One way to make an AI 'dumber' is to instead give it a handicap. You could limit its movement speed until it doesn't always win.
Another option is to randomize the ball trajectory after it bounces, and then give the AI limited information about the ball trajectory, but rather represent it as a collection of potential states, which gets narrowed down as the ball gets closer to the AI paddle.
In order to use an approach like this, you need to make the AI smarter first, so it can do trajectory prediction, and still be able to respond realistically with limited information.
Needless to say, this is harder.