The results from refinements with six different weighting schemes in PROFFT, SHELXL93 and X-PLOR show that experimental weights, and especially unit weights, tend to overweight the low-angle data at the expense of the higher resolution data. This suggests that the use of unit weights are inappropriate except during initial cycles of refinements. Two alternative weighting schemes, one-line and two-line empirical weights, are shown to produce a relatively even distribution of the weighted r.m.s. error [r.m.s. Err = ( summation operator wDeltaF(2)/n)(1/2)] as a function of sin theta/lambda and result in linear delta(R) normal probability plots. Free R values and free weighted r.m.s. errors from the refinement of a scorpion toxin structure show that the alternative weighting schemes are superior to unit weights, and the r.m.s. deviations of bond distances and angles from the target values and G factors, as calculated by PROCHECK, confirm the superiority of the empirical weighting scheme. It was also observed that changing the value of WA, the contribution of the X-ray gradient to the total function minimized, produced little variation in R free but was directly proportional to the r.m.s. deviation of the bond distances from the target values.