The “PII” label is taint that is probably impossible to dispel completely/perfectly, and difficult to dispel sufficiently (and deanonymising is an arms race).

Lossless techniques do nothing to dilute that taint.

Lossy techniques are necessary to get anywhere, such as disregarding certain bits of the address, or Bloom filters.

The problem, in general with hashing IP addresses (especially ipv4) is that there's not that many of them.

If I tell you the value is either 1 or 2, but I hashed it with sha256 to make it secure, that's bullshit, right? You can just hash both and see which it is.

Same concept applies regardless of the hash algo, and still applies if you have more than 2 possible values, 4 billion or so possible ipv4 addresses is _not_ that many values to a computer.

Other common places this problem occurs is with any other restricted set of values, eg phone numbers and email addresses (most are at like 5 domains and are easy to guess/know).