§ Claim under review · Capability
"Johns Hopkins astrophysicist and Anthropic researcher Brice Ménard used Claude Science to create the first complete ultraviolet map of the sky, filling in regions that space telescopes had never observed and revealing new details of the Milky Way."
Verdict
Partially accurate but misleading
Confidence
MediumSummary
This is a real project, but the headline oversells it. Brice Ménard really does hold positions at both Johns Hopkins and Anthropic, he really did use Anthropic's Claude Science to assemble a full-sky ultraviolet map, and the map is real and documented on his university page. The overstatement is in two places. First, his own page calls it "the first detailed version with no holes in it," while the post calls it the first complete UV map outright, and a peer-reviewed 2021 paper already built an all-sky far-UV map the same way, using real observations plus machine-learning predictions for the uncovered part. Second, "revealing new details of the Milky Way" sounds like a discovery, but Anthropic itself describes the map as an educational tool, and Ménard's documentation specifically warns users to check whether a region was observed or predicted before trusting a faint feature in it. About a third of the map is predicted rather than photographed, at roughly 20 to 35 times coarser resolution than the observed parts, and the technical paper is listed as in preparation with the data files not yet posted. Nothing here has been peer reviewed or independently checked, and the only account of what the AI did comes from the company that sells the tool.
The readings
key figures from the evidencevendor-stated accuracy of AI-filled UV gaps vs held-out data
technical spec median uncertainty where GALEX did not image (~17%)
Why this verdict
Evidence
Anthropic published a guest post in which Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, explains how he worked with Claude Science to produce the first complete map of the sky in UV light, and the post states the map will be a valuable educational tool allowing students to see in detail the rich structure of the Milky Way visible at this wavelength .
On method, the post states: roughly a third of the sky has never been observed in UV, and to account for those blank spots Claude Science was asked to use inpainting, a machine learning technique in which a model trained on millions of photos learns how each part of an image relates to its surroundings and uses that knowledge to restore missing portions . Where UV data is missing, observations exist at other wavelengths including visible, infrared and radio; using the two-thirds of the sky that has been mapped in UV, Claude learned how UV brightness relates to these other wavelengths and applied that relationship to the third with no UV data . The post reports that after rounds of refinement it was able to estimate the hidden data to within about 10% of the real UV measurements, a difference almost imperceptible to the human eye, and that Claude then added estimates on top of this diffuse inpainted background .
Ménard's own project page is more precisely worded than the social post. It describes the result as the whole sky in ultraviolet light, the first detailed version with no holes in it, most of it photographed by NASA's GALEX between 2003 and 2013, with the parts no ultraviolet telescope ever saw filled in by a statistical model and every such pixel labelled . On why the gap existed: almost all deep, wide-area ultraviolet imaging comes from one mission, GALEX, with about 38,000 snapshots of 1.2 degrees covering 64% of the sky in the far-UV and 76% in the near-UV; detector safety kept it away from ultraviolet-bright stars and the Galactic plane, and after 2009 it observed in the near-UV only .
Critically, the gaps were not filled by prediction alone. The page describes adding other ultraviolet data on the GALEX brightness scale: Swift-UVOT imaging, the FIMS/SPEAR far-UV level where GALEX never looked, TD-1 fluxes for stars GALEX saturated on, and far-UV translated from near-UV-only imaging , and specifies that FIMS/SPEAR far-UV spectro-imaging sets the degree-scale far-UV level over the 24% of the far-UV sky where GALEX never looked . A provenance layer is built in: a five-column weights file is the provenance mixture, with a one-byte provenance code per pixel (0 GALEX, 1 NUV-informed far-UV, 2 UVOT, 3 FIMS-constrained, 4 prediction plus stars) . The page tells readers to use it: it describes GALEX snapshots with model-filled gaps between them as the check to make before trusting a faint feature .
On precision, the technical summary states typical 1 sigma of about 0.02 dex (about 5%) where GALEX imaged, about 0.07 dex median where it did not, largest in the inner plane, with heavy tails , and resolution that is pixel-limited at 1.7 arcminutes where imaged, 0.5 to 1 degree effective where predicted, and degree-scale where only FIMS/SPEAR constrains the far-UV .
On publication status, the page instructs users to cite the technical paper as Ménard, B. (2026), Full-sky ultraviolet maps from harmonised GALEX, Swift-UVOT, FIMS/SPEAR, TD-1 and Gaia data, in preparation , and states that download links for the maps and their documentation will be added to the page .
On prior art, a refereed 2021 paper already did the same category of thing in the far-UV: a far-ultraviolet all-sky map was constructed from FIMS observations aboard the Korean microsatellite STSAT-1, with predictions from a deep artificial neural network used for the roughly 20% of the sky not covered by FIMS , with predicted FUV intensities showing good agreement with GALEX observations at high Galactic latitudes . Earlier, a map of the diffuse ultraviolet background in two bands over almost 75% of the sky was presented from archival GALEX data .
Secondary coverage echoes the vendor framing and also notes the stated purpose: the map is meant to serve as teaching material, and Ménard suspects many scientists have been putting off similar projects that AI could now make possible .
Findings
✓ What's accurate 7
- Brice Ménard is described, by both Anthropic and his own Johns Hopkins project page, as an astrophysicist at Johns Hopkins University and a researcher at Anthropic. The dual affiliation in the claim is correct.
- A real project exists, with a real artifact: a full-sky two-band ultraviolet map dated September 16, 2026, on a HEALPix grid, with a published technical specification.
- Claude Science was used, according to the author's own first-person account, to coordinate the downloading, calibration and merging of data from multiple space missions.
- GALEX did leave large gaps. The author's own figures are 64% far-UV and 76% near-UV coverage, with avoidance of UV-bright stars and the Galactic plane, which is consistent with the post's "roughly two-thirds."
- Regions never seen by any ultraviolet telescope were filled by a statistical model trained on the relationship between UV brightness and visible, infrared and radio data, and each such pixel is labelled.
- The "over 100M stars" figure in the caption is consistent with the project page, which cites 119 million Gaia DR3 stars with measured or predicted UV fluxes.
- Anthropic's own headline does use the words "first complete map of the sky in UV light," so the post is reporting the source's framing rather than inventing it.
≈ What's misleading 7
- The claim says "the first complete ultraviolet map of the sky." The author's own page says the narrower thing, that this is "the first detailed version with no holes in it." Dropping "detailed" turns a qualified first into an absolute one. A refereed 2021 MNRAS paper already constructed an all-sky far-UV map by the same general recipe, observations plus neural-network prediction for the roughly 20% of sky not covered, and earlier GALEX diffuse maps covered almost 75% of the sky. The defensible first is about resolution, two-band scope and per-pixel provenance, not about being the first gap-free UV map ever assembled.
- "revealing new details of the Milky Way" reads as new astronomical discovery. The author's page and Anthropic's post both frame the map's value differently: Anthropic calls it a valuable educational tool for students, and press coverage records it as teaching material. The project page goes further and warns that checking the provenance layer is the check to make before trusting a faint feature in a model-filled region. A map whose own documentation tells you to verify before trusting faint structure is not presented by its maker as a source of new findings.
- **Demo to product conflation, applied to a scientific result:** the claim presents a finished, usable scientific product. The technical paper is listed by the author as in preparation, and the map and documentation downloads were announced as forthcoming rather than posted. Nothing here has been peer reviewed or independently reproduced.
- The only account of what Claude Science did, and the only assertion that the result is a first, come from the tool vendor's own blog and from a researcher affiliated with that vendor. There is no independent evaluation of either the workflow or the map. The post cites "Source: Anthropic," which is accurate, but a vendor's self-report about its own product's scientific output is not independent corroboration.
- **Omitted qualifier, on the filling method:** the framing that unobserved regions were filled by AI prediction understates the layered reality. Much of what GALEX missed was supplied by other telescopes' real observations, with FIMS/SPEAR setting the degree-scale far-UV level over the 24% of the far-UV sky GALEX never looked at, plus Swift-UVOT imaging and TD-1 stellar fluxes. Pure prediction applies only to the residual.
- **Omitted qualifier, on the 10% figure (secondary claim in the caption):** the caption's "within roughly 10% of actual readings" is the vendor's description of test performance against held-out real measurements. The map's own technical summary gives a different and larger number for the delivered product, about 0.07 dex median uncertainty where GALEX did not image, which is roughly 17%, largest in the inner Galactic plane, with heavy tails. Those two numbers measure different things, and quoting only the first makes the filled regions sound more precise than the specification says they are.
- **Omitted qualifier, on resolution:** the predicted regions are not at the same fidelity as the imaged ones. The specification gives 1.7 arcminutes where imaged against 0.5 to 1 degree effective where predicted, a difference of roughly a factor of 20 to 35 in angular scale. A reader told the sky is now "completely mapped" would not infer that a third of it is at vastly coarser resolution.
? What's uncertain 5
- Whether the "first complete UV map" claim survives a full prior-art review. One refereed counterexample of the same type was found, and earlier partial-sky diffuse maps exist, but no systematic survey of all-sky UV map construction was completed, and no independent astronomer has publicly assessed the priority claim.
- What underlying model or models Claude Science used, and how much of the pipeline was autonomous versus human-directed. The claim is version-ambiguous and neither primary source pins it.
- Whether the map's accuracy claims hold up. Without the paper, the posted data files, or any independent reanalysis, the stated uncertainties cannot be checked by anyone outside the project.
- Whether any specific Milky Way feature in the map constitutes a genuinely new scientific finding. Neither primary source names one, and neither source makes that assertion.
- Whether a preprint has been posted. None was located, but the search budget was exhausted before this could be settled, so this is an open question rather than a verified absence.
Sources
6 of 7 linked to recordsMénard's own project page of record, "The full sky in the ultraviolet," Johns Hopkins, October 2026
Anthropic, "Using Claude Science to produce the first complete map of the sky in UV light," published 2026-10-08/09
Jo, Choi, Kim, Woo, Min, Seon, "Construction of a far-ultraviolet all-sky map from an incomplete survey: application of a deep learning algorithm," MNRAS 502, 3200 (2021)
Murthy, Henry, Sujatha, "Mapping the Diffuse Ultraviolet Sky with GALEX"
The Decoder, "Anthropic's Claude Science creates the first complete ultraviolet map of the sky," 2026-10-09
Anthropic, Science hub and related-content pages
Endpoints News / STAT / Northeastern coverage of the Claude Science launch, 2026-06-30