This is just showing steganography with plaintext payloads. If you use only ciphertext payloads (with the keys exchanged out of band) you sidestep this problem.
Not really. Encrypting the message will yield uniformly distributed noise and that is a very rare in nature. So if you attempt to hide an encrypted message in the least significant bits of images, audio recordings or video it is as easy to detect as plain text messages if not even easier.
Then don't use every LSB in the image; use a low percentage. Just a guess, but I bet if you applied your stego detection algorithm to a large sampling of random images on the internet, you'd find a significant false positive rate. Just hide your messages in the false positives.
As already mentioned by others - hiding small amounts of data is easy, the challenge is to hide nontrivial amounts of data. There are algorithms that (try to) compensate statistical changes introduced by hiding data. One approach is to only use half of the available bit for hiding data and modifying the other half in a way to preserve a set of statistical properties.
It is probably not a good idea to hide data in images available on the internet because this enables direct comparison of the same image with and without hidden data.
