Dielectric absorption looks like an R+C across your ideal capacitor, with the R relatively large (giving a long time constant, minutes to weeks) and C some small fraction of the bulk capacitance (usually 1% or less).
Consider what happens if the capacitor sits on a shelf. Over time, the two capacitances equalize in voltage. Then you pick it up and apply a voltage: you get two time constants, one dominant, another extremely slow and weak. The gain difference (between, say, 1Hz and 0.001Hz -- at least, one frequency well above and one well below the absorption's cutoff frequency) between these is very small, but since you're measuring the difference between signal and a reference, this is an error in your measurement. And even from the best capacitors, 0.01% is a whole lot of LSBs at 20 bits or more.
Tim