I was asked to expand on my earlier comments, but in the meantime you have seen two excellent responses from Bruce King and James Lang. I agree with what they say.
The point is that fitting a model to transformed data and then transforming back to the original scale is not the same as fitting a model directly to the untransformed data. If SSEL denotes the sum of squared residuals for the linearized data and SSEU for the untransformed data, a technique (as used by TI in many cases) that minimizes SSEL will not, in general, simultaneously minimize SSEU. There are non-linear techniques available to minimize SSEU, but they are beyond what is covered in an introductory course (and are computer-intensive). So, r^2 does not have the same interpretation for, say, an exponential model fit by linearizing as it does for, say, a quadratic model fit by least squares, as is correctly pointed out by Bruce.
Incidentally, a LINEAR MODEL means that the model is linear in the parameters, not necessarily linear in its functional form. So, the quadratic model is still a linear model (and fit by methods of linear algebra).