Well, it turns out that depends on what you mean by "the AES cipher".
If you are talking about the block cipher primitive, that is, if you define an alternate block cipher by taking AES, and swapping the 'encrypt' and 'decrypt' directions, well, that alternative block cipher is precisely as strong as AES. It can be used in any mode of operation we would normally use with AES, and work just as well. In fact, if we had any evidence that your block cipher had any specific weakness, that would immediately translate into a weakness within AES.
However, if you're talking about AES in a specific mode of operation, and use the 'decrypt' mode to encrypt, you may run into unexpected issues, depending on the mode (or it might not work at all; it won't for any authenticated-encryption mode). For CBC mode, yes, you can run into issues (and see below for the extended discussion); for CTR (counter) mode, you don't (in fact, encrypt and decrypt are the exact same operation, so swapping the two don't really change anything).
Now, for the discussion of CBC mode and the weakness you can run into:
CBC mode during encryption is immune to chosen plaintext attacks (assuming that the IV is selected randomly); that is, if the attacker specifies the plaintext, and then examines the encrypted ciphertext, that (with high probability) gives him no information on any other ciphertext he might have. In fact, even if you guesses the exact plaintext that his ciphertext under attack might be, and asks that to be encrypted, the resulting ciphertext will be unrelated (with high probability) to the ciphertext he already has. This is because the IV acts as a randomizer on the entire message.
This is not true for CBC mode in decrypt mode; the IV affects only the first block; hence (in this model) if the attacker submits the guessed plaintext, that will be exactly the same ciphertext (except for possibly the first block); hence CBC mode decryption is not secure in this model.
Side note (getting off topic, but this is the answer to a rather obvious question), this observation does not affect CBC mode decryption as it is normally used, because we add a Message Authentication Code to the message; if the attacker submits his own ciphertext, the MAC will fail, and so he doesn't learn anything. However, if you're using CBC mode decryption to "encrypt", we don't have a MAC.