Will the any published 3GPP 6G specification prior to 2030 include the use of "AI" in the physical layer?

This means that some part of the physical layer specification assumes the use of data-driven learned algorithms and makes provision for same. Note that there could still be alternative provision for non-AI approaches for certain deployments or in fallback scenarios.

This could include removal of pilots on the assumption that the channel state can be inferred, removal of cyclic prefix on the basis that interference can be mitigated by deep learning techniques, anything involving an auto encoder architecture which spans base station and user equipment, etc.

The 3GPP spec must include this; it's not sufficient for specific implementations to use deep learning to which the spec is agnostic.

The phy is everything from the ADC/DACs through to decoded bits.