NASA wants your help finding the mistakes in telescope AI

An invitation to inspect the imperfect side of astronomy

The most useful contribution to a space mission is not always spotting something spectacular. Sometimes it is recognising that a mark in the data should not be there—or that a computer is about to remove something it should keep. That is the practical task behind NASA's September 11 invitation to join Artifact InSPECtor.

The project asks volunteers to examine real observations from Euclid and help assess artifacts identified by AI. The new announcement is a call for participation in an existing project, not evidence of a new telescope launch or a scientific discovery made that day.

For anyone curious about scientific AI, it is a revealing example. The work begins by assuming that automated decisions need checking. The human role is not to applaud an impressive output, but to look closely at where it may be wrong.

Sources: NASA: September 11 invitation to help refine telescope data; NASA: Artifact InSPECtor participation guide

Not every signal in a telescope's data comes from the sky

NASA lists several ways unwanted features can enter observations: stray light, reflections involving instrument structures, cosmic-ray hits and detector electronics. Such artifacts can interfere with the information researchers want to extract. The project's ambition is to help train automated tools to identify them more reliably.

The word artifact does not mean the telescope is useless or the observations are fabricated. It means that an observed feature may have an origin other than the astronomical phenomenon being studied. Distinguishing those origins is part of interpreting measurements.

Our explanatory example is a cleaning decision with two possible errors. Leave an unwanted feature in place and a later analysis might be misled. Remove a genuine feature and useful evidence may be lost. A good system needs to handle both risks, rather than simply remove as much as possible.

Sources: NASA: September 11 invitation to help refine telescope data

A spectrum answers questions a photograph cannot

NASA's spectroscopy guide explains how studying light at different wavelengths reveals information about matter, including composition, temperature and motion. A familiar astronomical image helps show where objects are and what they look like. A spectrum provides a different set of clues about what produced the light and what happened to it.

The project's name emphasises that spectral information. For a reader, the important point is that a subtle data-cleaning decision can matter even when it would not make a dramatic before-and-after photograph. Scientific value is not measured only by how attractive an image looks.

The archive image accompanying this article is a visible- and near-infrared composite of the Perseus galaxy cluster released in 2023. It provides Euclid context; it is not an example of an Artifact InSPECtor classification screen or a newly obtained 2026 result.

Sources: NASA: spectroscopy explained

What volunteers are actually asked to do

NASA's project guide says a tutorial takes around 10–15 minutes and no specialist knowledge is required. People can participate on a computer, tablet or phone. The task involves examining highlighted regions and judging whether the AI's identification of an artifact is appropriate, with a discussion area for questions.

The safest route is to start with the official NASA project page linked below and follow its participation link. Reading the tutorial before judging an example matters more than trying to make a large number of rapid classifications. This article does not submit answers or enrol readers automatically.

Our practical advice is to treat uncertainty as useful information, not a personal failure. A volunteer is contributing an observation to a research process, not certifying an entire telescope exposure. An unfamiliar or ambiguous case is a reason to use the project's guidance and discussion tools.

Sources: NASA: Artifact InSPECtor participation guide

Reading the tutorial: object type comes before mask quality

On October 1, 2026, we read all six written steps of the project's public English tutorial and inspected its task definitions. We accessed the official public data feed; the browser interface was unavailable in our environment. We did not complete a rendered tutorial, inspect a live subject or submit a classification. The example below applies the written guidance to a new drawing.

The tutorial describes paired data and mask views, an earlier view before continuum subtraction, and controls for contrast and image layout. Its practical sequence is to identify the object in the green region, then assess the blue overlay. We verified that order in workflow 31550, version 49.111. A correct object label does not establish that a mask is good.

1. Compare the views before assigning a type. The research guide describes horizontal spectral traces and compact emission peaks. The FAQ explains that a horizontal bright stripe in the before-subtraction view identifies a spectrum; a tiny isolated peak can also be useful emission from a faint galaxy. An unfamiliar shape alone is insufficient reason to discard it.
2. Apply a stated assumption to this drawing. The research page describes a double-lobe or infinity-like zeroth-order shape created by the grism optics. We drew two lobes to represent that example. Assuming both belong to this one shape, the appropriate object category would be Artifact. This is our explanation of the guide, not a diagnosis of a real exposure.
3. Judge coverage separately. Our blue overlay includes the left lobe and misses the right. The written tutorial specifically treats an omitted lobe as an incomplete mask. Under our drawing's premise, the answer would be No, the object is not masked correctly. No answer was selected or sent to the project.
4. Keep mixed or empty regions visible in the reasoning. The inspected task definition includes separate categories for mixed object types and no clear subject. The FAQ allows a small border around an object while cautioning against detached blue pixels or an excessive mask. These cases need the project's examples; our clean two-lobe drawing cannot establish the right boundary for noisy data.

The useful distinction is between recognising a feature and deciding which pixels belong to it. A mask can look convincing at first glance and still miss a whole component. This demonstration explains that failure mode; it establishes no improvement in model accuracy, no classification consensus rule and no cosmological result. The dated reading record preserves the inspected sources, version, access limits and illustration provenance.

Source notes: 6, 7, 8, 9. The schematic is new; tutorial-media reuse rights were not established. Media guidance: 10.

Many judgements can be useful, but agreement is not infallibility

Zooniverse, the citizen-science platform, explains that classifications from multiple volunteers are combined so researchers can assess agreement and uncertainty. This is a general platform principle. The sources reviewed here do not establish a specific vote threshold or number of reviewers for each Artifact InSPECtor example.

Our assessment is that repeated judgements can make the process more informative than relying on one hurried click. They do not automatically make it correct. People can share a misunderstanding, particularly if a difficult case resembles a familiar example from a tutorial.

That is why the research question should not stop at whether volunteers agree with one another or with the model. The more useful question is whether their contribution improves the treatment of genuinely new observations when assessed against carefully checked reference data.

Sources: Zooniverse: how citizen-science contributions are combined

Scientific perspective: human feedback needs a measured outcome

Lumacta's evidence-based assessment—not an independent evaluation of the project—is that this is a sensible use of human feedback for a task where errors have interpretable consequences. The September invitation reports no measured project-specific accuracy gain. We therefore cannot say how much the volunteers have improved the AI or any resulting cosmological estimate.

A convincing evaluation would test updated models on observations kept separate from the examples used for training. It would report missed artifacts and wrongly rejected scientific features separately, and examine difficult cases rather than only an overall average. These are our suggested scientific checks, not published findings from Artifact InSPECtor.

The positive possibility is a feedback loop in which public participation helps researchers understand and reduce an automated system's errors. The limit is equally important: a larger collection of labels is an input to scientific work. Its value still has to be demonstrated in the quality of the resulting measurements.

Sources: NASA: September 11 invitation to help refine telescope data; Zooniverse: how citizen-science contributions are combined

The connection to dark energy is a research goal, not today's result

Euclid is an ESA mission with important NASA contributions. Its mapping of the universe is intended to help investigate dark matter, dark energy and cosmic evolution. Cleaner observations can support that programme; they do not constitute an explanation of dark energy by themselves.

NASA's invitation also says observations from the Nancy Grace Roman Space Telescope are planned to enter the project in early 2027. That is a stated future plan, not evidence that Roman observations are available in this workflow now.

Our conclusion is that the immediate significance is modest but meaningful: an accessible way to take part in checking scientific data, and a reminder that trustworthy AI depends on visible limits and careful validation. The exciting destination is a better understanding of the universe. The work in front of volunteers is more concrete—helping researchers decide which pieces of a measurement deserve to be trusted.

Sources: NASA: September 11 invitation to help refine telescope data; NASA: Euclid mission overview