An AES decryption with the correct key will return the original message, but an AES decryption with an incorrect key will produce garbage data as an output. The AES cipher itself provides no indication that the key was wrong - there's no point during the decryption at which the algorithm says "hey, wait a minute, this doesn't make sense!" and terminates with an error - it just returns pseudorandom data.
Of course, in practical implementations, we tend to use mechanisms such as padding, which can be used to validate that a decryption was correct. Padding is patterned data which is used to turn arbitrary sized messages into blocks of the correct size for the cipher. Since decryption with a "bad" key results in garbage data, it's highly likely that the padding patterns will be incorrect, which allows you to detect that the decryption operation did not succeed. Of course, this could mean that either the key was wrong, or that the ciphertext (i.e. encrypted) block was somehow damaged.
Normally, in order to provide proper validation of decryption with the correct key, the plaintext message starts with some known value which can be validated. For example, TrueCrypt's volume headers start with the string "TRUE". Additionally, message authenticity and integrity codes (MAC/MAIC) can be applied to to the message to ensure that they are not corrupted or modified.