded research on improving acoustic monitoring with machine learning, and secured $1.34 million from the Ministry of National Development to put drones and robotics into wildlife tracking. Same split as the Serengeti. The model takes the volume, the people take the odd ones.

We build an AI mentor for children, so our enthusiasm is not neutral. It is also part of why this corner appeals: there is no chatbot anywhere in it. Just a child, a pile of images a model could not resolve, and a scientist somewhere waiting on the answer.