That's because they are predicting phi, psi angles regarding reference databases. It just predicts phi, psi angle what energetically stable structures will have for sequence you put. You will never know how correct they are before you get the real structure from X-ray or NMR. For example, Glycine's phi, psi angle dispersion is so broad that it is not possible to predict correct. Try PACSY server (pacsy.nmrfam.wisc.edu). You can check statistics for each amino acid. For example, 

select avg(PHI),stddev(PHI),avg(PSI),stddev(PSI) from G_STRC_DB 

You will see how Glycine is freely rotating by seeing high deviation numbers.

If you see Isoleucine with the following query,

select avg(PHI),stddev(PHI),avg(PSI),stddev(PSI) from I_STRC_DB

You will see much lower deviation in phi angle than Glycine, but still that does not guarantee that your prediction is always correct. 

As said Woonghee Lee, It depends of  cutoffs selected and reference database used. Secondary structures are described as correct / outliers / "helix" / "beta" .... according to phi/psi angle cutoffs. Cutoffs are sometimes slightly different, depending of the software or the training database used to determine them. On several ramachandran plots made by different softwares, spots should be exactly at the same place, but sec.str. regions may slightly differs. Phi/Psi are fixed, interpretation is not.

In real structures, outlier residues are often associated with model errors or true constrained residues, like active residues.

It is the angular distributions and cut offs that were implemented in these programs which provides certain amino acids are outliers or non-outliers.

Lovell et al., 2003 stated that

"An acceptable structure at a given resolution can be defined by having suitably high overall percentages of , values, sidechain rotamers, and C deviations within allowed ranges. However, the truly important claim is that by examining the outliers for each of those criteria in conjunction with all-atom contacts and electron density, and by correcting them when appropriate, essentially any protein structure can feasibly be rendered significantly more accurate than without these tools."

Please refer and cross-refer this paper:

http://mordred.bioc.cam.ac.uk/~rapper/downloads/lovell-proteins-2003.pdf