Encrypting $M$ using $H(M)$ as the key is a natural and well-studied approach to deduplication. It is known in the literature as convergent encryption or message-locked encryption.
The natural problem with this approach is that it cannot achieve the standard notions of security for encryption (IND-CPA, IND-CCA, etc).
Indeed, anyone who knows $M$ will be able to verify whether a candidate ciphertext decrypts to $M$. The question then becomes: what is the "best possible" level of security that one could hope for, and can we achieve it?
The most thorough analysis of message-locked encryption that I know of is:
Mihir Bellare, Sriram Keelveedhi, Thomas Ristenpart: Message-Locked Encryption and Secure Deduplication, Eurocrypt 2013.
They define relevant security notions, and give corresponding constructions & proofs. They also consider some other security goals closely related to the deduplication application, separate from standard secrecy and integrity.
You can search ePrint and find many other papers with "message-locked" or "deduplication" (for other approaches) in the title. Note that message-locked encryption is non-interactive, but the problem of deduplication becomes a little easier when you allow interaction. Hence, many of the techniques that you find may be interactive.