There is a part of astronomy that never makes the news: the work of deciding which pixels are real. Every detector produces artefacts – reflections, cosmic-ray hits, sensor quirks – and those artefacts can look enough like real signals to mislead an automated pipeline. On 11 September, NASA introduced a citizen-science project called Artifact InSPECtor, which asks volunteers to help recognise these unwanted features in telescope data. It sounds minor. It is not.
The project supports work connected to the Euclid and Roman missions, and its output is intended to help train artificial-intelligence systems. That combination – humans teaching machines to clean data – says something useful about how modern science actually operates.
What artefacts are, and why they matter
An artefact is anything in an image that is produced by the instrument rather than by the sky. A reflection inside the optics, a charged particle striking the detector, a hot pixel that responds oddly – all can leave marks that resemble stars or galaxies. If those marks are counted as real, they corrupt the statistics that surveys depend on. In a mission designed to measure subtle cosmological signals, even a small rate of contamination can bias the result.
Why AI still needs humans here
Machine learning is good at recognising patterns it has been trained on. The problem is that artefact types are varied, sometimes rare, and sometimes new. A model trained only on known artefacts will not reliably catch an unfamiliar one. Human volunteers, by contrast, can reason about a strange feature and flag it for attention, providing the examples that later improve the automated system. The human contribution is how the training data gets built.
The Euclid and Roman connection
Both missions are survey telescopes, designed to image huge areas of sky and build catalogues of enormous numbers of objects. At that scale, manual inspection of every image is impossible, so automated processing is essential. But the accuracy of that processing depends on the quality of the examples it learns from. This is where citizen science fits: volunteers help create a reliable reference set, which makes the automated pipelines more trustworthy.
The task is classification, not discovery. Participants look at images and mark features that appear to be artefacts rather than real objects. It is repetitive, and that is the point – many small judgements, aggregated, produce a dataset that no single researcher could assemble alone. NASA’s announcement provides the current instructions and participation link, and anyone interested should start there.
The bigger picture: science as collective work
Citizen science is often framed as outreach, a way to let the public feel involved. That framing undersells it. Projects like this one produce data that researchers actually use, and the volunteers are contributors rather than an audience. The arrangement reflects a broader truth about large scientific collaborations: the work is distributed, and the people doing the distribution are not always professionals.
Why this kind of work is underrated
Data cleaning is unglamorous. It does not produce headlines, and it rarely wins prizes. But the reliability of a scientific result rests on it, and as datasets grow, the task grows with them. The fact that NASA is asking the public for help is a candid admission that this work matters and that there is more of it than the professional community can handle alone.
Participation is open, and the requirements are minimal: a browser and a willingness to look carefully. The project’s value to the volunteers is the chance to touch real data from missions at the frontier of astronomy. Its value to science is a cleaner foundation for everything built on top. Both are worth taking seriously.
How volunteer data is validated
Volunteer classifications are not taken at face value. Projects typically combine many independent judgements on the same image, use agreement rates to estimate reliability, and weight or discard responses that diverge. Some tasks include deliberately seeded examples to check that participants are following instructions. The result is a dataset with measurable uncertainty, which researchers can then use. This is why well-designed citizen science is not amateurish: it treats human judgement as a measurement to be calibrated, not as an opinion to be collected.
The economics of scientific attention
There is more data than there are scientists to examine it, and that gap is widening as instruments improve. Citizen science is one response: it distributes the work of looking. But it also raises a question about priorities. If data cleaning is important enough to crowd-source, it is important enough to fund properly. Volunteers can fill a gap, and relying on goodwill to sustain essential infrastructure is a fragile arrangement that says as much about research funding as it does about public enthusiasm.
This report draws on NASA Citizen Science.
Image: NASA Jet Propulsion Laboratory · Public domain · via Wikimedia Commons.