An AI-assisted workflow reduced the reported human time needed to describe NASA science images by 96%. Human editing was still required.
What researchers studied
Researchers developed a semi-automated process to write alternative text for publicly released Chandra X-ray Observatory images.
Alt-text conveys visual information to people using screen readers, but specialist scientific imagery can be difficult to describe accurately.
The researchers tested pretrained vision-language models combined with image metadata.
When that approach proved unreliable, they introduced checks comparing descriptions with both the image and the original caption.
They assessed outputs with a multidimensional rubric and human reviewers.
What they found
The authors reported that human and automated scores sometimes disagreed.
Their automated evaluator was relatively effective at recognising high-quality descriptions, but it was not a complete substitute for people.
The assisted process reduced the human time burden by 96% relative to writing alt-text from scratch.
That comparison concerns this particular workflow, not the accuracy of every generated sentence.
The authors explicitly concluded that descriptions still needed human editing.
What it means
Scientific institutions publish many visual materials, and accessible descriptions take specialist time.
A carefully checked drafting assistant may help observatories and museums address accessibility backlogs.
Metadata and verification can improve usefulness compared with relying on image recognition alone.
Human review remains essential for conveying the scientific meaning of visual features rather than merely naming objects.
Important limitation
The experiment centred on a specialised astronomy collection, not a representative sample of all images.
Evaluation methods and human judgments differed in places.
The reported time saving cannot automatically be transferred to medical, educational or everyday photographs.
Source Information
Study: Accessing the universe via algorithm: a baseline study for automating alt-text generation with NASA data and AI
Authors: Jessica Schonhut-Stasik, Tom Sgouros, Kimberly Arcand, Areshva Aisha Mir, Tianze Hua, Anna Ohrt, Sarah Kane and Nancy Wolk
Journal / Source: Frontiers in Artificial Intelligence
Published: 18 November 2026







