“The animal will expect a food reward every time it does something right, and will become upset and confused if the food reward does not come.”
Let’s take this apart. Often it’s helpful to invert the criticism to see how a traditional trainer would respond to it.
Does a traditional trainer correct every wrong response?
The answer is yes. Otherwise, the animal becomes upset and confused. Sometimes forging in heel is okay, but other times it gets a collar correction. Sometimes it’s okay to whine on a sit stay, but other times it earns a harsh “No!” (This is one of the reasons vocalization is such a hard problem to fix, regardless of your training paradigm, because the criteria is muddy) Everyone is in agreement that dogs perform better under clear criteria.
So yes, the traditional trainer corrects every failure, and the clicker trainer reinforces every success.
But, just like the traditional trainer has many different forms of correction available, the clicker trainer has many different forms of reinforcement available.
In addition to primary reinforcers like food and play, positively trained cues are reinforcers. I can reward autosits by resuming heeling. I can reinforce a whistle sit by giving my dog a cast to the bird. Stopped contacts are reinforced by continuing the agility run. I can smile at my dog on a long sit, or scratch his butt after the first scent article. Finishing to heel reinforces outing the dumbell reinforces a quick pick up reinforces a fast send. I can reinforce practically anything with a series of nose touches, which I take great pains to make highly reinforcing.
This is not a universal position. A more common answer is putting the animal on a variable reinforcement schedule. Often this is done with pet dogs, where owners are very concerned about getting their animals off treats as soon as possible. Competition trainers are more tolerant of having reinforcement (and punishment) tools in training sessions. However, I feel that while variable reinforcement schedules render the behavior very resistant to extinction, it does not necessarily improve reliability. I feel that goal is better met by a robust rate of reinforcement with a wide variety of reinforcers.