It's probably most noticeable when you are so familiar with someone's voice.
The main problem with auto-tuning live is that you're just trading the pitch control for everything else - intonation, vibrato, etc. Exposing that "robot sound" ends up being another way to screw up your performance, and so there's no way to cheat. You still need to be an excellent singer to get close to the studio-perfected sound.
(If you're in the studio, of course, you can do overdubs, and you can use the more sophisticated Auto-Tune competitor Melodyne, which uses batch processing to allow smoother and more detailed corrections including pitch, rhythm and timbre, and in the most recent versions, chord manipulations.)
It is surprising that this stuff is only recently (ish) in mainstream attention - the technology has been around for a long time and was used to good effect occasionally.
It's not as prominent as in the first one (and I'm prepared to believe it's not used hamhandedly by the skilled engineers), but it's there: bottom row, slightly right of centre.
But let’s assume that the Manhattan Transfer is trying to hide the use of Auto-Tune, in which case their recording engineer would presumably use a retune speed that approximates a “natural” value.
Fire and forget mode: http://www.antarestech.com/images/ATEvo_Auto_mode.jpg
Notice the prominent 'retune speed' adjustment knob.
And here is the way everyone uses auto-tune these days, when you aren't going for the robot hiphop/Cher effect: http://www.antarestech.com/images/ATEvo_Graphic_full.jpg
Notice there is no retune speed knob in the main panel at all. You control the pitches by adjust them directly in the graph.
[I've edited my original post above to include some of this information.]
And this is just for "fire and forget" auto-tune. If you work an auto-tune by hand – and most skilled engineers will do this, when it's called for – you can use your own judgment as a musician to figure out what to do. Antares Auto-Tune lets you do this by presenting a sort of X-Y plot of pitch over time which you can manipulate. Melodyne does something similar, with a little more sophistication (it actually pioneered this technique.)
Another one of his incorrect assumptions is that auto-tune still behaves like that patent from 1998. While some of the algorithms involved are the same, the systems of today [1] use lookahead on the audio source, so that the pitch correction isn't constantly playing 'catchup' (it takes at least a coupe of cycles of an oscillating wave to determine its pitch with reasonable accuracy.) So you would not see something start wrong and then deviate towards its ideal pitch – in fact, you might see it start too accurately.
The primary destruction auto-tune wreaks on vocal performances isn't just too-perfect pitch: it destroys the subtly and intonation that good performers can impart. A real vocalist doesn't hit the ideal pitch at the start of each note, because it's hard to make the human voice do that. They use the imperfection to their advantage. There's a lot of subtle complexity that auto-tune can just wipe out when you use it.
Of course, if the vocalist isn't that good, then there's not any real loss.
As for his article, even assuming the methods he used to analyze the songs were not based on a misconception, the resolution is too low. Perhaps surprisingly, the human ear is capable of distinguishing pitch much more finely than the graphs he plotted.
Edited addendum: here are a couple of screenshots of the actual Antares Auto-Tune software that will illustrate what I'm talking about.
Here is a picture of the Auto-Tune software set in fire-and-forget 'auto' mode. http://www.antarestech.com/images/ATEvo_Auto_mode.jpg
Notice the prominent 'Retune Speed' adjustment knob. This is how you used it in the late 90s and early 00s, or today when you want the robot or Cher effect.
Now here is a picture of Auto-Tune in use in Graphical mode, where you adjust the pitches directly. This is how most engineers today will use Auto-Tune when it's not for a deliberate, noticeable effect. http://www.antarestech.com/images/ATEvo_Graphic_full.jpg
Notice there is no retune speed adjustment knob. Instead, you edit the pitches directly in a graph.
[1] In a non-live-performance setting, anyway. It introduces latency, so if you used this live, it would add 'lag' to the performer's input. Bad.