Well let's look at how this actually played out.
- Defendant was in fact sending CP through his gmail.
- gmail correctly detects and flags it based on hash value
- Google sends message to NCMEC based on hash value
- NCMEC sends it to police based on hash value
Now police are facing the obvious question, is this actually CP? They open the image, determine it is, then get a warrant to search his gmail account, and (later) another warrant to search his home.
The court here is saying they should have got a warrant to even look at the image in the first place. But warrants only issue on probable cause. What's the PC here? The hash value. What's the probability of hash collisions? Non-zero but very low.
The practical upshot of this is that all reports from NCMEC will now go through an extra step of the police submitting a copy of the report with the hash value and some boilerplate document saying 'based on my law enforcement experience, hash values are pretty reliable indicators of fundamental similarity', and the warrant application will then be rubber stamped by a judge.
An analogous situation would be where I send a sealed envelope with some documents to the police, writing on the outside 'I believe the contents of this envelope are proof that John Doe committed [specific crime]', and the police have to get a warrant to open the envelope. It's arguably more legally consistent, but in practice it just creates an extra stage of legal bureaucracy/delay with no appreciable impact on the eventual outcome.
Recall that the standard for issuance of a warrant is 'probable cause', not 'mathematically proven cause'. Hash collisions are a possibility, but a sufficiently unlikely one that it doesn't matter. Probable cause means 'a fair probability' based on independent evidence of some kind - testimony, observation, forensic results or so. Even a shitty hash function that's only 90% reliable is going to meet that threshold. In the 10% of cases where the opened file turns out to be a random image with no pornographic content it's a 'no harm no foul' situation.
For reference, a primer on hash collision probabilities: https://preshing.com/20110504/hash-collision-probabilities/
and a more detailed examination of common perceptual hashing algorithms (skip to table 3 for the collision probabilities): https://ceur-ws.org/Vol-2904/81.pdf
I think what a lot of people are implicitly arguing here is that the detection system needs to be perfect before anyone can do anything. Nobody wants the job of examining images to check if they're CP or not, so we've outsourced it to machines that do so with good-but-not-perfect accuracy and then pass the hot potato around until someone has to pollute their visual cortex with it.
Obviously we don't want to arrest or convict people based on computer output alone, but how good does it have to be (in % or odds terms) in order to begin an investigation - not of the alleged criminal, but of the evidence itself? Should companies like Google have to submit an estimate of the probability of hash collisions using their algorithm and based on the number of image hashes that exist on their servers at any given moment? Should they be required to submit source code used to derive that? What about the microcode of the silicon substrate on which the calculation is performed?
All other things being equal, what improvement will result here from adding another layer of administrative processing, whose outcome is predetermined?