This is actually the magic of FHE. The ciphertext is indistinguishable from noise AND can be computed on, it just looks like different noise.

If you believe the underlying cryptographic hardness assumption of LWE/RLWE/etc, then yes Google cannot see any of the input or output of the model.