Cl, Br, and I have a "sigma hole" of positive charge along the C-Cl axis. This has been treated via a virtual particle with a positive charge in other force fields (CHARMM, OPLS) but perhaps the approach here is to assign a small positive charge to Cl to mimic this effect. I would say that the virtual site approach is more rigorously correct, though. If Cl bears a partial negative charge, you will get incorrect halogen bonding (repulsion from groups like carbonyls that are actually favorable due to the sigma hole).

Thank you, Justin for elaborating on the special case of the halogen bonding.

Does that mean that the effective charge on Cl should be negative and that on the C should be positive positive when using single atom to represent each Cl and carbon and not using virtual site?

Later I also looked at the charges for benzene in that GAFF package. Since all the carbon atoms in benzene are equivalent, I was expecting the same charge on each carbon but they were significantly different. Is there any explanation for that?

That's precisely the problem - if you assign delta+ on C and delta- on Cl without a virtual site, you will incorrectly model halogen bonding, which is essential for describing interactions with water and biomolecules.

I do not regularly use GAFF, but if its model of benzene has different charges on everything, I would say that is a poor model. But I am truly surprised it would be this way.