K2-18 b: The Biosignature That Did Not Hold
How a claim about life gets taken apart
K2-18 b is a sub-Neptune in its star's habitable zone, 124 light-years away, and for two years it carried the most widely reported biosignature claim in exoplanet science. In 2023 and again in April 2025 a team using JWST reported dimethyl sulfide, a gas that on Earth is produced almost entirely by marine life. What happened next is the part worth keeping: independent groups reanalysed the same public data, and could not reproduce it.
What survived the reanalysis
Methane. Confirmed at 4 sigma and not disputed by anyone.
A temperate sub-Neptune in the habitable zone. Unchanged and still the reason this planet matters.
Carbon dioxide. No statistically significant evidence in the reanalysis.
Dimethyl sulfide. No statistically significant evidence. Several other molecules fit the same features equally well.
The hycean hypothesis, an ocean under a hydrogen sky. Neither confirmed nor killed. Still open.
Radius
2.37 Earth radii, a sub-Neptune
Mass and density
8.92 Earth masses at 4.11 g/cm3
Equilibrium temperature
284 K, measured
One year
32.94 Earth days
Host star
K2-18 at 3457 K, a red dwarf
Distance
about 124 light-years
This is not an argument that the original team was careless. They published their data, which is exactly what allowed four other groups to check it, and the correction happened in under four months. The failure was not the paper. It was the distance between a marginal feature in a noisy spectrum and a headline about life in another solar system.
The technical objection is degeneracy. A transmission spectrum is a handful of absorption dips, and many molecules with methyl groups produce dips in the same places. Fitting dimethyl sulfide is possible; so is fitting several alternatives nobody would write a press release about.
Worth noticing how fast this worked. The claim was public in April, the data was public with it, and by August four independent groups had answered it. That is the self-correction people say science has, running visibly and on a timescale of months.
The data was public, the check took four months, and the correction got a \
fraction of the coverage the claim did. What would it take for the correction to travel as far as the claim?