The actual calculation is with the arctanh(correlation) -- quoting from the help
++ Then carry out the t-test between/among these 3D correlation maps,
possibly allowing for dataset-level covariates.
-- Actually, between the arctanh() of these maps:
cf. RA Fisher: [en.wikipedia.org]
[you could do the arctanh() conversion manually via 3dcalc;]
[then do the t-tests manually with 3dttest++; then convert]
[the t-statistics to Z-scores using yet another 3dcalc run.]
-- To be overly precise, if the correlation is larger than 0.999329,
then the arctanh is clipped to 4.0, to avoid singularities.
In a 1 sample test, in the near vicinity of the seed point (where you clicked), the voxels will perforce be highly correlated with the seed, and so the "correlations" will be nearly 4.
I usually talk about "t-testing the correlations" because saying "t-testing the Fisher transformation of the correlations" more than once in a presentation is too lengthy and repetitive.