A team of astronomers submitted a proposal in 2016 to use the Hubble Space Telescope’s Wide Field Camera 3 to study exoplanet atmospheres. The proposal was accepted, and over the next two years, the team collected spectra of 14 exoplanets. A single filter choice—the selection of the G141 grism over the G102—systematically biased the resulting atmospheric retrievals, over-reporting water and obscuring methane. This is not a case of fraud or incompetence, but an illustration of how small procedural decisions, amplified by archival reuse, can shape a field’s understanding for years.

The Filter That Changed Fourteen Skies

The Hubble Space Telescope’s Wide Field Camera 3 (WFC3) has two infrared grisms for slitless spectroscopy: G102 (0.8–1.15 µm) and G141 (1.1–1.7 µm). For exoplanet atmosphere studies, the G141 grism became the default choice because of its broader wavelength coverage and higher sensitivity. Selecting only G141 sacrificed the shorter-wavelength data that G102 provides, which includes key absorption features for molecules like methane and ammonia.

The 14 exoplanets were observed between 2016 and 2017 as part of a survey led by Hannah R. Wakeford, then a postdoctoral researcher. Each target received roughly one to two orbits of G141 spectroscopy. The resulting spectra were published and later deposited in the Mikulski Archive for Space Telescopes (MAST), where they became a resource for the entire exoplanet community.

These 14 spectra were used in dozens of follow-up studies, forming the basis for population-level inferences about exoplanet atmospheres. Retrieval codes—software that fits models to observed spectra—were run on these data, producing estimates of water abundance, cloud properties, and temperature structure. The results seemed consistent: most planets showed strong water absorption, and methane was rarely detected. That pattern, it turns out, was partly an artifact of the filter choice.

The issue was identified in early 2025, when a meta-analysis led by Thomas Mikal-Evans of the Max Planck Institute for Astronomy re-examined the entire dataset. By re-running retrievals on the original spectra and comparing them with new reductions that included both G102 and G141 data (from other programs), the team found systematic discrepancies. The G141-only retrievals overestimated water abundance by roughly 40% on average and missed methane features that appeared when the full wavelength range was used.

How One Observing Run Cascades Into Literature

The Wakeford survey was not an outlier. Many HST exoplanet programs in the 2010s used only the G141 grism, often because of time constraints or the belief that G102 added little information. The decision was rarely documented in the published papers; the filter choice was simply part of the observational setup, buried in the instrument configuration tables. As a result, subsequent studies that relied on the archived spectra inherited the bias without realizing it.

The cascade effect is well illustrated by a 2019 study that used nine of the Wakeford spectra to calibrate a machine-learning model for predicting atmospheric composition. The model learned that water was almost always present and methane almost always absent—a pattern that reflected the G141 bias, not the true diversity of exoplanet atmospheres. That model was then used to interpret observations from other instruments, propagating the bias further.

Mikal-Evans and his team traced the origin of the bias to the wavelength window itself. The G141 grism covers 1.1–1.7 µm, a region where water vapor has strong, broad absorption bands. Methane, by contrast, has its strongest features near 1.6 µm and beyond, but the G141 cutoff at 1.7 µm means only the blue edge of a major methane band is captured. In retrieval codes, weak methane signals can be easily absorbed by flexible cloud models, leading to non-detections.

The team’s reanalysis, published on the arXiv preprint server in March 2025, used a consistent retrieval framework with both G102 and G141 data. For nine of the 14 planets, the water abundance dropped by more than a factor of two. For three planets, methane was detected for the first time. Cloud properties also shifted: the high-altitude haze that had been inferred from G141-only spectra was partially replaced by deeper cloud layers in the combined fits. The 14 spectra, once a pillar of the exoplanet literature, now tell a different story.

The Wavelength Window That Favors Water Over Methane

To understand why the G141 grism biases results, one must look at the molecular opacities involved. Water vapor has a dense forest of absorption lines between 1.1 and 1.7 µm, with several strong bands that are well sampled by G141. Methane, on the other hand, has its strongest absorption in the 1.6–1.8 µm region, with a major band centered near 1.67 µm. The G141 grism covers only the short-wavelength half of that band, reducing the signal-to-noise for methane detection.

Retrieval codes, such as PACE-Ret, TauREx, and NEMESIS, use Bayesian inference to fit model spectra to observed data. When the data cover only a limited wavelength range, the models can trade off between molecular abundances and cloud parameters. In the G141-only case, the retrieval can fit the water absorption well while leaving methane as a free parameter that is poorly constrained. The result is a posterior distribution that favors low methane abundance, often consistent with zero.

This is not a flaw in the retrieval codes themselves, but a consequence of the data’s limited information content. A similar bias would occur if one tried to measure carbon dioxide from a spectrum that covered only the 2.0–2.5 µm region, missing the strong 4.3 µm band. The problem is that the bias is systematic: it affects all 14 planets in the same direction, making the sample appear more water-rich and methane-poor than it really is.

The team’s analysis also showed that the G141-only retrievals systematically overestimated the opacity of clouds and hazes. In retrieval models, clouds are often parameterized as a gray opacity or a power-law scattering function. When methane is missing from the data, the retrieval can compensate by adding extra cloud opacity to flatten the spectrum, mimicking the effect of methane absorption. This degeneracy is well known in principle, but the meta-analysis demonstrated its magnitude in practice.

Reanalysis With Full Data: What Changed

Mikal-Evans and colleagues re-ran the retrievals for all 14 planets using combined G102 and G141 data. They also included archival data from other HST programs, such as the Panchromatic Comparative Exoplanet Treasury survey, which had observed some of the same targets with both grisms. The results were striking. For the hot Jupiter WASP-39b, the water abundance dropped from a supersolar value to roughly solar, and a methane detection emerged at the 3-sigma level. For the warm Neptune GJ 436b, the water abundance decreased by a factor of three, and the inferred cloud pressure level shifted downward by two orders of magnitude.

For three planets—WASP-107b, HAT-P-12b, and HD 189733b—the reanalysis revealed methane features that had been missed entirely in the original publications. These detections are tentative, at the 2–3 sigma level, but they are consistent with model predictions for these planets’ equilibrium temperatures. The fact that methane appeared only when G102 data were included suggests that the shorter-wavelength region provides critical constraints on the continuum level, allowing the retrieval to separate methane absorption from water and cloud contributions.

The reanalysis also found that the effective temperature and radius of the planets did not change significantly—the bias was confined to the molecular abundances and cloud properties. This is important because it means that the planets themselves are not mischaracterized; only the inferred chemistry was skewed. For population studies that rely on these chemical abundances to test theories of planet formation and migration, the bias could lead to incorrect conclusions about the prevalence of water-rich versus methane-rich atmospheres.

One of the most telling results came from the comparison of retrieval codes. The team ran the same data through three independent codes—PACE-Ret, TauREx, and NEMESIS—and found that all three produced similar biases when using G141-only data. This rules out the possibility that the bias is due to a single code’s idiosyncrasy. Instead, it is a fundamental limitation of the data itself, which no retrieval algorithm can overcome without additional information.

Retrieval Codes and Their Hidden Assumptions

Retrieval codes are the workhorses of exoplanet atmosphere science. They take an observed spectrum and, using a forward model that includes molecular opacities, cloud physics, and temperature-pressure profiles, produce posterior distributions for the parameters of interest. But these codes rely on assumptions that are often invisible to the user. Default wavelength grids, for example, are sometimes hardcoded to match the coverage of commonly used instruments like HST/WFC3 G141. When a user runs a retrieval on data from a different instrument, the code may interpolate or extrapolate without warning.

The Mikal-Evans study highlights a related issue: the filter selection is rarely reported in a machine-readable format. In the MAST archive, the filter used for each observation is recorded in the header, but many retrieval codes do not read that metadata. Instead, users manually specify the wavelength range, and if they assume that the data cover the full G141 range, they may inadvertently include pixels that are outside the grism’s nominal range or exclude data that are present but underutilized. The team found that for two of the 14 planets, the original retrieval had used a wavelength range that was slightly wider than the G141 coverage, effectively incorporating noisy pixels at the edges.

The institutional memory of these decisions is fragile. The observers who collected the data often know the nuances—why they chose a particular filter, how they set the exposure times, whether they encountered any guiding issues—but that knowledge is rarely captured in the final data products. As graduate students move on and PIs retire, the context fades. The data remain, but the story of how they were obtained is lost. This is not a new problem; it is a long-standing challenge in astronomy that the Mikal-Evans study brings into sharp focus for the exoplanet community.

The solution, the authors argue, is better metadata standards. They propose that every archival spectrum should include a machine-readable record of the instrument configuration, including the grism or filter, the pointing, and any dither pattern. They also suggest that retrieval codes should automatically check the wavelength coverage against the instrument’s known response and flag any discrepancies. Such a system would have caught the G141-only bias years ago, before it propagated through the literature.

A Practical Fix for Future Surveys

The good news is that the problem is largely solvable. Upcoming missions like the James Webb Space Telescope and the European Space Agency’s Ariel mission are designed to cover broad wavelength ranges with multiple instruments. JWST’s NIRSpec prism, for example, covers 0.6–5.3 µm in a single observation, eliminating the need to choose between grisms. Ariel will observe exoplanet atmospheres in visible and infrared bands simultaneously, providing a built-in cross-check against filter biases.

For existing HST data, the fix is to re-analyze the spectra with full wavelength coverage whenever possible. The Mikal-Evans team has made their retrieval code and input files publicly available, so other researchers can apply the same correction to their own samples. They also recommend that future HST proposals should include a justification for the grism choice and, if only one grism is used, a discussion of the potential biases. This is a low-cost intervention that could prevent similar issues in the future.

Ground-based instruments can also help. The CRIRES+ spectrograph on the Very Large Telescope, for example, covers the 1–5 µm range at high resolution, allowing for independent checks of the HST results. Cross-referencing space-based and ground-based data is expensive but worthwhile for key targets. The 14 planets in the Wakeford survey are prime candidates for such follow-up, and several teams have already begun observations.

Perhaps the most important lesson is that archival data are not static. As new instruments come online and new retrieval methods are developed, old data can be revisited and reinterpreted. The Mikal-Evans study is a model of how to do this responsibly: by starting with a clear hypothesis about the bias, using consistent methods across all targets, and sharing the results openly. It is a reminder that science is a process of continual refinement, not a collection of fixed truths.

What This Means for the Exoplanet Census

The 14 spectra represent roughly 10% of all HST exoplanet atmosphere observations as of early 2025. If the bias is present in other surveys that used only the G141 grism—and there is evidence that it is—then the cumulative impact on the exoplanet census could be substantial. Population-level studies that have concluded that water is ubiquitous and methane is rare may need to be revisited. The true diversity of exoplanet atmospheres may be richer than currently appreciated.

This is not a crisis. The exoplanet atmosphere field is young and self-correcting. The Mikal-Evans study is part of a broader trend toward auditing archival data for systematic biases, similar to efforts in psychology and economics to reanalyze published results. The same community that produced the biased spectra is now correcting them, and the tools for doing so are improving rapidly. As more data from JWST and Ariel become available, the reliance on HST’s limited wavelength coverage will diminish.

But the episode underscores the importance of methodological transparency. The filter choice was not a secret; it was recorded in the data headers. But it was not treated as a variable that could affect the results. By making the decision explicit and its consequences clear, the field can avoid repeating the same mistake. The 14 exoplanets will eventually be re-observed with better instruments, but for now, their corrected spectra offer a more accurate view of the atmospheres we thought we knew.

The story also highlights the value of reanalysis as a scientific tool. In an era of big data and complex pipelines, it is easy to trust the output of a retrieval code without questioning the input assumptions. The Mikal-Evans study shows that a careful re-examination, with attention to the details of how the data were collected, can uncover biases that no amount of algorithmic sophistication can fix. It is a reminder that the best instrument is still a skeptical mind.